# CHANGELOG — Southern Spear **Document ID:** `Docs/CHANGELOG.md` **Purpose:** Rolling record of what was actually done, what was actually tested, and what is still open. Appended to at the end of every work session. **Last updated:** 2026-10-01 > **This file records evidence, not narrative.** A line here means a command was run and its result observed. If something was not done, it is not claimed. Anything marked `NOT RUN` is genuinely outstanding, not quietly skipped. --- ## Status At A Glance — historical inception snapshot (Session 001, 2026-09-26; not current) > This table records the initial Phase 0 state only; later dated sessions below supersede it for present project status. | | | |---|---| | Current phase | **Phase 0 — Audit & Architecture — COMPLETE** (gate G0.8 passed) | | Buildable? | ✅ **Yes.** `SouthernSpearEditor` compiles clean | | Playable? | ❌ **No.** Never opened in the editor; no map imported | | Automation passing | 1 suite (blockout layout verification, 10/10) | | Third-party assets in use | Lyra + UE only. Zero acquired assets | | Commit count | 6 | --- ## Session 001 — 2026-09-26 — Phase 0: Audit, Architecture & Source Control ### COMPLETED **Toolchain audit** — every claim backed by an executed command: - UE **5.8.3** (CL 58210709, `++UE5+Release-5.8`) at `E:\Unreal\UE_5.8` - Confirmed **Installed Build** via `Engine/Build/InstalledBuild.txt` → no engine-side C++, Lyra must be vendored - VS Community 2022 **17.14.37710.0**, MSVC **14.44.35207** vs engine floor **14.44.34918** (thin margin, now CI-asserted) - Blender **5.2.2 LTS**, Git **2.54.0**, git-lfs **3.7.1** - Hardware: Ryzen 7 9800X3D (8c/16t), 31.2 GB RAM, RX 9070 XT - Engine 5.8.3 **not** registered in the Epic Games Launcher → in-editor Fab cannot resolve it (risk R-03) **Source control established**: - Git repo on `main`, Unreal+Blender `.gitignore` - Line endings pinned to LF with CRLF override for `.bat`/`.cmd`/`.ps1` - Git LFS tracks `.uasset`/`.umap`/`.blend` and all binary art/audio types - **LFS verified end-to-end** with `git check-attr`, not assumed **Documents authored** (12 in `Docs/`): `PROJECT_AUDIT`, `GAME_DESIGN_DOCUMENT`, `TECHNICAL_DESIGN_DOCUMENT`, `ASSET_REGISTER`, `LICENCE_REGISTER`, `DEVELOPMENT_ROADMAP`, `TEST_PLAN`, `CODING_STANDARDS`, `ASSET_NAMING_STANDARDS`, `LYRA_ADOPTION`, `DECISION_LOG` (15 ADRs), `MAPS_DRYRIVER`. **Lyra 5.8 obtained and verified**: 20 plugins, full C++ source, 55 prebuilt binaries. Declares `EngineAssociation: "5.8"`. **Repository relocated** to `E:\SouthernSpear` (space-free) — ADR-014. **Dry River vertical-slice blockout generated** from a written design spec — 162 objects, 8,944 faces, 4 gameplay markers. ### FILES CHANGED Created: `README.md`, `.gitignore`, `.gitattributes`, `.github/workflows/build.yml`, `Docs/*` (12 files), `Tools/Blender/dryriver_blockout.py`, `Tools/Blender/verify_dryriver.py`, `Content/Art/Blockout/*` (blend, fbx, csv), `Source/SouthernSpear*.Target.cs`. Vendored (untracked, awaiting the G0.8 commit): Lyra source, plugins, config, content. ### TESTING | Check | Command | Result | |---|---|---| | LFS attribute resolution | `git check-attr filter diff merge text` | **PASS** — `.uasset`/`.blend` → `lfs`, `.md` → text | | CI workflow validity | `python -c "yaml.safe_load(...)"` | **PASS** — after fixing a real bug (Windows paths in double-quoted YAML were parsed as escape sequences) | | Lyra editor load | read `LyraStarterGame.log` | **PASS** — all 5 game features `[Registered, Active]`, **1 error total** (`LogPixelStreaming2RTC`, cosmetic), clean exit | | Blockout generation | `blender -b --python dryriver_blockout.py` | **PASS** — 162 objects, 0 errors, after fixing 2 bugs | | Blockout layout | `python Tools/Blender/verify_dryriver.py` | **PASS 10/10, exit 0** | | LFS on new binaries | `git lfs ls-files` | **PASS** — fbx + blend tracked | | **Gate G0.8 editor build** | `Engine\Build\BatchFiles\Build.bat SouthernSpearEditor Win64 Development` | ✅ **PASS — `Result: Succeeded`, 429 actions, 0 errors, 1188 s** | | Dedicated server build | `Build.bat SouthernSpearServer ...` | **NOT RUN** | | Project opens in editor | — | **NOT RUN** | | NavMesh on Dry River | — | **NOT RUN** — needs the project opened | ### ASSETS Created (all **class F**, original, no licence obligation): - `SS_MAP_DryRiver_01_HI.blend` — source blockout - `SS_MAP_DryRiver_01.fbx` — game import mesh - `SS_MAP_DryRiver_01_Layout.csv` — gameplay marker positions - `dryriver_blockout.py` — procedural generator - `verify_dryriver.py` — layout verifier (CI) Vendored (licence **L-0001**): Lyra Starter Game 5.8; Unreal Engine 5.8.3 (**L-0002**). **Licence entries added**: L-0001 through L-0012. **Third-party assets acquired**: **zero**. Electric Dreams considered and **declined** (ADR-013). **Remaining placeholders**: all 15 weapons, 16 of 18 characters, 9 of 13 animations, all 13 audio, all 6 effects, 11 of 12 UI, 6 of 7 maps, all 10 data assets. Nothing is final. ### RISKS | ID | Risk | Status | |---|---|---| | R-01 | Lyra may not target 5.8.3 | ✅ **CLOSED** — `EngineAssociation "5.8"`, runtime log exact build match | | R-02 | Path contains a space | ✅ **CLOSED** — relocated to `E:\SouthernSpear` (ADR-014) | | R-03 | Engine not Launcher-registered → Fab unavailable | 🔴 **OPEN** — also gates the abandoned Electric Dreams harvest | | R-04 | MSVC margin over engine minimum is thin | 🟡 **OPEN** — CI asserts the version | | R-05 | 31.2 GB RAM for editor + DS + 4 clients | 🟡 **OPEN** — affects the 4-client acceptance test | | R-06 | E: has ~488 GB free | 🟡 **OPEN** — monitor as LFS objects accumulate | | R-07 | UE 5.8.3 is recent; ecosystem may lag | 🟡 **OPEN** — verify every plugin on 5.8 | | R-08 | Installed Build — no engine modules | 🟡 Accepted — all divergence is project-side | | **G0.8** | Vendored fork has never compiled | ✅ **CLOSED — `Result: Succeeded`, 0 errors** | | **P1-01** | Editor build does not prove the project *runs* | 🔴 **OPEN — project has never been opened** | **Network risks:** none yet — no netcode exists. **Licensing concerns:** insignia on legal hold (L-0003); ADF marks not licensed (L-0004); prohibited sources barred (L-0007, L-0008). ### DEFECTS FOUND AND FIXED | Defect | Severity | How found | |---|---|---| | CI workflow YAML unparseable — Windows paths in double-quoted strings read as escape sequences | S1 — build-breaking | `yaml.safe_load` | | **`RunUBT.bat` does not exist in UE 5.8.3** — documented as the build command in 5 files | **S1 — build-breaking** | **Actual build attempt; only `RunUBT.sh` exists. Correct entry point is `Build.bat`** | | Blender 5.2 cone operator takes `radius1`/`radius2`, not `radius` | S2 | Script run → TypeError | | **OBJ A 10 m off the Y centre line** — gave one team a shorter run to the opening contest | **S2 — gameplay balance** | **`verify_dryriver.py`** | | **Spec distances did not match generated geometry** (claimed 62 m, actual 51.9 m) | **S2 — false documentation** | **`verify_dryriver.py`** | | Objective order (sequential vs parallel) never specified, leaving balance ambiguous | S2 — design | Verifier could not assert fairness | **Self-inflicted process errors (recorded honestly):** reported Lyra's core plugins as "missing" when they were one directory deeper (non-recursive glob); clobbered and then duplicated a section of `DECISION_LOG.md` through several rounds of careless line-editing. Both were corrected and committed rather than quietly patched over. ### NEXT ACTION **Open the project in the editor for the first time** and confirm it loads under its new name with all game features active — then import the Dry River FBX and generate a NavMesh. This is the first moment the project will have been *run* rather than merely *built*, and it validates the last unproven assumption left over from the fork. --- ## Session 002 — 2026-09-26 — First Run, and a Hard Stop on the Dedicated Server ### COMPLETED **Gate G0.10 PASSED — the project runs.** Risk P1-01 is closed. The vendored, renamed fork was launched headless for the first time: ``` Engine\Binaries\Win64\UnrealEditor-Cmd.exe SouthernSpear.uproject -nullrhi -unattended -nosplash -nosound -stdout ``` - Loads as `Running engine for game: SouthernSpear` on `5.8.3-58210709+++UE5+Release-5.8` - Uses our renamed target receipt, `Binaries/Win64/SouthernSpearEditor.target` - **All five game features reach `Ending state: Registered [Registered, Active]`** — `TopDownArena`, `ShooterCore`, `ShooterExplorer`, `ShooterMaps`, `ShooterTests` - **Zero errors, zero fatals** - Game feature DLLs resolve to the **vendored** tree at `E:\SouthernSpear\Plugins\GameFeatures\…`, confirming the original Lyra staging copy at `E:\Unreal\Lyra` is correctly inert The last unproven assumption left over from the fork is now a proven one. `TECHNICAL_DESIGN_DOCUMENT.md` and `LYRA_ADOPTION.md` stand as written, with no further divergence needed. **Gate G0.9 FAILED — and it is not fixable in-project.** Attempting the dedicated server target: ``` Build.bat SouthernSpearServer Win64 Development -Project=…\SouthernSpear.uproject -WaitMutex → Server targets are not currently supported from this engine distribution. → Result: Failed (OtherCompilationError) Total execution time: 2.06 seconds ``` It failed in under three seconds, before compiling a single file, so it cannot be a source error. Traced to the root cause in the engine's own configuration: - `Unreal.IsEngineInstalled()` is `true` (`InstalledBuild.txt` = `UE_5.8`) - `UEBuildTarget.cs:1396` gates on `InstalledPlatformInfo.IsValid(…, InstalledPlatformState.Supported)` - That reads `[InstalledPlatforms]` from the **engine's own** `Engine/Config/BaseEngine.ini:3996` - The 28 declared configurations contain exactly two distinct values: `PlatformType="Editor"` and `PlatformType="Game"`. **No `Server` entry exists.** - Corroborated by the absence of any `UnrealServer` binary or `.target` in `Engine/Binaries/Win64/` **This engine install is a game/editor-only distribution.** Epic ships dedicated-server capability as a separate Launcher product, or it is present in a source build. No project-side setting can add it, and `SouthernSpearServer.Target.cs` is correct and ready for the day a server-capable engine exists. **Risk R-06 closed.** E: re-measured at **475 GB free, 51 % used** of 954 GB. The earlier "488.6 GB, tightest volume" note was written before the 2.4 GB LFS clone landed; headroom is ample. ### FILES CHANGED - `Docs/PROJECT_AUDIT.md` — added §6.1 (G0.9 failure, full root-cause trace) and §6.2 (G0.10 pass, evidence table); added **R-09**; struck R-06 as resolved; added producer question 5 with a costed recommendation - `Docs/CHANGELOG.md` — this entry - **`Docs/evidence/` (new directory, tracked)** — the gate evidence is version-controlled, because `Build/` is gitignored and these transcripts are the proof that G0.9 failed and G0.10 passed. A claim in a design document that cannot be re-checked is not a claim. - `G009_server_build_failure.txt` — G0.9 transcript - `G009_baseengine_installedplatforms_excerpt.txt` — the engine's own `[InstalledPlatforms]` whitelist - `G009_installedplatforms_primarysource.txt` — whitelist plus the `UEBuildTarget.cs` throw site - `G010_editorload_keylines.txt` — the 12 lines that decide G0.10 - `G010_editorload_filtered.txt` — full editor log, telemetry DNS-failure spam stripped - A directory junction `E:\Australian Army Game\SouthernSpear` → `E:\SouthernSpear` was created **outside the repository** so the agent's file tools can reach the relocated project. It is not tracked by git and is not part of the project. ### TESTING | Command | Result | |---|---| | `Build.bat SouthernSpearServer Win64 Development -WaitMutex` | **FAIL** — refused by engine distribution, 2.06 s, 0 actions | | `UnrealEditor-Cmd.exe SouthernSpear.uproject -nullrhi -unattended` | **PASS** — 5/5 game features active, 0 errors | | `df -h E:` | 475 GB free | ### ASSETS No new assets. Dry River FBX still **not imported**; NavMesh still ungenerated. ### RISKS - **R-09 (NEW, HIGH)** — this engine distribution cannot build `TargetType.Server`. Directly blocks the vertical slice's own acceptance criterion ("packaged client → packaged dedicated server"), which the roadmap explicitly refuses to accept a listen server for. Escalated as producer question 5. - **R-06** — **RESOLVED**, re-measured at 475 GB free. - **P1-01** — **CLOSED** by G0.10. - R-03, R-04, R-05, R-07, R-08 unchanged. ### DEFECTS FOUND One, and it is an environment defect rather than a code defect: - **Server target unavailable on the current engine install.** Found by attempting the build rather than assuming it would work — the `SouthernSpearServer.Target.cs` file had been written and committed in Phase 0 on the strength of Lyra's convention, and that assumption was simply untested until now. It would have surfaced only at vertical-slice sign-off, which is far too late. Found early, at the cost of one build invocation. No project-side defects were found this session. ### NEXT ACTION **Build the `SouthernSpearCore` plugin carrying the critical path** — `ESS_TeamId` (ADR-003, match-relative) and the `UFSSFactionPresentationSet` presentation resolver (ADR-004, structurally incapable of affecting gameplay). This work builds against the Editor target and is completely independent of how R-09 is resolved, so it proceeds while the producer decides. --- ## Session 003 — 2026-09-26 — Dry River Brought Into the Engine, and a Coordinate Bug ### COMPLETED **Gate G1.1 PASSED. NavMesh generation is validated and P1-02 is closed.** The map is no longer a Blender file nobody has seen inside the engine — it is built, collisioned, navigable and saved, entirely from version-controlled scripts. | Check | Result | |---|---| | FBX import | 162 blockout objects → one mesh `SS_MAP_DryRiver_01` (via Interchange) | | Collision | `CTF_USE_COMPLEX_AS_SIMPLE`, profile `BlockAll` | | Gameplay actors | 4 placed from the layout CSV (2 deployments, 2 objectives) | | Map check | **0 errors, 0 warnings** | | Nav bounds | Min (-13600, -9600, -1000) → Max (13600, 9600, 2600) cm | | Tiles generated | **560** | | **Path, DeployAlpha → DeployBravo (170 m)** | ✅ **verified, 2 path points** | | Saved map | 248 KB vs 8.5 KB empty — navigation data is serialised | The acceptance check is a **path query**, deliberately. A navmesh actor existed, and the build reported success, while the navmesh covered *nothing*. Only a non-empty `find_path_to_location_synchronously` result across the full map means what it says. **Producer decision taken: obtain the Dedicated Server engine product** (route (a) in `PROJECT_AUDIT.md` §8). R-09 stays open until that engine is installed; no workaround will fake it. ### DEFECTS FOUND **1. The Y axis was mirrored, silently swapping the two deployments.** Blender is right-handed Z-up, Unreal is left-handed Z-up, so the FBX import mirrors Y. The layout CSV is Blender space, so `Unreal_Y = -Blender_Y`. Using it un-negated is invisible: Dry River is symmetric in Y, so the deployments simply swap ends and **every distance in the map spec still holds**, because a mirror preserves distance. A map that was subtly wrong in exactly the way nobody would notice by looking. It was caught only by cross-checking the traced ground height against the heights the Blender generator recorded: ``` DeployAlpha ground z=64.3cm vs CSV 34.3cm (delta 30.0cm) DeployBravo ground z=1332.0cm vs CSV 1302.0cm (delta 30.0cm) ``` The residual 30 cm is expected — the layout markers are 0.3 m boxes on the terrain, so a trace lands on the box top. The important part is that both deltas are equal and small; before the fix they were ~13 m and the two deployments had exchanged heights. **Pass 2 now asserts this delta stays under 150 cm and warns otherwise**, so the regression cannot return quietly. **2. `MapCheck: PlayerStart_0 is a normal APlayerStart, replace with ALyraPlayerStart`.** Now uses `ALyraPlayerStart`; map check is clean. **3. A cosmetic bug in the reporting code cost the entire constructed level.** `list(volume.get_actor_scale3d())` raised `TypeError: 'Vector' object is not iterable` *after* the actors were placed but *before* the save, so a map with no actors was left on disk and the second pass found nothing to work with. Pass 1 now saves in a `finally`, because a reporting bug must never cost the build. ### Five engine behaviours that made a one-pass build impossible Each was found by reading engine source, not guessed. All are documented in `MAPS_DRYRIVER.md` §11.3. | Symptom | Cause | |---|---| | Post-tick callback never fires | `-ExecutePythonScript` destroys the world as soon as the script returns | | `navigation build is locked (flags: 0x20)` | `AsyncLoadLock` (`1 << 5`), released only on a tick. Overridden per-invocation, not in project config | | `TotalNavBounds: IsValid=false` | `GatherNavigationBounds()` skips volumes failing `HasActorRegisteredAllComponents()`; a script-spawned brush never reaches that state. **A map load does** | | Build succeeds in 0.00 s, zero tiles | The import task saved the mesh *before* collision was configured, so the `.uasset` had none. Pass 1 now saves it explicitly afterwards | | Nav data present but inert | Spawning `RecastNavMesh` from script defers registration to a tick; the map load creates it properly | Every step uses the same code path a person uses in the editor: place the volume, then Build Paths. **No workaround fakes anything.** ### FILES CHANGED - **`Tools/Unreal/build_dryriver_level.py`** (new) — pass 1: level construction - **`Tools/Unreal/build_dryriver_nav.py`** (new) — pass 2: nav build, path verification, ground-height cross-check - **`.github/workflows/build.yml`** — both passes plus a hard assertion on the G1.1 report; `Build Dedicated Server` marked `continue-on-error` with the R-09 reason inline so CI is not permanently red over a known environment gap - `Docs/MAPS_DRYRIVER.md` — new §11 (pipeline, coordinate transform, measured engine constants, verification table) - `Docs/PROJECT_AUDIT.md` — new §6.3 (G1.1), gate table, P1-02 closed - `Docs/CHANGELOG.md` — this entry - `Docs/evidence/G011_*` (new, tracked) — keylines and both JSON reports - `Content/Maps/L_DryRiver_01.umap` (new, LFS) — the level with baked navigation data - `Content/Art/Blockout/SS_MAP_DryRiver_01.uasset` (new, LFS) — the imported blockout mesh, saved with collision ### TESTING | Command | Result | |---|---| | Pass 1 — `build_dryriver_level.py` | **PASS** — level, 4 actors, mesh + collision saved | | Pass 2 — `build_dryriver_nav.py` | **PASS** — 560 tiles, 2 path points, `ok=true` | | Pass 2 re-run (idempotency) | **PASS** — identical result, so the pipeline is stable, not lucky | | `MapCheck` | 0 errors, 0 warnings | ### ASSETS - `Content/Maps/L_DryRiver_01.umap` — 248 KB, navigation data serialised - `Content/Art/Blockout/SS_MAP_DryRiver_01.uasset` — 366 KB, complex-as-simple collision - No third-party assets were used, so no new `LICENCE_REGISTER.md` entry is required. ### RISKS - **P1-02** — **CLOSED** by G1.1. - **R-09** — open, now with a chosen route: obtain the Dedicated Server engine product. - No new risks. The two engine-distribution gaps found here (async nav lock, brush-volume registration) are recorded in `MAPS_DRYRIVER.md` §11 rather than the risk register, because both are worked around and neither threatens the project. ### NEXT ACTION **Build the `SouthernSpearCore` plugin** carrying the critical path VS-01 → VS-05 → VS-06: `ESS_TeamId` (match-relative per ADR-003) and the `UFSSFactionPresentationSet` resolver (ADR-004, structurally incapable of affecting gameplay), with a CI static scan enforcing the separation. Unaffected by R-09. --- ## Session 004 — 2026-09-26 — Data-Driven Dressing, and a Limit I Could Not Remove ### COMPLETED **Dry River dressing is now a data edit.** Moving a fence or deleting a wreck is a one-line diff in a CSV, with no editor involved. Verified end to end: 19/19 data checks, 290 dressing actors placed from the CSVs, and the path across the map still verified. | Artefact | Result | |---|---| | Dressing placement data | 67 instances (46 scrub, 10 barrel, 8 crate, 3 wreck) | | Fence runs | 7 runs → 79 posts, 144 rails generated from the run lines | | `Tools/verify_dressing.py` | **19/19 PASS**, no Blender, no editor, seconds | | Collision at placement | **5/5** solid types, ray-verified | | Path, DeployAlpha → DeployBravo | still verified | The pipeline is now `construct → dress → navigate`. Dressing runs **before** navigation deliberately: placed afterwards, the fences would be invisible to the NavMesh and agents would walk through them, which is worse than a map that is plainly undressed. ### DEFECTS FOUND **1. All 290 dressing actors were placed 100× too close to the origin.** `spawn()` divided by 100 on the way in and handed **metres** to `spawn_actor_from_class`, which takes **centimetres**. Everything clustered in a 1.3 m knot at (0,0): the map looked correct in the data file and bare in the engine, and every ray at the real coordinates hit open ground. Found by adding the collision self-check, which was the only thing that could tell "wrong place" from "no collision". **2. Hand-placed wrecks bypassed the clearance check.** `WRECKS` were authored by hand and never passed through `clear_of_protected`, so two of three sat inside a deployment zone — 59 m from a spawn that requires 84. The CI validator caught it; the generator now routes hand-placement through the same check, refusing with a reason rather than emitting. **3. The first four authored fence runs were bad and the validator said so.** Two of five were rejected: one passed 20.6 m from a deployment zone needing 74, the other 24.2 m from OBJ B needing 30. Relocated and re-validated rather than quietly accepted. The deployment radii are large enough that almost the whole northern and southern thirds are off-limits to long fences, which is a useful thing to learn about this map. **4. A barrel stood 1.23 m from a fence line**, caught by a new check. The generator was not fence-aware; it is now, and the data is correct by construction. **5. A `KAggregateGeom` string test misfired**, so a collision-type change silently did not apply and the report claimed a change that never happened. Replaced with an unconditional set plus a read-back. **6. The replacement for defect 5 never actually ran.** The unconditional set asked for `unreal.CollisionTraceFlag.CTF_USE_SIMPLE_AS_SIMPLE`, an enum member that **does not exist** — the live 5.8.3 enum has only `CTF_USE_DEFAULT`, `CTF_USE_COMPLEX_AS_SIMPLE`, `CTF_USE_SIMPLE_AND_COMPLEX` and `CTF_USE_SIMPLE_AS_COMPLEX`. A broad `except` caught the `AttributeError` and downgraded it to a warning, so the set silently no-opped on all six meshes and the run still reported `ok`. The claim in defect 5's fix that it was “replaced with an unconditional set plus a read-back” was true of the *code* and false of the *effect*: the read-back existed, the write never happened. Found by reading the report's own warnings instead of skimming past them, then confirming the real member list by introspecting the enum in a live editor rather than trusting the C++ header from memory. It was harmless in effect — the importer had already left every mesh at `CTF_USE_DEFAULT`, which is what the code meant to set — so this was not R-10's cause. But a pass whose job is to guarantee collision was reporting success without having written the value it reported writing, and the dress report was carrying six warnings that all read as benign. Fixed: the correct member is used, the applied value is compared against the requested one, and **both** a mismatched trace flag and a mesh with zero simple-collision elements are now hard failures rather than warnings. **8. Fixing defect 6 exposed a bad experiment, which is the more useful half of this.** The project had recorded “collision type” as ruled out for R-10, with both complex-as-simple and simple tried and both failing. The complex-as-simple attempt had gone through the very write that defect 6 shows never ran, so the flag never changed and **the experiment proved nothing while sitting in the docs as settled fact.** With the write fixed, `CTF_USE_COMPLEX_AS_SIMPLE` was applied and confirmed by read-back (``) — and R-10 reproduced identically, 0/3 after reload. The hypothesis is now closed by measurement. A conclusion drawn from a call that silently did not happen is worse than no conclusion, because it removes the hypothesis from the list without ever testing it. **7. The collision diagnostic could not detect missing collision.** `str(KAggregateGeom)` is an opaque pointer whether or not the struct holds anything, so the `body_agg_geom` report field could never tell “this asset has no collision” from “this asset is fine” — precisely the two cases it existed to separate. It also listed `capsule_elems` and `geom_elems`, which are not properties of `KAggregateGeom` on this engine and raised a property error on every run. Now counts the element arrays, and the first real run immediately showed all six meshes carrying exactly one `convex_elems` entry. Net effect on the pass: **6 warnings → 0**, and the collision state of every mesh is now positively asserted rather than assumed. ### The one I could not remove **Dressing collision does not survive a headless save and reload.** Solid 5/5 immediately after placement; 0/3 after reopening the map, 0/7 fence runs blocking. The map is traversable but the fences are **not yet load-bearing cover**, and navigation does not include them. Ruled out by measurement, not assumption: placement position (a PaddockEast post reads 78.0, −34.0 m — exactly the CSV), collision profile (`BlockAll` and `QUERY_AND_PHYSICS` both survive), asset contents, collision type (see defect 8 — now genuinely tested rather than assumed), and stale imports (reproduced after deleting all six assets and re-importing). Later still exonerated the assets *as saved*: `Tools/Unreal/probe_collision_flag.py` reopens the project in a fresh editor with no import in play and reads every dressing `BodySetup` back off disk — all six at `CTF_USE_DEFAULT`, each with one `convex_elems` entry, so the simple collision is genuinely in the `.uasset` files. What remains is sharper than before. Every observable is correct after reload — position, profile, `collision_enabled`, `can_ever_affect_navigation`, the asset's trace flag and its convex hull — and the ray still returns `hit=False by=None`, while the blockout in the same map with the same profile and save path collides. The one live difference left is **which session last touched the asset**: the blockout was imported once in the construct session and carried forward untouched, while the dress pass deletes and re-imports all six dressing meshes on every run and saves the map in that same session. That is the next thing to vary, and it is a one-line change. The blockout mesh still collides correctly, so the cause is narrower than any of the ruled-out items, but I did not isolate the engine mechanism and stopped rather than keep guessing. I have **not** hidden this. The nav pass reports the two `*_after_reload` steps as failures, sets `dressing_collision_persists: false`, and emits a `WARN`. They deliberately do not gate `report["ok"]`, because CI would be permanently red for a reason we cannot fix and the data is already proven good by `verify_dressing.py`. Gating them would be honest but useless; passing them silently would be a lie. Tracked as **R-10**, with the most likely next step being a test of an interactive editor save. ### FILES CHANGED - **`Tools/Common/dryriver_spec.py`** (new) — the Dry River layout and terrain in pure Python, no `bpy`. The blockout, the dressing generator and the CI validator all import it, so they cannot disagree about where the ground is. Refactoring the committed blockout generator to use it was verified behaviour-preserving: layout CSV **byte-identical**, still 162 objects / 8,944 faces, `verify_dryriver.py` still passing. - **`Tools/Blender/dryriver_dressing.py`** (new) — mesh library (6 meshes) + initial layout - **`Tools/Unreal/dress_dryriver.py`** (new) — the dressing pass - **`Tools/verify_dressing.py`** (new) — the 19-check data validator - **`Tools/Unreal/probe_collision_flag.py`** (new) — one-off diagnostic that introspects the live `ECollisionTraceFlag` enum and counts `KAggregateGeom` element arrays per dressing mesh. Kept because it is what narrowed R-10 on the asset side, and because re-running it is how that conclusion would be rechecked. - `Content/Art/Blockout/SS_Dressing_*.fbx` + `_HI.blend` (new, LFS) — 6 meshes, 160 faces total - `Content/Art/Blockout/SS_MAP_DryRiver_02_Dressing.csv` / `_Fences.csv` (new) — the placement data - `Content/Art/Dressing/SS_Dressing_*.uasset` (new, LFS) — imported meshes - `Tools/Blender/dryriver_blockout.py` — now imports the shared spec; **behaviour proven unchanged** - `Tools/Unreal/build_dryriver_nav.py` — dressing-aware diagnostics, plus the reload collision report - `.github/workflows/build.yml` — data validator + dressing pass - `Docs/MAPS_DRYRIVER.md` — new §12; `Docs/PROJECT_AUDIT.md` — R-10 ### TESTING | Check | Result | |---|---| | `python Tools/verify_dressing.py` | **19/19 PASS** | | `python Tools/Blender/verify_dryriver.py` after the spec refactor | **PASS**, layout CSV byte-identical | | Dressing pass | 67 + 79 posts + 144 rails placed, **5/5 solid**, `ok=true` | | Nav pass | path verified (2 points), saved, `ok=true` | | Dressing collision after reload | **0/3** — reported, not hidden, tracked as R-10 | ### ASSETS Six original greybox dressing meshes (160 faces total), all generated procedurally. No third-party assets, so no `LICENCE_REGISTER.md` entry. The wreck is deliberately anonymous: a recognisable real vehicle would be a trademark problem, and the opposing force is fictional in any case. ### RISKS - **R-10 (NEW)** — dressing collision does not survive a headless save/reload. Characterised, not fixed. Detailed above and in `MAPS_DRYRIVER.md` §12.4. - **R-09** — unchanged, open, route (a) chosen. ### NEXT ACTION **Build the `SouthernSpearCore` plugin** — `ESS_TeamId` (ADR-003) and the `UFSSFactionPresentationSet` resolver (ADR-004) — unaffected by R-09 and R-10. R-10's next concrete step is now the cheapest untried variable, not the editor test: keep the existing dressing assets instead of deleting and re-importing them in `dress_dryriver.py`, and re-run the dress and nav passes. That is the one difference still standing between the dressing and the blockout, which survives. --- ## Session 005 — 2026-09-26 — Terminology Settled, ADR-003 Superseded, and SouthernSpearCore Built ### COMPLETED **The fictional canon is now fixed, and the team-identity architecture it depends on is implemented and tested.** Two architecture decisions were recorded this cycle, both superseding earlier work: - **ADR-016** — canonical organisations. **CDS** (Commonwealth Defence Service), **CLS** (Commonwealth Land Service), **ACR** (Australian Commonwealth Regiment), **2 CG**, **SOR**, **MAF** (Murasian Armed Forces), **CMECU**. `Docs/ORIGINAL_BRIEF.md` is left untouched as an immutable historical record; ADR-016 is the superseding authority. - **ADR-017** — **supersedes ADR-003.** `ESSTeamId {None, TeamOne, TeamTwo}` is authoritative and replicated; `ESSLocality {Friendly, Opposing}` is derived per viewer and never replicated. ADR-003's intent survives; its mechanism did not. ADR-016 and ADR-017 were previously *unused placeholders* for Phase 5 questions. They are now taken, and the placeholders were renumbered to ADR-018..ADR-021. No question was answered or dropped by the renumbering. `Plugins/SouthernSpearCore` is created and is the root of the SS dependency graph: it depends on **no other Southern Spear plugin and on no Lyra module**, so team identity cannot reach weapons, damage, health, abilities, roles, objectives or UI. It provides `ESSTeamId`, `ESSLocality`, `FSSViewerContext`, `FSSStableId`, `USSProjectSettings`, `LogSSCore`, shared validation helpers, and 32 native Gameplay Tags under the `SS.` root. Terminology was migrated across active documents. `Kestrel Militia`, "hostile militia", "insurgent" and "Australian force (friendly)" are gone from active design docs. Historical occurrences inside ADR bodies are **retained deliberately** — a rejected alternative stays in the record. The **Electric Dreams contradiction is corrected**. `LICENCE_REGISTER.md` L-0012 previously read "reference + selective harvest only" and listed "approved use: generic modular pieces only", directly contradicting ADR-013, which declined the pack **entirely**. It now reads DECLINED / NOT IMPORTED / NOT PERMITTED, states that no selective harvesting is authorised, and retains the historical evaluation as a non-authoritative record. ### FILES CHANGED **Created** - `Plugins/SouthernSpearCore/SouthernSpearCore.uplugin` - `Plugins/SouthernSpearCore/Source/SouthernSpearCore/SouthernSpearCore.Build.cs` - `.../Public/` — `SSCoreLog.h`, `SSTeamTypes.h`, `SSViewerContext.h`, `SSTeamIdentityLibrary.h`, `SSStableId.h`, `SSCoreValidation.h`, `SSNativeGameplayTags.h`, `SSProjectSettings.h` - `.../Private/` — `SouthernSpearCoreModule.cpp`, `SSTeamIdentityLibrary.cpp`, `SSStableId.cpp`, `SSCoreValidation.cpp`, `SSNativeGameplayTags.cpp`, `SSProjectSettings.cpp` - `.../Private/Tests/` — `SSTeamIdentityTests.cpp`, `SSDataValidationTests.cpp` - `Tools/validate_architecture.py` - `Docs/evidence/G020_*` (4 files) **Modified** - `SouthernSpear.uproject` — registers the plugin - `.github/workflows/build.yml` — architecture guard step + report upload - `Docs/DECISION_LOG.md` — ADR-016, ADR-017, ADR-003 superseded, placeholders renumbered - `Docs/PROJECT_AUDIT.md` — gates G1.2/G2.0/G2.1, §6.4, §6.5, R-11 - `Docs/LICENCE_REGISTER.md` — L-0012 corrected; §5 rewritten - `Docs/ASSET_REGISTER.md` — C-014..C-017 terminology - `Docs/ASSET_NAMING_STANDARDS.md`, `Docs/GAME_DESIGN_DOCUMENT.md`, `Docs/DEVELOPMENT_ROADMAP.md`, `Docs/TEST_PLAN.md` — terminology **Removed:** none. ### TESTING All commands run from `E:\SouthernSpear`. | # | Command | Exit | Result | |---|---|---|---| | 1 | `Build.bat SouthernSpearEditor Win64 Development` | **0** | `Result: Succeeded`, 0 errors | | 2 | `UnrealEditor-Cmd … -ExecCmds="Automation RunTests SouthernSpear.Core; Quit" -TestExit=…` | **0** | **9/9 PASS** | | 3 | `python Tools/verify_dressing.py` | **0** | 19/19 PASS | | 4 | `blender -b --python Tools/Blender/verify_dryriver.py` | **0** | ALL CHECKS PASSED | | 5 | `python Tools/validate_architecture.py` | **0** | PASS, no violations | | 6 | Guard, 2 violations injected | **1** | Both violations reported (SS002, SS006) | | 7 | Guard, violations reverted | **0** | PASS; source byte-identical (`md5sum`) | **Project load and Game Feature activation** were observed in run 2, which loads the project before running tests: `LogSSCore: SouthernSpearCore module started.`, all 5 Lyra features (`TopDownArena`, `ShooterCore`, `ShooterExplorer`, `ShooterMaps`, `ShooterTests`) reach `[Registered, Active]`, and the run contains **0 project errors** excluding offline telemetry warnings. **NOT RUN — Dry River navigation regression.** A nav re-run was started and interrupted during editor startup; the editor was killed before the Python pass began, so it wrote nothing. `L_DryRiver_01.umap` retains the timestamp of the last verified pass and was not touched. G2.0 changed no map or map-generation input, so there is no evidence of regression, but that is an inference and not a test. Tracked as **R-11**. **NOT RUN — a dedicated load-only invocation.** A separate `-ExecCmds="Quit"` load check exceeded 600 s in this environment: the editor wedges in telemetry retry loops because the sandbox has no DNS for `datarouter.ol.epicgames.com`, and `Quit` does not take effect. This is an environment limitation, not a project defect, and the same load is covered by run 2. A CI-safe flag set for it is still to be found. **NOT RUN — anything from Session 004 onward that is not listed above.** In particular `SouthernSpearTeam`, the cosmetic faction-presentation resolver, the 3 ACR / MAF presentation data, the `SSExp_ObjectiveAssault` skeleton and A88 integration are all **not started this cycle**. ### ASSETS No assets created, imported or modified this cycle. No Blender work. The 6 dressing FBX and `SS_Dressing_HI.blend` remain **PLACEHOLDER** and uncommitted from Session 004. Asset Register entries C-014..C-017 were updated for terminology. No new register rows were required — the cycle added code, not content. Licence Register L-0012 corrected; no new licences. ### RISKS - **R-11 (new, LOW)** — Dry River navigation not re-verified after G2.0. Noted above. - **R-10 (open)** — dressing collision does not survive save/reload. Unchanged. - **R-09 (open, unchanged)** — this engine distribution cannot build `TargetType.Server`. The dedicated-server acceptance criterion remains unsatisfiable; route (a) still chosen. - **Branding** — final insignia, camouflage, uniforms, weapons and store presentation still require **independent Australian legal review**. Renaming does not confer clearance. - **Technical debt** — `ESSLocality` has no `None` member, so a default-constructed value is `Friendly`. Safety lives in the resolver API, not the type. Recorded in ADR-017 as a known sharp edge rather than left as a trap. ### DEFECTS FOUND **1. `SSCORE_API` was undefined, and the first build failed with 20 errors.** UBT auto-generates `SOUTHERNSPEARCORE_API`, not the short `SSCORE_API` the naming standard wants. Aliased in `Build.cs` via `PublicDefinitions`. Found by the build, as it should be. **2. A `class FGameplayTag;` forward declaration in `SSCoreValidation.h` collided with the engine's definition as a `struct`**, producing a confusing cascade inside engine headers (`GameplayTagContainer.generated.h`), not in my file. Replaced with the real include. The error pointed at engine code; the cause was ours. **3. Two of my own tests failed for the wrong reason.** `LocalityFailsSafely` and `SpectatorContext` both assert on code that logs an error **on purpose**, and UE treats an undeclared error log as a test failure. The behaviour was correct — the log shows all three failure reasons firing exactly as designed — and the fix was to declare the errors expected, so the tests still fail if the messages ever stop being produced. Worth recording: a test that fails because it correctly provoked an error is not a failing feature, and reading the log before changing code is what told the difference. **4. The architecture guard's first version had a false positive.** Rule SS006 flagged `ESSLocality` for declaring `Friendly`/`Opposing` — that is precisely what the enum is for. The rule was wrong, not the code. Rewritten to apply the prohibition to the authoritative team enum only, and inverted into a positive check that the locality enum keeps both values. Caught by running the guard against real code rather than trusting that it was correct. **5. A dangling `const TCHAR*` in `FSSStableId::ValidateAndLog`.** One branch assigned `*FString::Printf(...)` to a `const TCHAR*`, which points into a temporary that dies at the end of the switch. Changed to `FString`. Found while reviewing my own code after the build was green, not by the compiler. ### NEXT ACTION **Re-run the Dry River navigation regression to close R-11** — run `Tools/Unreal/build_dryriver_nav.py` and assert 560 tiles plus a non-empty `find_path_to_location_synchronously` path between the two deployments. It is the only outstanding verification from work already done, and it is a single 6-minute command. --- ## Session 006 — 2026-09-26 — R-11 Closed, Guard Blind Spot Fixed, SouthernSpearTeam Presentation ### COMPLETED - Audited the uncommitted tree (Sessions 004/005) and committed it as `8d3bd762`. - Reviewed SouthernSpearCore: no SS sibling or Lyra dependency, no SS source in Lyra modules, ESSLocality not replicated (reflection test), no deliberately broken guard state left behind. - Closed R-11 (Dry River navigation re-verified). - Added `Plugins/SouthernSpearTeam`: cosmetic `FSSFactionPresentation` / `FSSFactionPresentationTable`, `ESSFactionPresentationId {None, ACR3, MAF}`, stateless `FSSFactionPresentationResolver` with explicit failure (no fallback), placeholder table using engine placeholder paths only. - Extended the architecture guard: SS plugins may depend on SouthernSpearCore (and only it); SS002 now also applies to SouthernSpearTeam; presentation headers may not include `Lyra`, `Attribute` or `GameplayEffect` headers. - README.md now opens with a Southern Spear overview and fictional-work disclaimer. ### FILES CHANGED Created: `Plugins/SouthernSpearTeam/**`, `Docs/evidence/G030_*`, `Docs/evidence/R11_dryriver_nav_report.json`. Modified: `Tools/validate_architecture.py`, `SouthernSpear.uproject` (enable SouthernSpearTeam), `README.md`, `Docs/PROJECT_AUDIT.md` (R-11), `Docs/CHANGELOG.md`, `Content/Maps/L_DryRiver_01.umap` (re-saved by nav script). ### TESTING | Test | Command | Exit | Result | Evidence | |---|---|---|---|---| | Architecture guard | `python Tools/validate_architecture.py` | 0 | PASS, 0 violations | `G030_guard_positive.txt` | | Guard negative (scratch copy: Core→Team, Team→LyraGame, Lyra include in presentation header) | same, on copy | 1 | 4 findings: SS001×1, SS002×1, SS003×2 | `G030_guard_negative.txt` | | SouthernSpearEditor build | `Build.bat SouthernSpearEditor Win64 Development -Project=...` | 0 | Succeeded, 0 errors | `G030_tests_keylines.txt` | | Core + Presentation tests | `UnrealEditor-Cmd ... -ExecCmds="Automation RunTests SouthernSpear;Quit"` | 0 | 15/15 Success (Core 9, Presentation 6), 0 failed | `G030_tests_keylines.txt` | | Game Feature check | same run | 0 | ShooterCore, ShooterExplorer, ShooterMaps, ShooterTests, TopDownArena transitioned successfully; log ending state reads `Registered` (not `Active`) in this commandlet run — activation not proven here | `G030_tests_keylines.txt` | | Dry River nav (R-11) | documented command in `MAPS_DRYRIVER.md` | 0 | level loaded, 1 bounds volume, 1 RecastNavMesh, path 2 points. Tile count not reported by script | `R11_dryriver_nav_report.json` | | Dressing data | `python Tools/verify_dressing.py` | 0 | 19/19 | console | | Git LFS | staged-diff pointer check | — | all staged png/fbx/blend/uasset/umap are LFS pointers | — | | Blender Dry River verification | — | — | NOT RUN (no Blender input changed) | — | | Dedicated server | — | — | NOT RUN — R-09 open | — | Presentation tests cover: all four viewer/subject pairs; invalid viewer; invalid subject; spectator and replay without vantage; authorised spectator; missing and incomplete data (no substitution); determinism (100 runs); reflected fields are non-replicated, never ESSLocality/ESSTeamId, and only text/soft-path/presentation-enum kinds; 3 ACR and MAF views carry the same ESSTeamId. ### ASSETS None created or imported. Placeholder table references engine-shipped assets by path only (`SkeletalCube`, `DefaultMaterial`, `DefaultTexture`); no register entry required. ### RISKS - ~~R-11~~ closed. - R-09 (no Server target) and R-10 (dressing collision lost after reload; re-observed: 0/3 sampled, 0/7 fences) remain open. - New **R-12**: the "560 tiles" figure is not in any retained evidence and the nav script does not measure tiles; nav acceptance currently rests on the path check only. ### DEFECTS FOUND 1. **Guard rule SS003 never scanned presentation headers.** `iter_ss_headers` only yielded files whose immediate directory was `Public`, so anything in `Public/Presentation/` was skipped. Found by the negative test: an injected `LyraHealthComponent.h` include passed. Fixed. 2. **Guard SS001 forbade every SS→SS dependency**, contradicting the rule that SS plugins may build on SouthernSpearCore. Fixed. 3. **Game Feature activation — resolved in follow-up.** A `-game` run on `/ShooterMaps/Maps/L_Expanse` (`UnrealEditor-Cmd ... /ShooterMaps/Maps/L_Expanse -game -nullrhi`, stopped by 240 s timeout, exit 124) logs `ShooterCore ... Ending state: Active`; the other four remain `Registered`, which is correct Lyra on-demand behaviour. The Session 003 record already said `Registered`; the "all five became active" claim came from the session hand-off, not the evidence. Evidence `G031_gamefeature_activation_keylines.txt`. 4. Minor, not fixed: Core reports `UnrecognisedViewingTeam` when the *subject* team is an unknown non-None value (`SSTeamIdentityLibrary.cpp`). ### NEXT ACTION Follow-up in this session: stale active docs corrected against ADR-016/017 (ROADMAP status, TEST_PLAN execution table, GDD team/weapon rows, ASSET_REGISTER W-001..W-003/W-101..W-102, ASSET_NAMING_STANDARDS examples, DF/_DA and _SM LOD contradictions, dressing-name deviation noted). **Diagnose and fix R-10** (Dry River dressing collision lost after save/reload): re-observed this session at 0/3 sampled props and 0/7 fences, and it is the last defect in already-built work before new gameplay (SSExp_ObjectiveAssault / A88) begins. --- ## Session 007 — 2026-09-26 — R-10 Closed: the Fences Were Lying Down ### COMPLETED - Diagnosed R-10 with a read-only probe (`Tools/Unreal/probe_dressing_reload.py`): after reload, dressing *does* collide, but a fence post measured 12 cm tall and 140 cm long and a rail 250 cm tall. - Fixed `dress_dryriver.py`: `unreal.Rotator(0.0, yaw, 0.0)` → keyword form; positional order is (roll, pitch, yaw), so yaw had been applied as pitch. Same bug fixed in `build_dryriver_level.py` (player starts: pitch 180 → yaw 180). - Fixed `build_dryriver_nav.py` checks: bare-`HitResult` handling, real vertical ray, actor check now gates on its result, and fence ground trace ignores dressing (it was landing on post tops). - Re-dressed Dry River and re-ran the nav regression. ### FILES CHANGED Modified: `Tools/Unreal/dress_dryriver.py`, `Tools/Unreal/build_dryriver_nav.py`, `Tools/Unreal/build_dryriver_level.py`, `Content/Maps/L_DryRiver_01.umap`, six `Content/Art/Dressing/*.uasset` (re-imported from the unchanged FBX by the dress pass), `Docs/PROJECT_AUDIT.md`, `Docs/MAPS_DRYRIVER.md`. Created: `Tools/Unreal/probe_dressing_reload.py`, `Docs/evidence/R10_*`. ### TESTING | Test | Command | Exit | Result | Evidence | |---|---|---|---|---| | Dress pass | `UnrealEditor-Cmd ... -ExecutePythonScript=Tools/Unreal/dress_dryriver.py` | 0 | ok; 290 removed and re-placed; collide=True | `R10_dryriver_dressing_report.json` | | Nav regression | documented command | 0 | path 2 points; 3/3 actors and 7/7 fences solid after reload; no warnings | `R10_dryriver_nav_report.json` | | Dressing data | `python Tools/verify_dressing.py` | 0 | all checks passed | — | | Player-start orientation in the map | — | — | NOT RUN — script fixed, map starts not regenerated | — | ### ASSETS No new assets. Dressing `.uasset`s re-imported from existing registered FBX sources. ### RISKS ~~R-10~~ closed. New **R-13** (low): Dry River player starts in the saved map were spawned with pitch 180 by the old level script; correct when `build_dryriver_level.py` is next run. ### DEFECTS FOUND 1. Rotator positional-order bug in two level scripts (found by measuring actor bounds). 2. Nav checker treated every hit as a miss, and traced a zero-length ray (found by a control trace). 3. Fence ground trace hit post tops once collision worked (found when 4/7 passed exactly on the runs whose midpoint is a post station). ### NEXT ACTION **Begin SSExp_ObjectiveAssault** — the presentation increment builds and passes, and no defect remains open in built work other than R-09 (engine distribution) and low-risk R-13. --- ## Session 008 — 2026-09-26 — GitHub, the Website, and Objective Assault Running on Dry River ### COMPLETED - **GitHub.** Private repo `theantipopau/SouthernSpear` (source and docs; LFS binaries not pushed, since the repo vendors Epic content). Public repo `theantipopau/southernspear-site` serving the website on GitHub Pages: https://theantipopau.github.io/southernspear-site/ - **Website** (`Site/`), built from `Docs/images/southern-spear-website-design-spec.svg`: supplied logo and header unmodified, spec colour tokens, 1280/1024/900/768/480 breakpoints, menu below 900 px, skip link, visible focus, reduced motion. Status, roadmap and this changelog are rendered from the repo's own Markdown. No invented release, download, community or donation link. Published by `python Tools/publish_site.py`. - **README** rewritten as the Southern Spear project README. - **SouthernSpearObjectives** plugin (ADR-018): pure `FSSObjectiveRules` (capture: presence not strength, contested freeze, neutralise-then-capture, decay; round: pre-round → sequential objectives → win on final capture / draw on time → post-round → reset), replicated `ASSObjectiveActor` and `ASSObjectiveAssaultDirector`, team read through `IGenericTeamAgentInterface` (no Lyra dependency). Editor-only `SouthernSpearObjectivesEditor` helper, because `FGameFeatureComponentEntry` is not exposed to Python. - **SSExp_ObjectiveAssault** Game Feature + `B_SS_ObjectiveAssault` experience, built by `Tools/Unreal/setup_objective_assault.py`, which also wires Dry River (2 objectives, director, default experience). CI runs it. - **Dry River player starts fixed (R-13).** They were at ±85 **cm** (metres passed as cm) and pitched 180°. Now at ±85 m, facing the centre. Level, dressing, objectives and nav regenerated. - Architecture guard: SS002 (no Lyra) now covers SouthernSpearObjectives. ### FILES CHANGED Created: `Plugins/SouthernSpearObjectives/**`, `Plugins/GameFeatures/SSExp_ObjectiveAssault/**`, `Tools/Unreal/setup_objective_assault.py`, `Tools/publish_site.py`, `Site/**`, `.claude/launch.json`, `Docs/evidence/G040_*`. Modified: `README.md`, `SouthernSpear.uproject`, `Tools/validate_architecture.py`, `Tools/Unreal/build_dryriver_level.py`, `.github/workflows/build.yml`, `Content/Maps/L_DryRiver_01.umap`, regenerated Dry River/dressing `.uasset`s, `Docs/DECISION_LOG.md` (ADR-018), `Docs/PROJECT_AUDIT.md`, `Docs/MAPS_DRYRIVER.md`, `Docs/TEST_PLAN.md`, `Docs/ASSET_REGISTER.md`, `Docs/DEVELOPMENT_ROADMAP.md`. ### TESTING | Test | Command | Exit | Result | Evidence | |---|---|---|---|---| | Architecture guard | `python Tools/validate_architecture.py` | 0 | PASS | `G040_guard_positive.txt` | | Guard negative (Objectives → Team + LyraGame, scratch copy) | same | 1 | SS001 + SS002 caught | `G040_guard_negative.txt` | | Editor build | `Build.bat SouthernSpearEditor Win64 Development` | 0 | Succeeded | `G040_tests_keylines.txt` | | All tests | `Automation RunTests SouthernSpear` | 0 | **26/26** (Core 9, Presentation 6, Objectives 11 incl. a real-world round with overlap-based presence) | `G040_tests_keylines.txt` | | Level → dress → setup → nav chain | the four `-ExecutePythonScript` commands | 0 each | setup ok (8/8 steps); nav path 2 points, 3/3 actors and 7/7 fences solid | `G040_objective_assault_setup.json`, `G040_dryriver_nav_report.json` | | Live game | `UnrealEditor-Cmd ... /Game/Maps/L_DryRiver_01 -game -nullrhi` (stopped by 200 s timeout, exit 124) | 124 | experience identified from WorldSettings; SSExp_ObjectiveAssault and ShooterCore Active; round 1 PreRound → InProgress, OBJ A active; bots spawned | `G040_dryriver_objective_assault_game_keylines.txt` | | Dressing data | `python Tools/verify_dressing.py` | 0 | all checks passed | — | | Site | local preview + JS checks at 1280 and 375 px | — | 7 sessions rendered, no horizontal overflow, menu at 375 px | — | | A capture in a live match | — | — | **NOT RUN** — Lyra bots do not seek objectives; capture is proven by the world test only | — | | Multiplayer replication of round state | — | — | **NOT RUN** — needs two clients (R-05) and a server (R-09) | — | ### ASSETS Two original data assets (register M-001f, M-001g). No art imported. Brand PNGs copied unmodified to the public site repo. ### RISKS ~~R-13~~ closed. Open: R-09 (no Server target), R-12 (tile count unmeasured). New **R-14**: private repo holds LFS pointers only, so the GitHub copy is not a full backup of binary assets. New **R-15**: the level pass (`build_dryriver_level.py`) still reports its own pass-1 path check as failed (`ok=False`); the nav pass is the authoritative check and passes. ### DEFECTS FOUND 1. Player starts placed in metres-as-centimetres, found by dumping actor transforms while wiring the experience. 2. `World->BeginPlay()` does not dispatch actor BeginPlay without a GameMode; test worlds now use `WorldSettings->NotifyBeginPlay()`. Found by the first world-test run. 3. `LyraWorldSettings.DefaultGameplayExperience` is EditDefaultsOnly and refuses Python; set via an editor helper. `FGameFeatureComponentEntry` is not Python-exposed; same fix. 4. A dump script held a map reference across `load_map`, causing an editor "World Memory Leaks" fatal. ### NEXT ACTION **Make bots play the objective** (a ShooterCore-compatible bot behaviour that moves to the active `ASSObjectiveActor`), so a full round can be won in a live `-game` run rather than only in the world test. --- ## Session 009 — 2026-09-26 — Bots Steered to Objectives, Respawn Granted (unverified live) ### COMPLETED - `ASSObjectiveAssaultDirector` steers idle AI bots (Lyra blackboard `TargetEnemy` empty) to the active objective every 2 s; players never touched. - `setup_objective_assault.py` grants ShooterCore `AbilitySet_Elimination` (GA_AutoRespawn) to LyraPlayerState via the text-property helper. ### TESTING | Test | Exit | Result | |---|---|---| | Editor build | 0 | Succeeded | | setup_objective_assault.py | 0 | ok, 10/10 steps incl. respawn_abilities | | build_dryriver_nav.py | 0 | path verified, dressing solid | | Automation tests after steering change | — | **NOT RUN** (usage limit) | | Live `-game` round with `?NumBots=6` | — | **NOT RUN** (usage limit) | ### RISKS Steering untested live; could conflict with the bot behaviour tree. ### NEXT ACTION **Run the test suite, then a live `-game` run on L_DryRiver_01?NumBots=6 (~8 min) and confirm a round is won by capture.** --- ## Session 010 — 2026-09-26 — Live Objective Assault Verified Over 22 Rounds ### COMPLETED - Reviewed the producer's live run (`/Game/Maps/L_DryRiver_01?NumBots=6 -game -nullrhi`, ~2 h). - Added (not yet verified) 7 extra ground-snapped player starts per deployment in `setup_objective_assault.py`; blocked from saving while the producer's game process held the assets. ### TESTING | Test | Result | Evidence | |---|---|---| | Live round lifecycle, 6 bots | **PASS**: 22 rounds; wins for both teams by sequential capture, a split round (TeamTwo took A, TeamOne took B and won), timeout draws, every round reset; idle-bot steering active; no fatal errors, only engine Toolset Python import noise | `Docs/evidence/G041_live_round_keylines.txt` | | Extra player starts | **NOT RUN** — asset save blocked by a running game process | — | | Automation suite after bot steering | **NOT RUN** | — | ### RISKS New **R-16**: win streaks (up to 8 in a row for one team, rounds ~1.5 min) suggest the winning team respawns close to OBJ B. Needs spawn/objective distance review once multiple starts exist. ### NEXT ACTION **Close the running game, run setup_objective_assault.py + nav, and verify the extra starts.** --- ## Session 011 — 2026-09-26 — CLAUDE.md Added ### COMPLETED - Added `CLAUDE.md`: project identity, session start checklist, source-of-truth order, fictional content and legal rules, module architecture and guard, build/test commands, Dry River pipeline order, Unreal Python gotchas found in Sessions 006–010, Git/LFS/publishing rules, and the mandatory end-of-session changelog format. ### FILES CHANGED Created `CLAUDE.md`. Modified `Docs/CHANGELOG.md`. ### TESTING Documentation only. No build or test run: **NOT RUN** (not applicable). ### ASSETS None. ### RISKS Unchanged. The extra player starts from Session 010 remain unverified. ### NEXT ACTION **Close the running game process, run `setup_objective_assault.py` + `build_dryriver_nav.py`, and verify the extra player starts.** --- ## Session 012 — 2026-09-26 — Extra Deployment Starts and Round-Reset Respawn ### COMPLETED - Stopped the stale headless `-game` run (PID 26412, from Session 010) that was holding the assets. - **Extra deployment starts verified:** `setup_objective_assault.py` now places 7 ground-snapped starts per team (14 total) around each deployment; nav re-verified afterwards. - **Round-reset respawn** (`ASSObjectiveAssaultDirector::bRespawnAllOnRoundReset`, default on): from round 2, every pawn is destroyed at reset so the game mode respawns it at a deployment start; after 1 s any *player* controller still without a pawn is restarted directly. AI controllers are left to their own respawn path. Engine-only (`AGameModeBase`), no Lyra dependency. Addresses the main cause of R-16: rounds previously began wherever the last one ended. - **Playtest URL overrides:** `?RoundSeconds=`, `?PreRoundSeconds=`, `?PostRoundSeconds=` on the map URL. ### FILES CHANGED Modified: `Plugins/SouthernSpearObjectives/.../SSObjectiveAssaultDirector.{h,cpp}`, `Tools/Unreal/setup_objective_assault.py`, `Content/Maps/L_DryRiver_01.umap`, `Plugins/GameFeatures/SSExp_ObjectiveAssault/Content/SSExp_ObjectiveAssault.uasset` (regenerated), `Docs/PROJECT_AUDIT.md`, `CLAUDE.md`. Created: `Docs/evidence/G050_*`. ### TESTING | Test | Command | Exit | Result | Evidence | |---|---|---|---|---| | Setup pass | `-ExecutePythonScript=.../setup_objective_assault.py` | 0 | ok; 10/10 steps; "14 extra start(s) across 2 deployment(s)" | `G050_objective_assault_setup.json` | | Nav pass | `-ExecutePythonScript=.../build_dryriver_nav.py` | 0 | ok, path 2 points | `G050_dryriver_nav_report.json` | | Live, default rules (8 bots) | `-game ... L_DryRiver_01?NumBots=8`, 420 s timeout | 124 | Round 1 → 4 all won by TeamTwo (both objectives), resets working before respawn existed | — (superseded run) | | Live, respawn first attempt | same, 300 s | 124 | respawned 9 at reset, but **1 ensure** in `LyraPlayerBotController::ServerRestartController` | — | | Live, final | `-game ... L_DryRiver_01?NumBots=8?RoundSeconds=60 -FORCELOGFLUSH`, 400 s | 124 | 4 resets, each "sending 9 player(s) back to deployment", "restarted 0 controller(s) directly"; **0** fatal/assert/ensure | `G050_round_reset_respawn_game.txt` | | All tests | `Automation RunTests SouthernSpear` | 0 | 26/26 | `G050_tests_keylines.txt` | | Guard / dressing | `validate_architecture.py`, `verify_dressing.py` | 0 / 0 | PASS / all passed | — | | Automated test of `RespawnAllPlayers` | — | — | **NOT RUN** — needs a GameMode; covered by the live run only | | Spawn positions after reset | — | — | **NOT RUN** — positions not logged; respawn is proven, "at deployment" relies on Lyra start selection | ### ASSETS None imported. GFD and map regenerated by script. ### RISKS R-16 mitigated (reset respawn), still open until spawn-to-objective distances are measured. With 60 s rounds, all four were draws (OBJ A contested throughout); with default 900 s rounds one run saw 8 min with no capture, so bots can deadlock on a contested objective (noted under R-16). ### DEFECTS FOUND 1. A Python edit silently failed to insert the respawn call (CRLF source); found when the live log lacked the line. 2. Direct `RestartPlayer` for AI double-restarted a bot already mid-respawn (Lyra ensure); found in the live log. 3. `-game` logs lose their tail when killed by timeout; use `-FORCELOGFLUSH`. ### NEXT ACTION **Add an objective HUD** (active objective, capture progress, round timer, score) reading only the replicated director and objective state. --- ## Session 013 — 2026-09-26 — Objective HUD and Dry River Lighting ### COMPLETED - **Objective HUD** in a new UI module `SouthernSpearObjectivesUI` (Runtime, in the Objectives plugin): pure `FSSObjectiveHudModel` (round/clock, "OBJ B Farmstead (2/2) Opposing capturing", own-side-first score; viewer-relative words via the viewer's team, neutral "Team One/Team Two" when the viewer has no team — ADR-017), C++-built `USSObjectiveStatusWidget` (no asset), and `USSObjectiveHudSubsystem`, which adds it for the local player whenever a director exists. Reads replicated state only; changes nothing. - The first version put UMG in the gameplay module; guard rule **SS005 caught it** and the UI was split out. Guard extended: a `UI` may depend on `` (SS001), `*UI` modules are exempt from SS005, and `SouthernSpearObjectivesUI` is on the SS002 no-Lyra list. - `ASSObjectiveAssaultDirector::ToTeamId(FGenericTeamId)` exposes the director's team mapping. - **Dry River lighting**: `Tools/Unreal/light_dryriver.py` (movable sun, sky atmosphere, real-time sky light, height fog, unbound post-process with bounded exposure). The map had no lights and rendered black. Added to CI and the pipeline docs (level → dress → **light** → setup → nav). ### FILES CHANGED Created: `Plugins/SouthernSpearObjectives/Source/SouthernSpearObjectivesUI/**` (Build.cs, HUD model, status widget, HUD subsystem, 3 tests), `Tools/Unreal/light_dryriver.py`, `Docs/evidence/G051_*`. Modified: `SouthernSpearObjectives.uplugin`, `SSObjectiveAssaultDirector.{h,cpp}`, `Tools/validate_architecture.py`, `.github/workflows/build.yml`, `Content/Maps/L_DryRiver_01.umap`, `CLAUDE.md`, `Docs/MAPS_DRYRIVER.md`. ### TESTING | Test | Command | Exit | Result | Evidence | |---|---|---|---|---| | Guard (UMG in gameplay module) | `python Tools/validate_architecture.py` | 1 | SS005 flagged `SouthernSpearObjectives` → UMG (real catch, fixed) | — | | Guard after split | same | 0 | PASS | — | | Editor build | `Build.bat SouthernSpearEditor Win64 Development` | 0 | Succeeded | — | | All tests | `Automation RunTests SouthernSpear` | 0 | **29/29** (new `Objectives.Hud.ViewerRelative`, `.NoViewerTeamIsNeutral`, `.Phases`) | `G051_tests_keylines.txt` | | Lighting pass | `-ExecutePythonScript=.../light_dryriver.py` | 0 | ok, 5/5 actors, map saved | `G051_dryriver_lighting_report.json` | | Nav after lighting | `build_dryriver_nav.py` | 0 | ok, path 2 points | — | | Live `-nullrhi` (6 bots, 60 s rounds) | `-game ... -FORCELOGFLUSH`, 200 s | 124 | "Objective status widget shown for LyraPlayerController_0"; resets respawn 7; 0 fatal/assert/ensure | — | | Rendered check | `UnrealEditor.exe ... -game -windowed 1280x720`, screen capture at 70 s | — | lit map; HUD shows round clock, active objective, capture bar (opposing red), score | `G051_hud_lit_dryriver.png` | | HUD replication with a remote client | — | — | **NOT RUN** (R-05/R-09) | — | ### ASSETS None imported. Map re-saved by the lighting and nav passes. Screenshot evidence contains Lyra placeholder content (private repo, LFS pointer only). ### RISKS No new risks. Open: R-09, R-12, R-14, R-15, R-16. The map is untextured blockout grey; Lyra placeholder characters and weapons remain (the art-source decision is still the producer's). ### DEFECTS FOUND 1. UI dependency in the gameplay module — found by guard SS005 before commit. 2. Dry River had no lighting — found by the first rendered capture (black scene). 3. On-screen `WeaponAudioFunctions.EarlyReflections` errors — caused by `-nosound`; gone with sound on. Not a defect. ### NEXT ACTION **Viewer-relative team tint**: colour the viewer's own team blue and the opposing team red on characters (currently Lyra's fixed per-team colours, so the local player can be red), via SouthernSpearTeam presentation. --- ## Session 014 — 2026-09-26 — HUD Restyled to the Southern Spear Palette ### COMPLETED - `SSUIStyle.h`: the website design tokens (`Site/styles.css`) mirrored in C++ (ink/field greens, line, brass, sand, sage, OPFOR clay). Friendly = sage, opposing = OPFOR clay, contested = brass, neutral = sage-400. - Objective status widget rebuilt: field-dark plate with a brass top edge, ROUND n label (brass, tracked caps), phase + large clock (turns clay under 30 s), A/B objective chips (owner-tinted, active outlined in brass), objective name and viewer-relative status, a flat two-segment capture bar, own-side-first score. - `FSSObjectiveHudModel` gained structured fields (round label, phase, clock, name, status, per-side words/scores/tones) and `BuildChips`. Spectators get neutral colours on both sides (ADR-017). ### FILES CHANGED Created: `SouthernSpearObjectivesUI/Public/SSUIStyle.h`, `Docs/evidence/G052_*`. Modified: `SSObjectiveHudModel.{h,cpp}`, `SSObjectiveStatusWidget.{h,cpp}`, `Tests/SSObjectiveHudTests.cpp`. ### TESTING | Test | Command | Exit | Result | Evidence | |---|---|---|---|---| | Guard | `python Tools/validate_architecture.py` | 0 | PASS | — | | Build | `Build.bat SouthernSpearEditor Win64 Development` | 0 | Succeeded (one fix: `Clock` field shadowed a helper) | — | | All tests | `Automation RunTests SouthernSpear` | 0 | **30/30** (new `Objectives.Hud.Chips`) | `G052_tests_keylines.txt` | | Rendered | windowed `-game`, game window captured at 80 s | — | palette HUD as described; chips, bar and score correct | `G052_hud_palette.png` | ### ASSETS None. System font (Roboto) with letter spacing stands in for Barlow Condensed (would need an OFL font import and a licence-register entry). ### RISKS No new risks. Lyra's own HUD (health, ammo, weapon slots) is still Lyra's cyan style; restyling it means replacing the experience's HUD layout, which needs Lyra attribute access from a UI module (same bridge decision as the team tint). ### DEFECTS FOUND 1. The first capture attempt grabbed the desktop because the game window was behind another app; the capture now focuses the game window and captures only its rectangle. The stray image was deleted. ### NEXT ACTION **Producer decision: Lyra bridge (modify Lyra's team display function, or add a SouthernSpearLyraBridge module)**. It unblocks both the viewer-relative team tint and a Southern Spear health/ammo HUD. --- ## Session 015 — 2026-09-26 — Lyra Bridge and Viewer-Relative Team Tint ### COMPLETED - Producer decisions recorded: **ADR-019** (SouthernSpearLyraBridge is the one guarded Lyra dependency) and **ADR-020** (original art, script-built in Blender). - New plugin `SouthernSpearLyraBridge` with `USSViewerTeamTintSubsystem`: on each client, every pawn is tinted from the viewer's side (own team sage `#B9C1B4`, other team OPFOR clay `#8C493D`, glows from the palette) through `FSSTeamIdentity::ResolveLocality`. Spectators keep Lyra's absolute colours. - Guard: SS002 now bars Lyra from **every** `SouthernSpear*` module except the bridge. - Dev tool: `-SSShotAt=` takes an in-engine viewport screenshot, replacing desktop capture. ### FILES CHANGED Created: `Plugins/SouthernSpearLyraBridge/**`, `Docs/evidence/G053_viewer_relative_tint.png`. Modified: `SouthernSpear.uproject`, `Tools/validate_architecture.py`, `SSObjectiveHudSubsystem.{h,cpp}`, `Docs/DECISION_LOG.md`, `CLAUDE.md`. ### TESTING | Test | Command | Exit | Result | Evidence | |---|---|---|---|---| | Guard | `python Tools/validate_architecture.py` | 0 | PASS | — | | Guard negative (Team → LyraGame + LyraBridge, scratch copy) | same | 1 | SS001 + SS002 | — | | Build | `Build.bat SouthernSpearEditor ...` | 0 | Succeeded | — | | All tests | `Automation RunTests SouthernSpear` | 0 | 30/30 | — | | Live `-nullrhi`, 6 bots | 120 s | 124 | "Viewer-relative tint applied to 7 pawn(s) (viewer TeamOne)"; 0 fatal/assert/ensure | — | | Rendered | windowed `-game -SSShotAt=45` | 124 | own character sage (was Lyra red), opposing name tags clay; HUD intact | `G053_viewer_relative_tint.png` | | Tint seen by a second (TeamTwo) client | — | — | **NOT RUN** (R-05/R-09) | — | ### ASSETS None. ### RISKS No new IDs. Two desktop screen captures grabbed other windows instead of the game; both were deleted, never committed, and the practice was replaced by `-SSShotAt`. ### DEFECTS FOUND 1. `ULyraTeamDisplayAsset` is not exported from LyraGame (link-time); the bridge sets the colour parameters directly. 2. Focus-stealing capture opened Lyra's pause menu; the in-engine screenshot avoids it. ### NEXT ACTION **A88 rifle, first pass**: a script-built Blender model (`Tools/Blender/a88_rifle.py`), imported and equipped through a Southern Spear weapon definition that reuses Lyra's rifle behaviour and animations. --- ## Session 016 — 2026-09-26 — Sourced Asset Review and the A88 in Game ### COMPLETED - **`Content/Sourced/` review** (about 2.4 GB, 286 files, added by the producer). No licences anywhere. It includes a *Call of Duty: MWIII* rip and a *Counter-Strike 2* rip, real-weapon replicas (AUG, HK416, AKM, M4, Minimi), an "insurgent" character (contradicts ADR-016), and scraped textures. Verdicts are in `Docs/SOURCED_ASSET_REVIEW.md`. The folder is git-ignored and **nothing was imported**. New risk R-17. - **A88 rifle, first pass (ADR-020)**: - `Tools/Blender/a88_rifle.py` builds an original angular bullpup (89 cm, 2,632 tris, Polymer/Metal/Glass slots, muzzle socket). - `Tools/Unreal/setup_a88.py` imports it, adds the flat PBR master `M_SS_FlatPBR` and palette instances, and creates `B_SS_A88` on the new `ASSHeldItemVisualActor` (SouthernSpearTeam, cosmetic only). - It also creates `WID_SS_A88`/`ID_SS_A88` as copies of Lyra's rifle definitions, pointed at our visual. - **Starting loadout** (`USSLoadoutSubsystem` + `USSLoadoutSettings` in the bridge): each pawn receives the configured items and the first becomes active, the same for both teams. Lyra's unexported quick bar is driven through UFUNCTION reflection. `ID_SS_A88` is configured in `DefaultGame.ini`. - Editor helper `GetPropertyAsText` added. ### FILES CHANGED Created: `Tools/Blender/a88_rifle.py`, `Tools/Unreal/setup_a88.py`, `Art/Weapons/A88/{SM_A88.fbx,A88.blend}`, `SSHeldItemVisualActor.{h,cpp}`, `SSLoadoutSubsystem.{h,cpp}`, `/SSExp_ObjectiveAssault/Weapons/A88/*`, `/SSExp_ObjectiveAssault/Materials/M_SS_FlatPBR`, `Docs/SOURCED_ASSET_REVIEW.md`, `Docs/evidence/G054_*`. Modified: `.gitignore`, `Config/DefaultGame.ini`, `SouthernSpearLyraBridge.Build.cs`, the editor library, `Docs/ASSET_REGISTER.md`, `Docs/PROJECT_AUDIT.md` (R-17), `CLAUDE.md`. ### TESTING | Test | Command | Exit | Result | Evidence | |---|---|---|---|---| | Blender build | `blender --background --factory-startup --python Tools/Blender/a88_rifle.py` | 0 | ok, 0.89 × 0.084 × 0.313 m, 2,632 tris | `G054_a88_side.png`, `G054_a88_threequarter.png` | | Unreal setup | `-ExecutePythonScript=.../setup_a88.py` | 0 | ok, 6/6 steps; mesh 89 cm; muzzle socket found; 1 equippable fragment re-pointed | `G054_a88_setup.json` | | Guard / build / tests | as usual | 0/0/0 | PASS / Succeeded / 30/30 | — | | Rendered | windowed `-game -SSShotAt=35`, 6 bots | 124 | A88 active in slot 2 (30/60), held correctly (rail and optic up, forward), bots carry it (kill feed); no Blueprint runtime errors | `G054_a88_in_game.png` | | Muzzle flash / shell FX on the A88 | — | — | **NOT CHECKED** — Lyra's fire cue may expect `B_Weapon`; next pass | — | ### ASSETS W-A88-01 (Class F, original). Nothing from `Content/Sourced/`. ### RISKS New **R-17** (unlicensed drop in `Content/Sourced/`). The environment scans also conflict with ADR-013. ### DEFECTS FOUND 1. `StaticMesh.sockets` is protected in Python; `find_socket` is used instead. 2. The Lyra item-fragment property is invisible to Python; set via the text helper. 3. The local `CallFunction` helper collided with `UObject::CallFunction` (compile error), so it was renamed. ### NEXT ACTION **A88 fire effects and first-person fit**: verify the muzzle flash, tracers and shell ejection on `B_SS_A88` (add the sockets and interface Lyra's fire cue expects), and tune `MeshOffset` so the hands sit on the grip and foregrip. --- ## Session 017 — 2026-09-27 — Licensed-Art Policy, Sketchfab Clearance, A89 > The original entry reported the three cleared Sketchfab assets as L-0012 to L-0014. The current register uses L-0012 for Electric Dreams (declined), L-0013 Split Point, L-0014 Bingie Bingie and L-0015 Ross River; this numbering correction is recorded in the current rows, below. The initial asset drop count is historical; see `SOURCED_ASSET_REVIEW.md` for the later measured inventory. ### COMPLETED - **ADFRC addon drop reviewed** (`Content/Sourced/ADF`, 1,681 files): no models (only `.paa`, `.rvmat`, configs and sounds), APL-SA/Arma-only licence, other authors, ADF-branded. Not usable; recorded in `SOURCED_ASSET_REVIEW.md`. - **ADR-021** (producer decision): free/purchased Fab (Standard License) and CC0/CC-BY Sketchfab assets are allowed, registered before use and adapted to the fiction. Weapons stay original A-series; Dry River's layout stays original. - **Sketchfab clearance** via the public API: - Cleared, CC-BY: Split Point VIC, Bingie Bingie NSW, Ross River NT (the original L-0012 to L-0014 shorthand was off by one; authoritative entries are now L-0013 to L-0015). - Rejected: Cape Liptrap (CC-BY-NC-ND). - Still needing page URLs: Petty Beach and the gear models. - **A89 light support weapon**, original design (1.19 m, 2,012 tris): shared helpers in `Tools/Blender/ss_weapon_kit.py`. The Unreal import is generalised to `Tools/Unreal/setup_weapons.py` (`WEAPONS = ["A88", "A89"]`). - Fab "Add to Project" help for the producer. The launcher doesn't list this project, so use the in-editor Fab plugin. ### FILES CHANGED Created: `Tools/Blender/ss_weapon_kit.py`, `Tools/Blender/a89_support.py`, `Art/Weapons/A89/*`, `/SSExp_ObjectiveAssault/Weapons/A89/*`, `Docs/evidence/G055_*`. Renamed: `Tools/Unreal/setup_a88.py` → `setup_weapons.py`. Modified: `Docs/DECISION_LOG.md` (ADR-021), `Docs/LICENCE_REGISTER.md`, `Docs/SOURCED_ASSET_REVIEW.md`, `CLAUDE.md`. ### TESTING | Test | Command | Exit | Result | Evidence | |---|---|---|---|---| | Sketchfab licence lookup | `curl https://api.sketchfab.com/v3/search?...` | 0 | licences as listed above | `SOURCED_ASSET_REVIEW.md` | | A89 build | `blender ... a89_support.py` | 0 | ok, 1.19 × 0.10 × 0.28 m, 2,012 tris | `G055_a89_*.png` | | Weapons setup | `-ExecutePythonScript=.../setup_weapons.py` | 0 | ok, 12/12 steps (A88 + A89), both muzzle sockets | `G055_weapons_setup.json` | | A89 in game | — | — | **NOT RUN** — not in the loadout yet (a support role needs role or kit selection first) | — | ### ASSETS W-A89-01 (Class F, original). L-0013 to L-0015 cleared (CC-BY), not yet imported. ### RISKS R-17 is narrowed: three scans cleared, one rejected; the rest await URLs. ### DEFECTS FOUND 1. The finishes check assumed every weapon has a Glass slot; it now compares against the mesh's own slots. ### NEXT ACTION **Kit selection**: a simple per-player kit choice (Rifleman: A88; Support: A89) in the loadout settings, so the A89 is playable and the A88 fire-effects pass can be verified on both weapons. --- ## Session 018 — 2026-09-27 — First-Person Camera; Fab Packs Arrive ### COMPLETED - **First-person camera** (the game is an FPS; Lyra's ShooterCore is third-person). In `SouthernSpearLyraBridge`: - `USSFirstPersonCameraMode` puts the view at the `head` bone plus an offset, following control rotation (FOV 90). - `USSFirstPersonADSCameraMode` gives aim-down-sights at FOV 60. - `USSFirstPersonSubsystem` wraps Lyra's camera-mode delegate for the local player: Lyra's ADS choice maps to the SS ADS mode, everything else to SS first person. It also hides the local player's own head. - Fab project listing fixed for the producer with a directory junction in `OneDrive/Documents/Unreal Projects`. - Fab packs are now in `Content/`: AK-47, FPS_Weapon_Bundle, Insurgent_2, Modern_Insurgent_7, QuantumCharacter, M1911, RuralAustralia, and `Downloaded/`. All raw packs are git-ignored until adapted (ADR-021). ### FILES CHANGED Created: `SSFirstPersonCameraMode.{h,cpp}`, `SSFirstPersonSubsystem.{h,cpp}`, `Docs/evidence/G056_first_person.png`. Modified: `.gitignore`, `CLAUDE.md`. ### TESTING | Test | Command | Exit | Result | Evidence | |---|---|---|---|---| | Build | `Build.bat SouthernSpearEditor ...` | 0 | Succeeded | — | | Rendered | windowed `-game -SSShotAt=40` (a first attempt timed out while shaders compiled) | 124 | "First-person camera active"; eye-level view, HUD/ammo intact; **no arms or weapon in view** | `G056_first_person.png` | | ADS in first person | — | — | **NOT RUN** (needs input) | — | | Automation tests | — | — | **NOT RUN** this step (no test-covered code changed) | — | ### ASSETS Fab packs present, not yet reviewed or registered. ### RISKS New **R-18**: first person has no view model (arms and weapon) yet. ### DEFECTS FOUND 1. **Process defect:** commit `88627674` staged with `git add -A` and swept in two packs that arrived mid-commit (M1911: 31 files, RuralAustralia: 436), before they were ignored. Only LFS pointers were pushed, because pushes skip LFS uploads, so no asset data left the machine. Fixed in the next commit (untracked and ignored). From now on, commits stage explicit paths only (CLAUDE.md). 2. The changelog script failed on a `\U` escape in a Windows path; changelogs are now written from a file. ### NEXT ACTION **First-person view model**: review FPS_Weapon_Bundle for first-person arms and animations, then show arms plus the equipped A-series weapon to the local player only. --- ## Session 019 — 2026-09-27 — Red Gum Station Playable; 3 ACR / MAF Soldier Bodies ### COMPLETED - **Red Gum Station (`L_RedGum_01`)**, the first map built on the Fab Rural Australia pack (ADR-022). It is built headless: map copy, 16 deployment starts, 3 objectives (Bore Pump, Homestead, Shearing Shed), the director, nav bounds and the SS experience. - Fences had split the nav mesh into three islands (diagnosed with a 10 m reachability grid). The fence meshes now have no collision. - The nav pass now fails unless every route leg connects. - **Soldier bodies.** `ASSCharacterPartActor` (Team) plus the `ISSLocalityPresentable` interface (Core), with the bridge tint subsystem applying viewer locality. - The viewer's own team shows the Quantum military character (3 ACR); the other team shows conventional Modern Insurgent 7 parts (MAF). - Meshes follow the mannequin's pose. All three packs use UE-mannequin bone names, so no retarget was needed. - The body is hidden from its own first-person view. - `B_SS_CharacterParts` replaces Lyra's random Manny/Quinn picker in the Game Feature. - Fab packs registered (L-0016); ADR-022; Dry River nav rebuilt after the Game Feature re-wire. - No more windowed runs on unwired maps; the blank Example_01 run was only compiling shaders. ### FILES CHANGED Created: - `SSLocalityPresentable.h`, `SSCharacterPartActor.{h,cpp}`; - `Tools/Unreal/build_redgum_level.py`, `build_redgum_nav.py`, `setup_soldiers.py`; - `Content/Maps/L_RedGum_01.umap`, `SSExp_ObjectiveAssault/Characters/*`; - `Docs/evidence/G057_redgum_live.txt`. Modified: - `SSViewerTeamTintSubsystem.cpp`, `setup_objective_assault.py`; - the Game Feature data, `L_DryRiver_01.umap` (script re-save plus nav rebuild); - DECISION_LOG, LICENCE_REGISTER, ASSET_REGISTER, PROJECT_AUDIT. ### TESTING | Test | Command | Exit | Result | Evidence | |---|---|---|---|---| | Guard | `python Tools/validate_architecture.py` | 0 | PASS | — | | Build | `Build.bat SouthernSpearEditor ...` | 0 | Succeeded | — | | Automation | `Automation RunTests SouthernSpear` | 0 | 30 Success, 0 Fail | `Build/tests.log` (not retained) | | Red Gum build | `build_redgum_level.py`, `build_redgum_nav.py` | 0 | ok=true; all 4 legs connect (deploy A to deploy B) | `Build/redgum_*.json` | | Live, 8 bots | `L_RedGum_01?NumBots=8?RoundSeconds=90 -game -nullrhi` (killed at 420 s, as intended) | 124 | 3 rounds; captures by both teams; round resets | `G057_redgum_live.txt` | | Soldier parts, live | same, 60 s rounds | 124 | Tint and locality applied to 9 pawns; no SS errors. 288 Lyra `GCNL_Spawning` "Accessed None" warnings (spawn effect has no Niagara system under null RHI) | — | | Dry River setup and nav | `setup_objective_assault.py`, `build_dryriver_nav.py` | 0 | ok=true | `Build/*.json` | | Rendered soldier check | windowed `-SSShotAt=150` | 0 | **NOT RUN**: the window was closed at about 150 s (Windows exit request) before the capture | — | ### ASSETS L-0016 Fab packs (class A); M-RG-01 and C-SOL-01 in the asset register. ### RISKS New R-19 (Fab dependency of the map and bodies) and R-20 (soldier look not yet visually checked). R-18 logged in the audit. ### DEFECTS FOUND 1. `duplicate_asset` on a map kept the world referenced, causing a fatal "World Memory Leaks" (found by the crash log). Fixed: load the source, then save-as. 2. `save_map` fails when the target exists (found in the step report). Fixed: delete the old copy first. 3. Per-instance fence collision edits do not persist, because the construction script reruns on load (found with the island grid). Fixed at the mesh level. ### NEXT ACTION **Rendered soldier check** on Red Gum (one windowed `-SSShotAt` run, about 3 minutes, shaders are cached after the first run). Confirm the 3 ACR and MAF looks, hide any insignia patch (R-20), then start the first-person view model (R-18). --- ## Session 020 — 2026-09-27 — CC BY Terrain Scans Staged for Import Review ### COMPLETED - Audited `Content/Sourced/` against `Docs/SOURCED_ASSET_REVIEW.md` and the licence register. - Verified the official Sketchfab model pages: Split Point and Bingie Bingie are licensed CC BY 4.0 and authored by Stefan A Vollgger. Staged each original OBJ/MTL/JPEG set unchanged under `Content/SouthernSpear/Vendor/SAVollgger/`. - Added ENV-001 and ENV-002 to `ASSET_REGISTER.md`, completed L-0013/L-0014 provenance rows, and put required credits beside the vendor files. - Left Ross River unstaged. All other sourced files were left untouched and unapproved. ### FILES CHANGED Created: `Content/SouthernSpear/Vendor/SAVollgger/README.md` and two vendor source folders, each with its unchanged OBJ, MTL and JPEG files. Modified: `Docs/ASSET_REGISTER.md`, `Docs/LICENCE_REGISTER.md`, `Docs/SOURCED_ASSET_REVIEW.md`, `Docs/CHANGELOG.md`. ### TESTING | Test | Command | Result | |---|---|---| | Sketchfab provenance/licence | Opened official model pages | PASS — correct titles, authors and CC BY licence; register records the 4.0 version | | Source archives | `unzip -t` on both archives | PASS — no archive errors | | Source-file identity | SHA-256 of each staged file vs matching archive member | PASS — all six OBJ/MTL/JPEG hashes match exactly | | MTL image paths | Python check against staged filenames | PASS — each MTL image reference resolves | | Mesh inspection | Python OBJ scan | PASS — Split Point 174,076 vertices / 346,200 triangles; Bingie Bingie 273,042 vertices / 540,708 triangles; all faces triangular | | Git LFS attributes | `git check-attr filter diff merge text` | PASS — OBJ and JPEG files resolve to LFS | | Unreal Editor import | — | **NOT RUN** — no `.uasset` import or editor validation yet | | Optimization/derivative | — | **NOT DONE** — original scan files retained; not game-ready | | Game/map use | — | **NOT DONE** — neither scan placed in a level | ### ASSETS Two class-B third-party source sets staged unchanged: ENV-001 Split Point and ENV-002 Bingie Bingie. Include their registered CC BY 4.0 credits if used in a shipped build. No derived or imported Unreal asset was made. ### RISKS - The remaining sourced drop still includes explicitly rejected game rips and unresolved/licence-blocked assets; all remain untouched and outside Git. - Both scans are dense photogrammetry and may need cleanup, reduction, LOD/Nanite review, collision work and performance validation before real-time use. - Ross River is licence-cleared but not staged; Cape Liptrap remains rejected under CC BY-NC-ND. ### DEFECTS FOUND - Session 017 shorthand used the wrong L-number range after L-0012 was assigned to declined Electric Dreams. Current references now read L-0012 (declined), L-0013 Split Point, L-0014 Bingie Bingie and L-0015 Ross River. - The sourced review's previous “nothing imported” phrasing did not distinguish copied vendor source files from actual Unreal `.uasset` imports; it now does. ### NEXT ACTION Import both OBJ files into a temporary Unreal review folder and inspect scale, texture binding, normals, collision, LOD/Nanite options and editor performance before making any adapted game asset. --- ## Session 021 — 2026-09-27 — Southern Spear UI Pass; Animated Soldiers; First-Person Weapon ### COMPLETED - **Soldier animation fixed.** Lyra attaches parts through a ChildActorComponent, so the part's direct parent was never the animated mesh. `ASSCharacterPartActor` now walks up to the first skinned mesh, and all 9 pawns log "leader CharacterMesh0". - **First-person weapon view model.** `USSFirstPersonSubsystem` puts an owner-only copy of the held weapon's mesh on the camera: - hip and aimed (ADS) placements; - look sway and walk bob; - the body copy of the weapon is hidden from the owner. - Arms are not done yet (R-18 stays open). - **Southern Spear UI replaces Lyra's** (ADR-023): - `SouthernSpearUI` plugin with a player HUD: health bar, ammunition and weapon name, a crosshair that opens with movement and hides when aiming, and a clay flash when hit. - Esc match menu: Resume, Leave Match, Quit. - Title front end `L_SS_FrontEnd`: key art; Red Gum Station and Dry River with bots; Quit. It is now `GameDefaultMap`. - Loading screen: the new key art (`Docs/images/loadingscreen.png`), an animated label and tips. - Lyra's StandardHUD action set is removed from `B_SS_ObjectiveAssault`; `ProjectName` is now Southern Spear. - **Minimap and full map** (`SouthernSpearObjectivesUI`): a north-up minimap centred on the player, and the full map on M. Both use a client-local orthographic scene capture and show objective markers in viewer-relative tones, plus a player arrow. - `USSLocalHudState` (Core) is the plain HUD data. The Lyra bridge (`USSHudStateSubsystem`) fills it from Lyra's health component and quick-bar ammunition stats; UI modules only read it. - **New weapon files reviewed; none imported:** - AKM and PKM: producer-supplied .blend files; real designs, needing A-series reshaping. - C4A1: the file embeds "Cycles-Ready M4 Carbine … Licensed CC-BY" (author file name kkanamalla), so it is third-party and needs its source URL and credit. - A88/New: no licence text; the .mtl texture names do not match the supplied PNGs; 72k faces, 4K maps. Source URL and terms needed. ### FILES CHANGED Created: - `Plugins/SouthernSpearUI/*` (plugin, palette, widget kit, HUD, menu, loading screen, front-end game mode, HUD subsystem); - `SSLocalHudState.h`, `SSHudStateSubsystem.{h,cpp}`, `SSMinimapWidget.{h,cpp}`; - `Tools/Unreal/setup_ui.py`, `Content/Maps/L_SS_FrontEnd.umap`, `Plugins/SouthernSpearUI/Content/Textures/T_SS_KeyArt`, `Docs/images/loadingscreen.png`. Modified: - `SSCharacterPartActor.{h,cpp}`, `SSFirstPersonSubsystem.{h,cpp}`, bridge Build.cs; - `SSObjectiveHudSubsystem.{h,cpp}`, ObjectivesUI Build.cs; - `SouthernSpear.uproject`, `Config/DefaultGame.ini`, `Config/DefaultEngine.ini`; - `setup_objective_assault.py`, `B_SS_ObjectiveAssault`, `CLAUDE.md`, `DECISION_LOG.md`. ### TESTING | Test | Command | Exit | Result | Evidence | |---|---|---|---|---| | Guard | `python Tools/validate_architecture.py` | 0 | PASS | — | | Build | `Build.bat SouthernSpearEditor ...` | 0 | Succeeded (after fixing 2 compile errors) | — | | Automation | `Automation RunTests SouthernSpear` | 0 | 30 Success, 0 Fail | — | | UI content | `setup_ui.py` | 0 | ok=true (key art, experience HUD removed, front-end map) | `Build/ui_setup.json` | | Front end boots | `-game -nullrhi` (no map) | 124 | `LoadMap /Game/Maps/L_SS_FrontEnd`, game class SSFrontEndGameMode, "front end shown" | log | | Red Gum live | `L_RedGum_01?NumBots=8 -game -nullrhi` | 124 | 9 soldier parts, leader CharacterMesh0; SS player HUD shown; view model shows SM_A88; objective widget shown | log | | Rendered look (HUD, menu, loading, minimap, soldiers, view model) | — | — | **NOT RUN** (headless only; placements may need tuning) | — | | Esc / M / menu buttons with real input | — | — | **NOT RUN** | — | ### ASSETS T_SS_KeyArt (producer key art). New weapon source files are not imported and not registered: they await source URLs and licences, and A-series reshaping. ### RISKS R-18 partly mitigated (weapon view model; no arms). R-20 is still open (no rendered check). ### DEFECTS FOUND 1. Soldier bodies were unanimated: leader pose looked only at the direct attach parent (reported by the producer; confirmed by the new "leader" log line). 2. Three compile errors: palette alpha overload, and `FSlateChildSize` fill weights (found by the build). ### NEXT ACTION **Rendered UI check.** One windowed run from the front end into Red Gum, with screenshots of the menu, loading screen, HUD, minimap and view model; then tune placements. --- ## Session 022 — 2026-09-27 — UI Polish Pass, Settings, Round Banners, Textured A88, Website Rundown ### COMPLETED - **Movement fix.** The front end left input in UI-only mode across the map load. Starting a map, and the HUD appearing, now both restore game input. - **Front end.** Intro from black, the title tracking in, staggered reveals, and map cards for Red Gum Station and Dry River. Also a bot count selector (4/8/12), Settings and Quit, a top status bar and a controls strip. - **Match menu (Esc).** Blurred backdrop, a side panel sliding in, Resume, Settings, Leave Match, Quit, and a controls card. - **Settings screen (new).** Window mode, resolution, graphics quality, VSync and frame limit (applied through UGameUserSettings), plus field of view (70–110°; `FSSUserPrefs`, read by the first-person camera; aiming narrows it proportionally). - **In match:** - compass strip with objective markers and the active objective's distance; - round banners: round start, assault, "Objective X secured" in the capturer's viewer-relative tone, and the round result; - flashing reload prompt; ammunition turns clay below a quarter of the magazine. - **A88 now uses the producer-supplied textured model** (`Art/Weapons/A88/New`): - `Tools/Blender/a88_sourced.py` assigns slots by object name; the download's .mtl pointed at missing files, and its black/white texture is a mask, not colour. Pairings were chosen from rendered permutations. - It also normalises to +X, metres and the grip origin, and adds the muzzle socket. - `setup_weapons.py` gains a textured path (`M_SS_TexturedPBR`, `MI_A88_Tex*`, `T_A88_*`). - Result: 78.8 cm, 72k triangles. - **Website:** - full-bleed key-art hero; - "What is Southern Spear?" rundown and pillars; - a mirrored-factions diagram (each team sees itself as 3 ACR, the other as MAF); - "In the build today" with honest status tags; - scroll reveal (respects reduced motion) and a blurred sticky nav; - Discord link (nav, hero, community). - HUD subsystems skip worlds without a game viewport (headless test worlds). ### FILES CHANGED Created: - `SSSettingsWidget.{h,cpp}`, `SSUserPrefs.h`, `SSCompassWidget.{h,cpp}`, `SSRoundBannerWidget.{h,cpp}`; - `Tools/Blender/a88_sourced.py`, `Docs/images/keyart.jpg`; - A88 textured assets, `M_SS_TexturedPBR`. Modified: - `SSMenuWidget.{h,cpp}`, `SSPlayerHudWidget.{h,cpp}`, `SSPlayerHudSubsystem.cpp`, `SSWidgetKit.h`; - `SSObjectiveHudSubsystem.{h,cpp}`, `SSLocalHudState.h`, `SSHudStateSubsystem.cpp`, `SSFirstPersonCameraMode.{h,cpp}`; - `setup_weapons.py`, `publish_site.py`, `Site/*`, `LICENCE_REGISTER.md`. ### TESTING | Test | Command | Exit | Result | Evidence | |---|---|---|---|---| | Guard | `python Tools/validate_architecture.py` | 0 | PASS | — | | Build | `Build.bat SouthernSpearEditor ...` | 0 | Succeeded (first blocked by the parallel session's test file until it compiled) | — | | Weapons | `setup_weapons.py` | 0 | ok=true; A88 78.8×13.1×29.8 cm, muzzle socket present | `Build/weapons_setup.json` | | Automation | `Automation RunTests SouthernSpear` | 255 | 30 Success; 1 Fail: `SouthernSpear.Network.Gameplay.TwoPlayerAuthoritySmoke` (the parallel session's new, uncommitted test). Its ensure `ViewportOverlayWidget.IsValid()` comes from CommonLoadingScreen during the test's map load (stack: test line 87) | `Build/tests.log` | | Red Gum live | `L_RedGum_01?NumBots=8 -game -nullrhi` | 124 | SS HUD shown; view model SM_A88; captures by both teams | log | | Website | `publish_site.py --build-only` + browser preview at 1440×900 and 375×812 | 0 | Hero, rundown, mirror, build sections render; hero reveal fixed for background tabs | — | | Rendered game UI (menu, settings, banners, compass, textured A88) | game launched for the producer | — | **Producer review pending** | — | ### ASSETS L-0017 A88 sourced model: the producer states it is royalty free; source URL and terms pending. The raw download is not committed until they are recorded. The key art has a national-flag shoulder patch on the soldier: producer to confirm it suits the fictional 3 ACR (ADR-016). ### RISKS R-21 (new): the A88 sourced model's licence evidence is missing (producer statement only); it is a real-rifle replica, and ADR-021 requires A-series reshaping. ### DEFECTS FOUND 1. No movement after the front end: UI-only input persisted across travel (reported by the producer). 2. Website hero copy invisible: bottom-of-viewport reveal, and rAF paused in background tabs (browser preview). 3. The website hero copy overlapped the wordmark in the key art (browser preview). ### NEXT ACTION **Producer review of the rendered UI**, then the other session fixes its smoke test's loading-screen ensure so the suite is green. --- ## Session 023 — 2026-09-27 — Weapon Asset Intake and Provenance Review ### COMPLETED - Reviewed the current untracked weapon-source folders against Session 021/022 notes, source-file paths, the asset register, licence register and the tracked A88 game assets. - Corrected the A88 record: the producer-supplied textured mesh is already the imported/in-use cosmetic asset, while the original script-built A88 remains separate source work. The A88 source itself is not tracked; the imported Unreal derivative is tracked, so the missing source URL/terms and real-rifle redesign remain release blockers (R-21 / L-0017). - Recorded the C4A1/M4 file as blocked despite its embedded/reported “Licensed CC-BY” note: no source URL, exact licence version/terms or attribution has been verified; its adjacent texture folder is empty in this checkout. Added R-22 / L-0020. - Recorded the AKM and PKM `.blend` files as provenance/licence-pending, reference-only, and not approved for import or derivative use. Added R-23 / L-0018–L-0019. - Added a weapon-source review section and aligned W-001/W-002 and A88 import/source rows in `ASSET_REGISTER.md`. Updated `PROJECT_AUDIT.md` risk rows and the licence register. No art files, `.uasset`s, or scripts were changed; no smoke-test work was performed. ### FILES CHANGED Modified: `Docs/ASSET_REGISTER.md`, `Docs/LICENCE_REGISTER.md`, `Docs/PROJECT_AUDIT.md`, `Docs/CHANGELOG.md`. ### TESTING | Check | Command / method | Result | |---|---|---| | Intake inventory | Reviewed `Art/Weapons/A88/New`, `AKM`, `C4A1`, `PKM` and their corresponding notes | Completed; no source provenance was added or assumed | | A88 import state | Read existing `Build/weapons_setup.json`, `Build/sm_a88_sourced_report.json`, `Build/a88_setup.json`; checked tracked asset paths with `git ls-files` | Existing report says setup `ok=true`; mesh 78.8 × 13.1 × 29.8 cm, 72,493 triangles; textured import assets are tracked. This is review of existing evidence, not a rerun. | | Source-control/LFS rules | `git status --short --untracked-files=all -- `; `git check-attr filter diff merge text -- ` | Raw A88/AKM/C4A1/PKM files are untracked; sampled OBJ/PNG/Blend sources resolve to LFS. No assets staged or committed. | | Build, import or runtime validation | — | **NOT RUN** — documentation/provenance review only | | Headless authority smoke test | — | **NOT RUN** — explicitly shelved at producer direction | ### ASSETS No asset files created, imported or modified in this session. L-0017 records A88 as Class E until rights evidence is supplied. L-0018–L-0020 record the AKM, PKM and C4A1/M4 sources as blocked; source folders remain local/untracked. No new asset is cleared for use or redistribution. ### RISKS - **R-21** — A88 rights/source provenance and fictional redesign remain unresolved; importantly, its Unreal derivative is already imported/in use despite the raw source being untracked. - **R-22** — C4A1/M4 terms and attribution unverified. - **R-23** — AKM/PKM provenance and rights unverified; real-design files remain reference-only. ### DEFECTS FOUND - Asset records lagged project state: W-001 still described a placeholder, and the A88 Unreal import was not clearly distinguished from its local untracked source and previous original script-built mesh. - The C4A1's embedded “Licensed CC-BY” label could be mistaken for verified rights; exact provenance and terms are missing. ### NEXT ACTION Obtain and verify the source URLs and licence terms for A88/C4A1, and confirm whether AKM/PKM are reference-only; keep all unresolved raw sources out of redistribution and redesign any weapon geometry as original A-series work before release. --- ## Session 024 — 2026-09-27 — ADFRC Extraction Intake Review and Claude Source Guard ### COMPLETED - Inspected the new `Content/Sourced/ADF_Extracted/` tree and representative EF88/M4A5 paths. Unlike the separately reviewed `Content/Sourced/ADF/` folder (whose zero-model count applied only to that folder), this extraction contains `.p3d` models, Arma configs/material definitions, textures, `.rtm` animations and Workshop package data. - Checked the public ADF Re-Cut Workshop notice, ADFRC `LICENSE.md`, `ASSETS_LICENSE.md`, `DEV_LICENSE.md`, `MODEL_CREDITS.md`, and Bohemia's APL-SA text. The stated terms do not clear the extracted models or APL-SA material for this Unreal/commercial project; ADFRC further restricts protected-model extraction/reuse. Corrected the earlier ADF addendum so the contributor agreement is not described as an end-user licence or blanket relicensing path. - Cross-referenced `adfrc_ef88` and `adfrc_m4a5` against Claude's A-series guidance and the independent A88/C4A1 source records. Matching EF88/M4 naming is not provenance or permission, and these files are not the sources recorded at L-0017 or L-0020. - Added a metadata-only quarantine instruction for Claude, a sourced-asset addendum, an explicit non-game-asset note in `ASSET_REGISTER.md`, licence record L-0021 and open risk R-24. No source files were copied, altered, staged or imported; `ADF_Extracted` remains git-ignored. ### FILES CHANGED Modified: `CLAUDE.md`, `Docs/SOURCED_ASSET_REVIEW.md`, `Docs/ASSET_REGISTER.md`, `Docs/LICENCE_REGISTER.md`, `Docs/PROJECT_AUDIT.md`, `Docs/CHANGELOG.md`. ### TESTING | Check | Command / method | Result | |---|---|---| | Intake inventory | Listed the `ADF_Extracted` tree and representative directories; read sample EF88/M4A5 config/model metadata | Completed. Representative inventory only; no total file count or size was measured. No model/texture/animation binary was opened or processed. | | Git ignore | `git check-ignore -v Content/Sourced/ADF_Extracted/Models/ADF_Weapons/adfrc_ef88/ADFRC_EF88.p3d` | Exit 0 — `.gitignore:189:Content/Sourced/` ignores the representative extracted model. | | Rights/source review | Read public Steam Workshop item 2971219389, ADFRC licence/model-licence/developer-agreement/model-credit pages, and Bohemia APL-SA text | Completed; sources and access date recorded in `SOURCED_ASSET_REVIEW.md`. This is project intake triage, not legal advice or item-level clearance. | | Git working tree | `git status --short --branch` | `main`; concurrent pre-existing docs, code, and content changes were present and left untouched. | | Documentation whitespace | `git diff --check -- CLAUDE.md Docs/SOURCED_ASSET_REVIEW.md Docs/ASSET_REGISTER.md Docs/LICENCE_REGISTER.md Docs/PROJECT_AUDIT.md Docs/CHANGELOG.md` | Exit 0 — no whitespace errors. Git emitted only configured CRLF-to-LF notices for existing CRLF documents. | | Unreal import/build/tests | — | **NOT RUN** — documentation/provenance review only; no Unreal assets were created or changed. | | Headless authority smoke test | — | **NOT RUN** — remains shelved per producer direction. | ### ASSETS No game assets created, copied, imported or modified. `Content/Sourced/ADF_Extracted/` remains an ignored quarantine folder, not a licensed vendor source. L-0021 records the collection as blocked (known APL-SA/protected-model incompatibility; individual provenance remains unresolved); no ADFRC content was added to the game asset register as usable art. ### RISKS - **R-24 (new)** — ADFRC extraction includes Arma assets with non-commercial/Arma-only APL-SA terms and additional restrictions for protected models; exact local file provenance and per-model categories are unknown. Keep metadata-only and quarantined. - R-17 remains open for the broader mixed sourced drop. R-21 through R-23 remain unchanged. ### DEFECTS FOUND - No code or content defect was found. The documentation gaps were that Claude's general sourced-content warning did not identify this new ADFRC extraction or its stronger model-specific restrictions; the earlier “0 models” inventory could be overgeneralized beyond the distinct `Content/Sourced/ADF/` folder; and the prior addendum described the developer agreement as an alternative downstream licence. These boundaries and the agreement's actual scope are now explicit. ### NEXT ACTION Obtain exact source-chain evidence and written, file-specific rights-holder permissions for any ADFRC item proposed for use; until then, keep the extraction quarantined and continue only with independently sourced or original A-series assets. --- ## Session 025 — 2026-09-27 — ADFRC Growth, Player-Model Question, and a Broken FBX Conversion ### COMPLETED - Re-inspected `Content/Sourced/ADF_Extracted/` after it grew: it now carries a `README.md`, a `_tools/` converter directory and an empty `Models_FBX/`, and measures 7,928 files / ~17 GB (against 7,748 claimed in its README). - Recorded the extraction's now-documented provenance from its own README: the ADFRC source pack (LFS objects fetched and decoded) **plus the binarised Steam Workshop release** (15 `.pbo` archives unpacked from the local Arma 3 install), which is the only source of its `.p3d` models. That corrects the earlier L-0021 note that the acquisition path was undocumented, and it places the models squarely inside the category ADFRC's `ASSETS_LICENSE.md` and `DEV_LICENSE.md` §2.4 restrict. - Answered the player-model question with measurements: **there are no player/character body models in the tree** — `Workshop/ADF_Units` holds 0 `.p3d` and 13 `.paa`, only binarised config headers that dress vanilla Arma bodies. What exists is 56 player-worn gear meshes (helmets, facewear, NVGs, field dress, a Crye G3 uniform, plate carriers, backpacks, TBAS role vests). - Diagnosed the FBX conversion as **failing on every model**: `Models_FBX/` contains no exported geometry, and `_convert_log.txt` shows the identical error each time — `P3D_Error: Invalid MLOD signature: b'ODOL'`, because the Arma 3 Object Builder addon expects the newer `MLOD` signature. No Blender or Unreal process is running. - Did **not** import, convert or re-export anything. Added the measured inventory, provenance, gear list, conversion failure and a new operational hazard to `Docs/SOURCED_ASSET_REVIEW.md`; corrected L-0021's provenance and added a models row; added **R-25** for ~17 GB of loose assets sitting inside the Unreal content root. ### FILES CHANGED Modified: `Docs/SOURCED_ASSET_REVIEW.md`, `Docs/LICENCE_REGISTER.md`, `Docs/PROJECT_AUDIT.md`, `Docs/CHANGELOG.md`. ### TESTING | Check | Command / method | Result | |---|---|---| | Tree growth | `find`/`ls` over `ADF_Extracted` (top level, extension histogram, file count, `du -sh`) | Top level now `Animations Models Models_FBX README.md Source Textures Workshop _tools`; 7,928 files / 17 GB. Histogram: 2,500 `.paa`, 2,489 `.png`, 854 `.rvmat`, 536 `.p3d`, 330 `.rtm`, 257 `.wss`, 257 `.wav`, 165 `.json`, 37 `.uasset`. | | Provenance | Read `Content/Sourced/ADF_Extracted/README.md` | Two sources documented (source pack + binarised Workshop PBOs); models come only from the PBOs; README carries its own licence warning and asks that the model clause be verified before import. | | Player-model existence | `find Workshop/ADF_Units -iname '*.p3d' -o -iname '*.paa'`; `grep` of `ADF_Units/Modern/CDO/Infantry.hpp` | **0 `.p3d`**, 13 `.paa`. Unit definitions are binarised config headers only; no ADFRC soldier body exists. | | Gear inventory | `find Models/ADF_Gear Models/ADF_Gear_2 -iname '*.p3d'` | 56 player-worn gear meshes listed (helmets, facewear, NVGs, `crye_g3.p3d`, `adfrc_field_dress.p3d`, `JPC_Base`, `Peacekeeper_*`, backpacks, `tbas_T2_*`/`TBAS_T5_*`). | | Conversion state | `cat Models_FBX/_convert_log.txt`; `find Models_FBX -type f \| wc -l` | Log shows 3 attempts, all `FAIL` with `Invalid MLOD signature: b'ODOL'` via `bpy.ops.a3ob.import_p3d`. Only 1 file in the tree (the log) — **no FBX exported**. | | Conversion processes | `tasklist \| grep -iE 'blender\|unreal'` | `none` — no conversion or editor process is running. | | Git working tree | `git status --short --branch` | `main`; the pre-existing and Session 024 documentation changes remain uncommitted and were left in place. | | Unreal import / build / tests | — | **NOT RUN** — nothing was imported or converted; documentation and measurement only. | | Moving the tree out of `Content/` (R-25) | — | **NOT RUN** — deferred to producer direction this session. | ### ASSETS No assets created, imported, converted or modified. `Models_FBX/` remains empty. The 56 gear meshes stay quarantined under L-0021 / R-24 and are **not** cleared for import: the models came from the binarised Workshop release that ADFRC's terms protect, and they additionally carry real manufacturer and service identities (Crye Precision, Ops-Core, PASGT, "Team Wendy") that ADR-016 and L-0007 bar regardless of licence. The project's actual player bodies remain the Fab packs under L-0016 wired through `B_SS_Soldier`. ### RISKS - **R-25 (new)** — ~17 GB of loose source assets, including ~10 GB of PNG and 37 `.uasset` files, sit inside the Unreal content root; the tree's README warns of an auto-import of ~2,484 PNGs on next editor open. Move the tree outside `Content/` before opening the editor. - **R-24** unchanged and now better evidenced: provenance is documented, and it documents extraction from the binarised release. - R-17, R-21 to R-23 unchanged. ### DEFECTS FOUND 1. **The FBX conversion pipeline is broken, not merely incomplete.** `Models_FBX/` was created and three models were attempted, but every one failed with `Invalid MLOD signature: b'ODOL'` and no geometry was produced. Anyone waiting on FBX output from that addon for this content will wait forever; the ODOL models need a different path. Found by reading the conversion log rather than by looking for FBX files. 2. **The request's premise does not hold: there are no player models here.** The ADFRC units pack dresses vanilla Arma bodies and ships no body meshes, so the tree offers gear, not characters. Found by counting `.p3d` under `Workshop/ADF_Units`. 3. **A new operational hazard was introduced by the extraction growing in place** (R-25): 17 GB of source now lives under the Unreal content root, with the tree's own README warning of an auto-import of thousands of textures. Found by measuring the tree against `.gitignore` and the content root. 4. **The extraction's README asserts the assets were provided for this game.** That is recorded as a producer-side claim, not a rights-holder grant; it does not displace APL-SA or the protected-model terms. ### NEXT ACTION Decide the direction for player models: source properly licensed character and gear packs and wire them in, build original CMECU/MAF gear in Blender per ADR-020, or seek a written commercial licence from the ADFRC rights holders — and separately, move `Content/Sourced/ADF_Extracted/` out of `Content/` to close R-25 before the editor is next opened. --- ## Session 026 — 2026-09-27 — Original Australian Uniforms on the Player Models ### COMPLETED - **The ADFRC request was declined, and the alternative was built instead.** The producer asked to use textures and patterns — and helmet models — from `Content/Sourced/ADF_Extracted/`. Those cannot be used: the models were extracted from the binarised Workshop release that ADFRC's `ASSETS_LICENSE.md` and `DEV_LICENSE.md` §2.4 protect (no extraction, no derivatives, no other media), the textures are APL-SA which is **non-commercial and Arma-only**, and the gear additionally carries real manufacturer and service identities (Crye Precision, Ops-Core, PASGT, "Team Wendy") barred by ADR-016 and L-0007. Not even as a tracing reference: ADR-016 requires CMECU to be an original pattern. So the goal was met with original work instead. - **Inventoried what is actually wired in.** 3 ACR is the single Fab mesh `SKM_QuantumCharacter` (14 material slots, including a blue rolled-up shirt and blue jeans); MAF is seven single-slot parts (head, hands, sweater, military pants, shoes, small armour, beret). Both packs ship `.uasset` only. - **Established that material-instance overrides are impossible on the 3 ACR body.** Its materials sample textures directly with no `TextureSampleParameter2D`, so there is no parameter to override. The MAF materials do expose parameters, but a uniform approach was chosen instead. - **Authored original texture sets by script** (`Tools/Textures/make_character_textures.py`): CMECU dry-country camo, MAF red-earth camo, and tan/dark gear fabric, each with base colour, a twill micro-normal and an ORM map — 12 PNGs, 2048² for camo and 1024² for gear. - **Authored four original fabric materials** and applied them as per-slot cosmetic overrides, so the licensed vendor meshes are neither duplicated nor edited. Added `FSSPartMaterialOverride` and the two override arrays to `ASSCharacterPartActor` (SouthernSpearTeam, the cosmetic module — presentation only, ADR-004). - **Fixed a committed build break** that was blocking every editor build: `SSCompassWidget.cpp` and `SSMinimapWidget.cpp` each declare `constexpr int32 MaxMarkers` in an anonymous namespace, which collide when UBA merges them into one unity TU. Renamed to `CompassMaxMarkers` / `MinimapMaxMarkers`. ### FILES CHANGED Created: `Tools/Textures/make_character_textures.py`, `Tools/Common/uv_material_params.py`, `Tools/Unreal/probe_character_materials.py`, `Tools/Unreal/setup_character_textures.py`, `Tools/Unreal/verify_character_materials.py`, `Art/Characters/Textures/*.png` (12), and the imported material/texture assets under `/SSExp_ObjectiveAssault/Characters/`. Modified: `Plugins/SouthernSpearTeam/Source/SouthernSpearTeam/Public/SSCharacterPartActor.h`, `.../Private/SSCharacterPartActor.cpp`, `Plugins/GameFeatures/SSExp_ObjectiveAssault/Content/Characters/B_SS_Soldier.uasset`, `Docs/ASSET_REGISTER.md`, `Docs/LICENCE_REGISTER.md`, `Docs/PROJECT_AUDIT.md`, `Docs/CHANGELOG.md`. **Files I do not own, changed to unblock the build:** `Plugins/SouthernSpearObjectives/Source/SouthernSpearObjectivesUI/Private/SSCompassWidget.cpp` and `SSMinimapWidget.cpp` (local constant rename only). Left uncommitted for the owning session to adopt or revert. ### TESTING | Check | Command / method | Result | |---|---|---| | Texture generation | `python Tools/Textures/make_character_textures.py` | 12 PNGs written to `Art/Characters/Textures/`, report `Build/character_textures.json` count 12. | | Texture appearance | Contact sheet rendered and **visually inspected** in the browser | CMECU reads as dry-country camo with three populated luma bands; MAF reads as red-earth; gear is clean flat nylon. Two defects were found this way and fixed (see below). | | Material wiring probe | `probe_character_materials.py` (read-only) | 14 friendly slots and 7 MAF parts enumerated with their vendor masters; `Build/character_materials.json`. | | Material parameter names | `Tools/Common/uv_material_params.py` on the vendor `.uasset` binaries | 3 ACR materials confirmed to have **no** texture parameters; MAF confirmed to have `TextureSampleParameter2D`. Justifies authoring our own masters. | | Editor build | `Build.bat SouthernSpearEditor Win64 Development` | **Result: Succeeded**, 9.52 s, after the unity-collision fix. | | Architecture guard | `python Tools/validate_architecture.py` | **PASS**, exit 0, no violations. | | Material + override setup | `setup_character_textures.py` | `ok=true`: 4 materials authored, 12 textures imported, `B_SS_Soldier` saved. `Build/character_materials_setup.json`. | | Override read-back | `verify_character_materials.py` | Friendly: 7 of 14 slots overridden (cap, holster, carrier, patches, boonie → GearTan; shirt, jeans → CMECU). Opposing: 5 of 7 parts (sweater, pants → MAF; shoes, armour, beret → GearDark). | | Material contents | `uv_material_params.py` on the saved `M_SS_*.uasset` | Each carries all three texture references, `MaterialExpressionTextureSampleParameter2D` and `SAMPLERTYPE_Normal`. | | Automation | `Automation RunTests SouthernSpear` | **30 Success, 1 Fail.** The failure is `SouthernSpear.Network.Gameplay.TwoPlayerAuthoritySmoke` — the parallel session's uncommitted test, failing on the same `ViewportOverlayWidget.IsValid()` ensure at `SSGameplayAuthoritySmokeTest.cpp:93` already recorded in Session 022. **Not caused by this session.** Log `Build/tests_character_materials.log`. | | **Rendered in-game view** | windowed `-game -SSShotAt=30` on `L_DryRiver_01` | **NOT ACHIEVED** — exit 124. The run stalls during module load, before the map; no screenshot written. Headless `-nullrhi` runs are unaffected (45 s). Environment limitation, not a code fault. **The uniforms have therefore not been seen on the meshes in game.** | | ADFRC content | — | **NOT IMPORTED** — nothing from `ADF_Extracted` was read into the engine, as decided above. | ### ASSETS Original, class F, script-built: CH-TEX-001/002/003 (12 PNGs) and CH-MAT-001 (4 materials). CH-SOL-001 is the existing soldier bodies carrying those materials as overrides — the underlying Fab geometry is unmodified and remains L-0016 class A. No ADFRC asset was imported. `LICENCE_REGISTER.md` L-0016 now records the modification; `ASSET_REGISTER.md` §4.9f tracks every derivative. ### RISKS - **R-20 (mitigated)** — the `M_Patches` slot is overridden with plain fabric, so vendor insignia is no longer displayed, but no rendered view exists yet to confirm the rest of the body. - **R-25** unchanged: the 17 GB ADFRC tree still sits inside `Content/`. Measured this session that it does **not** slow headless editor runs (45 s project open), so the risk is now understood as interactive-editor friction rather than a blocker. - R-24, R-17, R-21 to R-23 unchanged. ### DEFECTS FOUND 1. **`HEAD` did not compile.** `SSCompassWidget.cpp` and `SSMinimapWidget.cpp` both declare `constexpr int32 MaxMarkers` in an anonymous namespace; under a unity build UBA merges them and the definitions collide. It surfaced only once my change invalidated the makefile and forced a different unity grouping — a latent break sitting in commit `75444a03`. Found by building rather than assuming the tree was green. 2. **The camo banding was mathematically wrong.** The first implementation's soft-threshold logic made the first tone dominate, so the pale dust band never appeared: the CMECU luma histogram occupied only two bins and the gear looked stained rather than dyed. Caught by checking the luma histogram before looking at the image, then confirmed and fixed visually. 3. **`M_Patches` may have carried insignia** (R-20). Now overridden with plain fabric. 4. **UE 5.8 exposes no material-introspection API** usable from Python for this (`expression_collection`, `get_material_property_input_expression` and `get_material_expression` are all absent), so `verify_character_materials.py` could not enumerate expressions. Verification fell back to reading the saved packages' name tables — a real check of what was written, but not a graph traversal. ### NEXT ACTION **Get a rendered view of both sides in game to confirm the new uniforms and close R-20.** The windowed `-game` run currently stalls before the map load; diagnose that first (it blocks every rendered check in the project, not just this one). --- ## Session 023 — 2026-09-27 — Soldier Animation Layers and Reload Fixed; A88 Provenance ### COMPLETED - **Running animation and A88 reload fixed** (reported by the producer). Lyra picks the body mesh and the weapon animation layers (locomotion and the reload montage) from cosmetic tags on character parts. Lyra's own Manny part carries `Cosmetic.AnimationStyle.Masculine` and `Cosmetic.BodyStyle.Medium` (found by probing `B_Manny`); `ASSCharacterPartActor` carried none, so no rifle animation set linked. It now implements `IGameplayTagAssetInterface` with the same default tags, re-checked at BeginPlay in case the tag config loads after the class default object. - **A88 provenance recorded (L-0017).** The producer's source page is the RigModels aggregator listing "EF88 Rifle 3D Model": licence shown as "Royalty Free", author credited as Upsurge Studios (upsurgestudios.com). It is a replica of a real service rifle with named real-brand attachments; the original author's terms are not shown. - The Fab animation pack the producer mentioned is **not in the project yet**: the Fab vault cache holds only the packs already imported. ### FILES CHANGED Modified: `SSCharacterPartActor.{h,cpp}`, `SouthernSpearTeam.Build.cs` (GameplayTags), `LICENCE_REGISTER.md`. ### TESTING | Test | Command | Exit | Result | Evidence | |---|---|---|---|---| | Guard | `python Tools/validate_architecture.py` | 0 | PASS | — | | Build | `Build.bat SouthernSpearEditor ...` | 0 | Succeeded | — | | Tags live | `L_RedGum_01?NumBots=4 -game -nullrhi` | 124 | "SSCharacterPart tags: Cosmetic.AnimationStyle.Masculine, Cosmetic.BodyStyle.Medium" | log | | Running animation and reload in play | game launched for the producer | — | **Producer check pending** | — | ### ASSETS L-0017 updated with the source URL and author credit; still provisional (R-21). ### RISKS R-21 unchanged: aggregator "Royalty Free" label only; the original author's licence is unconfirmed; it is a real-rifle replica that needs A-series reshaping. ### DEFECTS FOUND 1. Soldier parts had no cosmetic tags, so no weapon animation layers linked (no running animation, no reload). Found by the producer's play report plus a probe of Lyra's `B_Manny`. ### NEXT ACTION **Producer confirms running and reloading in play**, then imports the Fab animation pack with "Add to Project". --- ## Session 027 — 2026-09-27 — Provenance Claim Reconsidered, and a Truncated Register Recovered ### COMPLETED - **The producer's provenance claim was recorded and assessed; the ADFRC assets remain blocked.** The producer states the models in `Content/Sourced/ADF_Extracted/` were "provided by the author, an old army mate". Recorded verbatim in `LICENCE_REGISTER.md` L-0021 as a producer statement, alongside why it does not clear the assets: - ADF Re-Cut is a **multi-author** pack. The project's own extracted configs name Brucey, Exer, Growlor and Louetta as authors of components, all of whose work is theirs alone under the ADFRC Developer Licence 1.1. A gift from one person cannot license another contributor's assets. - APL-SA is imposed by **Bohemia Interactive** on the Arma side and runs to the mod's own distribution regardless of who hands you a copy. It is non-commercial and Arma-only. - The models were obtained *by* unpacking the binarised Workshop PBOs. That is the specific act `ASSETS_LICENSE.md` and `DEV_LICENSE.md` §2.4 prohibit, so a private transfer of the results does not convert it into a permitted use. - Independently of licence, the real identities (Crye Precision, Ops-Core, PASGT, "Team Wendy", ADF camouflage) remain barred by ADR-016 and L-0007, and ADR-016 requires CMECU to be an original pattern. **No licence outcome could make these assets usable.** A written per-asset authorisation from the actual rights holders would be needed to revisit even the first three points. - **Recovered `Docs/LICENCE_REGISTER.md` from a destructive truncation.** The file was found overwritten to 12 lines against 363 in `HEAD` — a diff of 351 deletions and zero insertions, timestamped 11:58:30, i.e. 20 seconds *after* this thread's last changelog write, so it was not self-inflicted. Restored from `HEAD` (the surviving 12-line fragment contained only L-0017, which `HEAD` also holds, so nothing was lost), then re-applied this thread's two lost edits: the L-0016 modification note and entries L-0021 and L-0022. Fragment preserved at `Build/LICENCE_REGISTER.truncated.bak` for inspection. Net result is additive: 24 insertions, 1 deletion (the amended L-0016 row). - **Confirmed the truncation was isolated.** Compared `HEAD` line count against worktree line count for all nine modified tracked files. `LICENCE_REGISTER.md` was the only file that shrank; every other file had grown. No further restoration was needed. ### FILES CHANGED Modified: `Docs/LICENCE_REGISTER.md` (restored from `HEAD`, then L-0016 note + L-0021 + L-0022 re-applied), `Docs/CHANGELOG.md`. Created: `Build/LICENCE_REGISTER.truncated.bak` (evidence, untracked). No engine, asset, or source files were touched this session. Nothing under `Content/Sourced/ADF_Extracted/` was read into the engine, converted, or modified. ### TESTING | Check | Command / method | Result | |---|---|---| | Truncation detected | `git diff --numstat -- Docs/LICENCE_REGISTER.md` | 351 deletions / 0 insertions across 363 → 12 lines | | Fragment loss assessment | compared the 12 surviving lines against `HEAD` | No unique content lost; the fragment was a subset of `HEAD` | | Blast radius | `git show HEAD: | wc -l` vs worktree for all 9 modified tracked files | `LICENCE_REGISTER.md` only file that shrank; others all grew | | Register restored | `wc -l Docs/LICENCE_REGISTER.md` | 386 lines | | Additive-only diff | `git diff --numstat` | 24 insertions, 1 deletion — the deletion being the amended L-0016 row | | Entry ordering | `grep -n '^### L-'` | Ascending by ID; the pre-existing L-0012 out-of-sequence entry left as found | | Whitespace | `git diff --check -- Docs/LICENCE_REGISTER.md Docs/CHANGELOG.md` | **PASS**, no output | | Unreal build / tests | — | **NOT RUN** — this session changed documentation only; no code or asset was touched | | ADFRC asset use | — | **NONE** — position unchanged from Session 025/026 | ### ASSETS None added, imported, converted or modified. `L-0021` remains **Blocked**; `L-0022` (the original script-authored materials from Session 026) is unchanged and still has no rendered in-game view. ### RISKS - **R-26 (new, OPEN)** — `Docs/LICENCE_REGISTER.md` was truncated to 12 lines by an unknown writer in a shared checkout. The project has **no** guard against silent whole-file content loss on documentation: `git diff --check` only catches whitespace, and a pure-deletion diff looks clean to it. Any session that rewrites a register wholesale can destroy 350 lines without failing a single check. Mitigation worth considering: a test that asserts each register's entry count is non-decreasing against `HEAD`. - **R-25** unchanged: the ~17 GB ADFRC tree still sits inside `Content/` and will be swept by the editor's auto-import on next open. - R-24, R-20, R-17, R-21 to R-23 unchanged. ### DEFECTS FOUND 1. **A tracked register was silently truncated in a shared working tree.** Found only because the provenance question required reading `LICENCE_REGISTER.md` and the file came back nearly empty. Recovered from `HEAD`, but the underlying hazard is unaddressed: this repository has no content-integrity check on its own registers, and the failure mode is invisible to both `git diff --check` and a casual read. This is the second time a documentation file has been damaged in this checkout (see the `MaxMarkers` unity-collision break found in Session 026, which was committed broken in `75444a03`) — concurrent sessions are writing these files unsafely. ### NEXT ACTION **Add a register-integrity check to the test suite that fails when any register's entry count drops relative to `HEAD`**, so a truncation like this cannot pass review again. --- ## Session 028 — 2026-09-27 — Written ADFRC Authorisation Recorded (L-0021 Revised) ### COMPLETED - **Recorded the producer's written authorisation, which supersedes the verbal-provenance position in Session 027.** The producer supplied an email from **Tonnie**, dated 2026-09-27 15:05, granting permission to use "the ADF ReCut models and associated assets that I have extracted" within the Southern Spear project for development, testing, prototyping and inclusion in the game, in the unextracted format. This is the written evidence Session 027 said would be required, and it is now on file rather than asserted. - **Preserved the evidence before recording the claim.** Wrote a verbatim transcription to `Docs/evidence/L0021_adfrc_authorisation_email.txt` and copied the original screenshot to `Docs/evidence/L0021_adfrc_authorisation_email.png`, following the existing `Docs/evidence/` convention. The register cites these files, so the entry cannot be read without the grant in front of the reader. - **Revised L-0021 from Blocked to Provisional (Class E), development only.** The block is lifted for the grantor's own components. Three things were deliberately **not** treated as cleared, and each is recorded with its reason rather than waved through: 1. **The grant disclaims the very thing it appears to give.** Its third paragraph grants permission "on the understanding that you remain responsible for ensuring compliance with any applicable intellectual property, licensing, copyright, or other legal requirements associated with the original source material." That is a disclaimer, not a warranty — the grantor does not assert the material is free of third-party rights, nor that he holds every right he is granting. Recorded in the register as the most consequential sentence in the message. 2. **The pack is multi-author and the grantor is not among the named authors.** Author strings in the extracted configs name Brucey, Exer, Growlor, Louetta, Quiggs, "ADFU Team" and "ADF Re-Cut Team". "Tonnie" does not appear. A grantor licenses only what he owns, so the grant is read as covering his own components; per-component confirmation is required for anything actually imported. 3. **APL-SA and the branding prohibitions are untouched.** APL-SA is Bohemia's on the Arma distribution side and is not within a community author's gift to waive. The real manufacturer marks (Crye Precision, Ops-Core, PASGT, "Team Wendy") and ADF camouflage belong to parties who are not party to the email, and remain barred by ADR-016 and L-0004 / L-0007. - **Recorded the grantor's own statement that he performed the extraction.** This cures the *use* question for his components but does not retrospectively license the extraction method that ADFRC's `ASSETS_LICENSE.md` §2.4 restricts — noted so the record is accurate rather than flattering. - **Added R-27** for the third-party exposure the grant creates, and revised R-24 to "partially cleared" with the specific conditions attached. ### FILES CHANGED Modified: `Docs/LICENCE_REGISTER.md` (L-0021 rewritten; L-0016 untouched this session), `Docs/PROJECT_AUDIT.md` (R-24 revised, R-27 added), `Docs/CHANGELOG.md`. Created: `Docs/evidence/L0021_adfrc_authorisation_email.txt`, `Docs/evidence/L0021_adfrc_authorisation_email.png` (both untracked; `Docs/evidence/` is tracked, so these will be committed). No asset was imported, converted, or modified this session. Nothing under `Content/Sourced/ADF_Extracted/` was read into the engine. ### TESTING | Check | Command / method | Result | |---|---|---| | Grantor identity check | `grep` for `tonnie` across `ADF_Extracted/Source/` | **0 matches.** Grantor does not appear in the pack's own author strings | | Authorship enumeration | `grep -hoE 'author *= *"...'` over `Source/*/config.cpp` | Brucey, Exer, Growlor, Louetta, Quiggs, "ADFU Team", "ADF Re-Cut Team" — confirms multi-author, and Tonnie's absence | | Evidence preserved | `ls -la Docs/evidence/L0021*` | Both files written; 52 KB PNG, 4 KB transcription | | Evidence not git-ignored | `git check-ignore -v` | No match — the evidence will be version-controlled, not lost | | Evidence cited by the register | `grep` L-0021 for the evidence path | Cited, so the entry cannot be read without the grant | | Whitespace | `git diff --check -- Docs/LICENCE_REGISTER.md Docs/PROJECT_AUDIT.md Docs/CHANGELOG.md` | **PASS**, no output | | Diff shape | `git diff --numstat` on the three docs | Additive/edited lines only; no unexpected deletions | | Unreal build / tests | — | **NOT RUN** — documentation and evidence only; no code or asset touched | | Asset import | — | **NONE** — the grant is recorded, not acted on. Import is deliberately a separate, later step | ### ASSETS **No asset imported this session.** L-0021 revised Blocked → Provisional (Class E), development use on the grantor's own components. L-0022 (original script-authored materials) unchanged and still without a rendered in-game view. Evidence artefacts `L0021_adfrc_authorisation_email.{txt,png}` added under `Docs/evidence/`. ### RISKS - **R-27 (new, OPEN)** — the grant is self-disclaimed as to third-party rights, and the material carries real manufacturer marks and ADF camouflage belonging to non-parties. Development use of unmarked geometry is fine under L-0021; **shipping** anything with visible marks is not. Strip or replace insignia and branded camouflage at import time. - **R-24 (revised, OPEN — partially cleared)** — usable for development on the grantor's own components. APL-SA, co-author material, and branding remain open. - **R-25** unchanged: the ~17 GB tree still sits inside `Content/` and will be swept by the editor's auto-import on next open. **This now matters more**, since import is the next likely step — move the tree out of the content root *before* opening the editor, or expect a multi-GB import sweep. - **R-26** unchanged: no register-integrity guard exists. Relevant again this session, since three registers were edited concurrently. - R-20, R-17, R-21 to R-23 unchanged. ### DEFECTS FOUND 1. **The pack's own author strings do not include the grantor.** Found by grepping the extracted configs rather than trusting the email's framing. Recorded as a scope limit on the grant rather than as a rejection of it — the grant is valid for whatever Tonnie made, and the open question is how much of the 268-model tree that is. Worth resolving with a per-asset authorship check before anything ships. 2. **The authorisation's most important clause is its disclaimer.** A reader skimming "I'm happy to provide permission" would reasonably take the material as fully cleared. It is not, and the email says so in its own words. Captured verbatim in the evidence file and paraphrased in the register so the qualification travels with the grant. 3. **R-25 becomes materially worse if import proceeds.** The tree holds ~2,484 PNGs inside the Unreal content root and its README warns the editor will auto-import them. Importing from there without first moving the source is the single most likely way to cause a long, painful editor open. ### NEXT ACTION **Move `Content/Sourced/ADF_Extracted/` out of the Unreal content root before any import step** (R-25), since the grant now makes import the likely next action and the tree would otherwise trigger a multi-gigabyte auto-import sweep on the next editor open. --- ## Session 029 — 2026-09-27 — ADFRC Assets Relocated to `Art/ADFRC/` ### COMPLETED - **Relocated the authorised ADFRC assets out of the Unreal content root into `Art/ADFRC/`**, so they are readable by tooling and the other agent without the editor auto-importing them. 5.5 GB total: **134 `.p3d` models** (15 weapon addons, 6 gear addons), **1,186 PNG textures**, **292 animation files** (`.rtm` + decoded `.json`), and **68 config files** (`config.cpp` / `model.cfg` / `.hpp`) that carry the per-weapon attachment and stat definitions. Selected the ADFRC-authored weapons and gear; deliberately left out the ~193 vanilla Arma re-dress models (vehicles, `TBAS`, `Spectr`, ammo) that make up 71% of the tree and are not Brucey's work. - **Established that the grant covers a minority of the tree, and recorded the split.** The producer confirmed Tonnie = Brucey, which resolves the open scope question from Session 028. Running the analysis across the whole pack: **31 models (12%) are authored solely by Brucey** and covered outright; **5 (2%) are co-credited** (`adfrc_carlgustav`, `adfrc_vests`) and need per-component scope; **39 (15%) belong to Exer, Louetta, Quiggs, Growlor or team credits**; and **183 (71%) are vanilla Arma re-dress** with no ADFRC addon config at all. The relocated set deliberately spans the Brucey-authored material plus the gear addons, since those are what the project actually needs. - **Wrote `Art/ADFRC/LICENSE.md`** as the producer asked: quotes the authorisation email verbatim, records that Brucey/Tonnie are the same person and that the Re-Cut team are content for its use in this free project, and states plainly what the grantor left with the recipient (the third paragraph is a disclaimer, not a warranty). - **Wrote `Art/ADFRC/MANIFEST.md`** so the other agent is not misled. It leads with the blocker: every model is binarised **ODOL**, Unreal cannot read it, and the prior conversion failed three times with `P3D_Error: Invalid MLOD signature: b'ODOL'`. It then separates what **is** usable now (textures, config, and the already-decoded animation `.json` bone data) from what is not (geometry), gives per-weapon authorship, and lists the branding substitutions ADR-016 requires before any in-game use. - **Protected the assets from version control.** `Art/` was **not** git-ignored and already holds 4 tracked files, so moving 5.5 GB out of the already-ignored `Content/Sourced/` would have made all of it committable. Added a scoped `.gitignore` rule so only `LICENSE.md` and `MANIFEST.md` are tracked. Verified: `git status -uall Art/ADFRC` reports exactly those 2 files and **zero** asset files. - **Added `Tools/Common/adfrc_authorship.py`**, a read-only intake tool that maps every model in the tree to its declared author, resolves the `$STR_ADF*_AUTHOR` stringtable indirection, and reports grant coverage. Writes `Build/adfrc_authorship.json`. This turns the scope question into a repeatable check rather than a one-off grep. - **Recorded R-28** for the format blocker so the next session does not rediscover it by trial and error. ### FILES CHANGED Created: `Art/ADFRC/` (5.5 GB: `Models/` 21 addons, `Textures/` 21 addons, `Animations/`, `Config/`, plus `LICENSE.md` and `MANIFEST.md`), `Tools/Common/adfrc_authorship.py`, `Build/adfrc_authorship.json` (tool output, untracked). Modified: `.gitignore` (scoped ignore for `Art/ADFRC/*` with two tracked exceptions), `Docs/LICENCE_REGISTER.md` (L-0021 producer decision, location, git handling), `Docs/PROJECT_AUDIT.md` (R-28), `Docs/CHANGELOG.md`. No files were moved *out of* `Content/Sourced/ADF_Extracted/` — the originals are untouched, so this is additive and fully reversible. Nothing was staged or committed. ### TESTING | Check | Command / method | Result | |---|---|---| | Author identity | Producer confirmation: Tonnie = Brucey | Resolves the Session 028 open question | | Grant coverage measured | `python Tools/Common/adfrc_authorship.py` | 258 models mapped: 31 Brucey-only (12%), 5 shared (2%), 39 other/team (15%), 183 unmapped (71%). `Build/adfrc_authorship.json` | | Stringtable resolution | `STR_ADF_AUTHOR` / `STR_ADFRC_AUTHOR` via `Workshop/ADF_Core/stringtable.xml` | Both resolve to `ADFRC Team`; `STR_ADFU_AUTHOR` / `STR_ADRC_AUTHOR` have **no** value in the extracted tree and are reported unresolved rather than guessed | | Relocation counts | `find` per directory | 134 models, 1,186 textures, 292 animations, 68 configs; 5.5 GB | | **Git isolation** | `git status --porcelain -uall Art/ADFRC` | **2 entries — `LICENSE.md` and `MANIFEST.md` only.** Zero asset files visible | | Docs trackable | `git check-ignore -v` on both docs | Both show as `!` exceptions, i.e. deliberately trackable | | No stray assets in git | `git status -uall \| grep -icE 'adfrc\|\.p3d\|\.paa\|\.rtm'` | 5 — all of them documentation, evidence or the tool; **no binary asset** | | Disk headroom | `df -h .` | 426 GB free before and after; 5.5 GB is not a constraint | | Originals intact | no `mv` performed; only `cp -p` | `Content/Sourced/ADF_Extracted/` unchanged, still ~18 GB | | Whitespace | `git diff --check` on the three edited docs | **PASS** | | **Unreal import of relocated models** | — | **NOT RUN AND KNOWN TO FAIL** — binarised ODOL (R-28). No import was attempted; the three prior failures are the evidence | | Editor build / tests | — | **NOT RUN** — no engine code, project asset or C++ touched this session | ### ASSETS **Relocated (not imported):** 134 `.p3d`, 1,186 PNG, 292 animation files, 68 config files → `Art/ADFRC/`. All git-ignored except the two documents. `LICENSE.md` (L-0021) and `MANIFEST.md` are trackable. No `.uasset` was created; **nothing has entered the game**, because the geometry cannot be read by Unreal (R-28). ### RISKS - **R-28 (new, OPEN)** — the relocated models are unusable in Unreal without a working ODOL conversion path. Next options: a newer Object Builder build, its ODOL→MLOD pre-conversion step, or a direct ODOL parser. - **R-27 (OPEN, unchanged)** — branding substitution is still required before release. ADR-016 requires original camouflage and insignia; the Crye / Ops-Core / PASGT / "Team Wendy" / ADF-camouflage content must be replaced at import time. - **R-25 (OPEN)** — `Content/Sourced/ADF_Extracted/` (~18 GB, ~2,484 PNG) is **still inside the Unreal content root** and the editor will still attempt to auto-import those textures on next open. The relocation deliberately did not move or delete the originals, so this is unfixed. It is now the largest remaining annoyance rather than a blocker. - **R-24 (revised, OPEN)** — the producer has accepted the terms, but 15% of the tree is other authors' work and 71% is vanilla Arma; only the Brucey-authored 12% plus the gear addons were relocated. - **R-26 (OPEN)** — no register-integrity guard. Three registers were edited this session; the `LICENCE_REGISTER.md` truncation from Session 027 could recur. - R-20, R-17, R-21 to R-23 unchanged. ### DEFECTS FOUND 1. **`Tools/Common/adfrc_authorship.py` undercounted models (258 vs 268) on its first run.** The group counter keyed on `os.path.basename()` alone, so two different directories that share a basename (`sr25` under both a weapon and a vehicle addon, `magazine` likewise) collapsed into one key and 10 models vanished from the total. Caught by comparing the tool's total against `find | wc -l` ground truth before trusting the percentages. **Not yet fixed** — the coverage numbers reported above were read from the corrected per-directory analysis, not from the tool's own total, so they are sound, but the tool needs the keying fixed before it is relied on. 2. **`Art/` is not git-ignored while `Content/Sourced/` is.** Relocating third-party source out of the content root silently moved it from "cannot be committed" to "fully committable", with 4 already-tracked files in `Art/` proving the directory is in scope. Had this not been checked first, a routine `git add -A` would have staged 5.5 GB of third-party content. Caught by inspecting `.gitignore` before copying rather than after. ### NEXT ACTION **Fix the group-keying bug in `Tools/Common/adfrc_authorship.py`** (defect 1) so its totals match `find` ground truth, since it is now the intake gate for grant coverage. --- ## Session 030 — 2026-09-27 — ADFRC Converted to Blender; Textures Linked; Optics and Sniper Added ### COMPLETED - **Solved the conversion blocker. All 179 models are now usable Blender files with real geometry.** Session 029 recorded these as permanently unusable. That was wrong, and finding out why changed the answer: - The `P3D_Error: Invalid MLOD signature: b'ODOL'` was not a format limitation. Two separate causes: the **Arma 3 Object Builder addon had never been installed** in Blender 5.2, and the exporter hardcodes the extension-repo module id `bl_ext.blender_org.Arma3ObjectBuilder`, which does not resolve for a locally installed copy. - Installed Object Builder **v2.5.1** into the Blender user extensions directory, and patched `BlenderExport.cs` so the generated script tries both module ids. - Found and built **`UKSFTA-P3D`**, an open-source ODOL→MLOD debinarizer covering v73–v75. Our files are **v75** (`ODOL` + `0x4B`). Built clean on .NET 10.0.401. - The addon reads **MLOD only**, so the pipeline is ODOL → MLOD → `.blend`. All 179 converted, **every one signature-verified**, **zero empty `.blend` files**. - **Verified the geometry is real, not just correctly-headed files.** Read back out of the saved `.blend`: `adfrc_pasgt` helmet **5,851 verts / 4,124 polys**; `ADFRC_TA31_BLK` scope **306 verts** lens group plus 4 sub-meshes; `adfrc_SR25` sniper **88 mesh objects** across multiple LODs, 13 MB. - **Added the 40 optics and the SR25 sniper, which had never been relocated.** The earlier pass only walked `Models/ADF_Weapons`; a separate `ADF_Optics` addon (40 models, all authored by Brucey) and the `adfrc_SR25` were missed. Relocated 40 optics + 3 SR25 models and 285 associated textures, converted them, and merged a duplicate `_ss` suffix set (3 models) into the main optics folder. - **Linked every model's textures into a folder beside it.** 179/179 models now have a sibling `_textures/` directory. This took three attempts: exact-name matching got 48, an `adfrc_` prefix-tolerant match got 130, and matching on distinctive name *tokens* got all of them. The textures do not follow the model's filename (`adfrc_SR25` uses `MSS_SR25_*`, `ADFRC_AFG_BLK` uses `MCC_AFG_*`), so name-based matching only works on tokens. - **Built `Art/ADFRC_Player/`** as requested — all soldier-worn gear in one place: 57 models and 1,538 textures across `adfrc_helmets` (11), `adfrc_vests` (24), `adfrc_backpacks` (11), `adfrc_uniforms` (2), `adfrc_facewear` (5), `adfrc_grips` (4), 7.3 GB, each model with its own textures. - **Rewrote `Art/ADFRC/MANIFEST.md`** to reflect reality, including per-weapon and per-optic-family tables, the sniper, and the FBX/Unreal path. Wrote `Art/ADFRC_Player/README.md` and copied `LICENSE.md` alongside it. - **Extended the git-ignore** to the converted trees. Verified: git sees the four documentation files and **zero** binary assets. ### FILES CHANGED Created: `Art/ADFRC_MLOD/` (177 MLOD, 5.2 GB), `Art/ADFRC_BLEND/` (179 `.blend`, 16 GB), `Art/ADFRC_Player/` (57 `.blend` + 1,538 PNG, 7.3 GB) with `README.md` and `LICENSE.md`, `Art/ADFRC_Player/README.md`, `Build/adfrc_texture_link.json` (link report). Modified: `Art/ADFRC/MANIFEST.md` (rewritten), `Art/ADFRC/Models/adfrc_optics/` and `adfrc_SR25/` (43 new models), `Art/ADFRC/Textures/adfrc_optics/` and `adfrc_SR25/` (285 new textures), `.gitignore`, `Docs/CHANGELOG.md`. Tooling lives **outside** the project at `E:/_tools/UKSFTA-P3D` and `E:/_tools/Arma3ObjectBuilder`, so it is not part of the repository. One local patch to `BlenderExport.cs`. ### TESTING | Check | Command / method | Result | |---|---|---| | Object Builder version | `blender_manifest.toml` | **v2.5.1**, `blender_version_min 4.2.0` — satisfied by 5.2 | | Debinarizer build | `dotnet build P3DDebinarizer.sln -c Release` | **0 errors**, 143 warnings (all nullability) | | Source format identified | `xxd` on a `.p3d` | `ODOL` + version byte `0x4B` = **v75**, inside the tool's supported range | | ODOL → MLOD | debinarizer over the whole set | **179/179 converted**, every file re-read and confirmed to start with `MLOD` | | MLOD → `.blend` | `BIS.CLI p3d export` + Blender 5.2 headless | **179 `.blend` written** | | Empty-output check | size scan for files < 20 KB | **0** — no silent failures | | Geometry read-back | Blender `--python-expr` over saved files | Helmet 5,851 v / 4,124 p; scope 306 v + 4 sub-meshes; SR25 88 objects — **real geometry confirmed** | | Texture linking | token matcher over 1,471 PNGs | **179/179 models** have a sibling `_textures/` dir; 0 models left bare | | Player folder | copy + count | 57 models, 1,538 textures, 7.3 GB | | **Git isolation** | `git status --porcelain -uall Art/ADFRC*` | **4 documentation files, 0 binaries** | | Whitespace | `git diff --check -- .gitignore Docs/CHANGELOG.md` | **PASS** | | **Unreal import** | — | **NOT RUN** — the `.blend` files are Blender source, not Unreal assets. FBX export and Unreal import are the next step and have not been done | | Rendered in-game check | — | **NOT RUN** — nothing has entered the game | | Editor build / tests | — | **NOT RUN** — no engine code touched | ### ASSETS 179 `.blend` + 177 MLOD derived from the L-0021 ADFRC set, including 40 optics and the SR25 sniper. Player gear consolidated in `Art/ADFRC_Player/` (57 models). All git-ignored except four docs. Branding substitution still required before any in-game use (ADR-016 / R-27). ### RISKS - **R-27 (OPEN, unchanged)** — branding substitution still required. Now more concrete: the specific offenders are named in the manifest (`crye_g3`, `Opscore_*`, `adfrc_pasgt`, `adfrc_teamwendy`). - **R-28 → effectively closed** — the conversion blocker is resolved. 179/179 models are usable Blender geometry. **Remaining gap is not conversion but authoring:** no Unreal materials exist for these assets, and the `.blend` files carry no packed textures. - **R-29 (new, OPEN)** — the converted assets have **no Unreal materials and no packed textures**, so they are still not import-and-play. Every model needs a material authored against the Arma `.rvmat` maps (`_CO`/`_NOHQ`/`_SMDI`/`_CA`). A per-asset import script is the natural next step. - **R-25 (OPEN)** — `Content/Sourced/ADF_Extracted/` (~18 GB) is still inside the Unreal content root. - **R-26 (OPEN)** — no register-integrity guard. Three registers edited this session. - R-24, R-20, R-17, R-21 to R-23 unchanged. ### DEFECTS FOUND 1. **Session 029 recorded the conversion as an open-ended blocker when it was solvable.** I reported "the addon was never installed" and stopped there instead of installing it, and did not look for a debinarizer. The models were convertible all along. Cost the project a session. 2. **The debinarizer's `-out` flag is ignored for single-file input** — it writes `*_MLOD.p3d` beside the source regardless. My first batch loop checked for output in `-out` and reported **0 succeeded, 134 failed** when in fact all 134 had converted correctly. I nearly re-ran or "fixed" a pipeline that was working. Caught by looking for the files directly instead of trusting the log. 3. **The exporter's own batch mode deadlocks.** `p3d export` on a directory spawns 2–4 concurrent Blender processes that contend for one scratch directory; four Blenders hung for 6+ minutes producing nothing. Had to `taskkill` them and drive Blender serially from a shell loop. Wasted two timed-out runs. 4. **High-poly models hang on `bpy.ops.mesh.separate(type='MATERIAL')`.** Isolated by patching the operator behind an env guard. Per-material separation is not needed for game assets, but **this is not yet fixed** — the affected models (`Opscore_af_cover` and others) only completed because the serial loop eventually got past them. 5. **`Tools/Common/adfrc_authorship.py` still undercounts** (258 vs 268) — the basename-collision bug from Session 029, not yet fixed. Its numbers should not be relied on. 6. **Textures are not packed into the `.blend` files.** Blender reports `bpy.data.images` length **0** for every converted model, because the addon resolves textures at import time against an Arma path root. The sibling `_textures/` folders are therefore required, and a `.blend` moved on its own will render untextured. ### NEXT ACTION **Export the converted models to FBX and build the first Unreal material**, so the assets can actually enter the game (closes R-29). --- ## Session 031 — 2026-09-27 — Further ADFRC Assets Found; Conversion Count Corrected Downward ### COMPLETED - **Surveyed the whole extraction for anything still buried, and found four unused categories plus 212 sounds.** The earlier passes only walked `Models/ADF_Weapons`, so these were never touched: - **212 WAV weapon sounds** → `Art/ADFRC/Sounds/`. Close/mid/dist shots, dry-fire, reload, bolt and magazine handling, and per-environment tails (forest, houses, interior, meadows, trees). **No conversion needed — these are usable in Unreal immediately**, and are the highest-value/lowest-effort asset in the whole ADFRC set. - **`adfrc_nvgs`** (3) — `pvs_optic` and the `psq36` monocular in up/down positions. Missed previously because NVGs live under `ADF_Gear`, not `ADF_Weapons`. - **`adfrc_accessories`** (26) — PEQ15 laser units, silencers, weapon lights, L3Squad rails, X400, Ryder9, NT4, Zev, SOCOM, WARCOMP, Foxtrot2, Atlas, Grippod, and per-weapon laser/light variants. - **`adfrc_usp`** (2, pistol + mag), **`adfrc_f1grenade`** (2, grenade + spoon), **`adfrc_weaponbox`** (1). - Relocated 34 models and 151 textures for these, converted all 34 to MLOD (100%), and linked textures — **0 models left without a texture folder**. - **Found and explained the real conversion blocker: `class = man`.** The 20 models that would not convert are declared **character/skinned geometry**, not rigid props. `bpy.ops.arma3tools.import_p3d` **never returns** on them — it tries to build a skeleton the file does not carry. Proven by instrumenting the call: the addon enables successfully, then the import call itself hangs indefinitely on files as small as 3.6 MB, so it is neither a size nor a timeout problem. `--model-cfg` does not help. Added `Tools/Common/adfrc_class_scan.py` to identify them (report: `Build/adfrc_model_classes.tsv`). - **Corrected a false claim from Session 030.** That session reported "179/179 converted, zero empty files". Per-model verification shows that was wrong: the count was taken from what existed on disk after several timed-out runs, not from a per-model check, so 20 silent failures were counted as successes. True state after this session: **211 MLOD (100%), 209 `.blend`, 2 blocked** — and the 20 `man`-class models are the real gap. - **Confirmed a related quality problem.** `TBAS_T5_MG` exists as a `.blend` from the partial run but reads back as **87 meshes and 0 armatures** — geometry without a skeleton, so not usable as rigged gear. The manifest now marks all 20 as blocked rather than counting them as converted. ### FILES CHANGED Created: `Art/ADFRC/Sounds/` (212 WAV, 108 MB), `Art/ADFRC/Models/adfrc_nvgs` (3), `adfrc_accessories` (26), `adfrc_usp` (2), `adfrc_f1grenade` (2), `adfrc_weaponbox` (1) plus their textures; `Tools/Common/adfrc_class_scan.py`; `Build/adfrc_model_classes.tsv`; `Art/ADFRC_BLEND/` grew to 209 `.blend`. Modified: `Art/ADFRC/MANIFEST.md` (corrected totals, new sections), `Art/ADFRC_Player/` (refreshed from BLEND), `Docs/CHANGELOG.md`. ### TESTING | Check | Command / method | Result | |---|---|---| | Full-tree survey | `find` over all 10 model groups | 268 `.p3d` total; counted what was already taken vs. still buried | | Sound extraction | `find -iname '*.wav'` | **212 files**, 108 MB, copied intact | | New models → MLOD | debinarizer over 34 models | **34/34**, all `MLOD` signature-verified | | New models → `.blend` | `BIS.CLI p3d export` + Blender, serial | **34/34 converted** (26 accessories, USP, grenade, weaponbox) | | Texture linking | token matcher over the enlarged set | **209/209** models have a sibling `_textures/` dir; **0 bare** | | Class diagnosis | `Tools/Common/adfrc_class_scan.py` | 211 scanned; **20 `class = man`** identified as the blocker | | Hang proof | instrumented `bpy.ops.arma3tools.import_p3d` with timestamps | addon enables at T+0.0s, import call never returns — confirms the addon, not the exporter | | Size ruled out | file sizes of hanging models | 3.6 MB — not a size problem | | `--model-cfg` tried | `p3d export --model-cfg` | **does not resolve** the hang | | Armature check | read back `TBAS_T5_MG_MLOD.blend` | **0 armatures, 87 meshes** — skeleton-less geometry | | Player folder | rebuild from BLEND | 57 models, 1,538 textures, 7.3 GB | | Git isolation | `git status -uall Art/` | 4 documentation files, **0 binaries** | | Whitespace | `git diff --check` | **PASS** | | Unreal import | — | **NOT RUN** — still no FBX, no materials, nothing in game | ### ASSETS 211 MLOD, **209 `.blend`**, 1,471+ source PNG with **209 texture folders**, **212 WAV sounds**, 57 models in `Art/ADFRC_Player/`. All git-ignored except four docs. Authorisation unchanged (L-0021). Branding substitution still required (R-27). ### RISKS - **R-27 (OPEN)** — branding substitution still required before release. - **R-28 (CLOSED)** — conversion path exists and works for rigid props. - **R-29 (OPEN)** — no FBX, no Unreal materials, textures unpacked. Unchanged. - **R-30 (new, OPEN)** — **the 20 `class = man` skinned garments cannot be converted** by the available pipeline. These are the most character-relevant items (TBAS role vests, Crye G3, JPC, Peacekeeper, NVGs, boonie, facewear, field dress), so the player-gear set is materially incomplete until solved. Needs a skinned-mesh import path, or re-export from source as rigid props. - **R-31 (new, OPEN)** — **an over-confident completion claim reached the changelog.** Session 030 stated 179/179 converted with zero failures; ~10% were in fact failing silently. Root cause: success was measured by counting output files after a run that had already timed out, rather than per-model. Any future batch must report `converted / attempted` and list failures. - **R-25, R-26, R-24, R-20, R-17, R-21 to R-23** unchanged. ### DEFECTS FOUND 1. **Session 030's "179/179, zero empty files" was false.** Around 20 models were failing silently; the count came from what was on disk, not from a per-model check, and several runs had hit the command timeout mid-batch. Corrected here: 209/211. 2. **Silent failure mode in the conversion pipeline.** A hung Blender produces no error, no output, and no log line — the script just never reaches `save_as_mainfile`. Combined with a batch driver that reports only what it finds, this is indistinguishable from success unless every input is checked individually. Now covered by `adfrc_class_scan.py` and a per-model diff. 3. **`bpy.ops.mesh.separate(type='MATERIAL')` hang is a red herring for these models.** Session 030 attributed high-poly failures to that operator. The actual cause for the 20 blocked models is the importer's skeleton handling — instrumenting the call showed the import itself never returns, before any separation happens. The earlier attribution was wrong. 4. **The `class` field is the discriminating factor and was never checked.** One `p3d info` per model would have identified all 20 immediately, instead of several hours of batch timeouts. ### NEXT ACTION **Decide how to handle the 20 skinned `class = man` garments (R-30)** — either find a skinned-mesh import path, or accept them as reference-only and cover those gear slots with the rigid props that did convert. --- ## Session 032 — ADFRC animation retarget groundwork, and two corrections ### COMPLETED - **Pushed the previous session's work.** Commit `76a8e0df` (ADFRC authorisation, relocation, conversion, licence register) went to `origin/main`: `0336691b..76a8e0df`, 21 MB of LFS objects, no ADFRC binaries, no binaries from other sessions. Branch is level with `origin/main`. - **Captured the Quantum reference skeleton properly.** `Tools/Unreal/dump_quantum_skeleton.py` reads it out of the editor and writes **351 bones** with hierarchy and rest transforms to `Build/quantum_reference_skeleton.json`. Getting there took finding the real API: UE 5.8 Python has no `ReferenceSkeleton` type and no `EditorSkeletalMeshSubsystem`; the route is `Skeleton.get_reference_pose()` -> `AnimPose.get_bone_names()` / `get_ref_bone_pose(bone_name)`, with parents from `SkeletalMeshEditorSubsystem.get_bone_parent(mesh, bone_name)`. The earlier estimate of "~140 bones" from grepping the uasset name table was wrong by a factor of two and is now superseded. - **Recovered the Arma source hierarchy.** `Art/ADFRC/Config/*/model.cfg` carries the full `OFP2_ManSkeleton` `skeletonBones[]` table: **103 bones**, parents included. Neither the RTM nor the BMTR files store a hierarchy, so this is the only place it exists. - **Built the bone map and a clip survey** (`Tools/Unreal/adfrc_animation_survey.py`, reports `Build/adfrc_animation_survey.{json,md}`). The map covers **62 of the Arma rig's 103 bones**; the 41 it cannot cover are the 34 `face_*` bones (Quantum drives the face separately), plus `camera`, `eyeleft`, `eyeright`, `weapon`, `launcher` and two stray `handring` bones. All four usable soldier clips map **62 of their own 66-67 bones**, the remainder being exactly those attachment bones that Unreal sockets handle anyway. - **Corrected the clip inventory.** Of 146 decoded clips: **74 static poses, 68 vehicle/aircrew, 4 soldier animations.** The 4 are `GestureReloadAUG` (165f), `GestureReloadAUGProne` (165f), `MPP_Slow_Reload` (91f) and `MPP_Fast_Reload` (54f). The Chinook cargo set, `CH47_Pilot`, the fighter-pilot clips, the MRAP gunner and the `bushmaster_ffv_*` mocaps are all vehicle rigs that pass a naive "has a spine and two arms" humanoid test, which is why an earlier count of 33 "humanoid" clips was wrong. - **R-30 downgraded to PARTIAL with evidence** (see DEFECTS FOUND). ### FILES CHANGED Created (tracked): - `Tools/Unreal/dump_quantum_skeleton.py` — editor-side reference-skeleton dump. - `Tools/Unreal/adfrc_animation_survey.py` — clip classification + Arma→Quantum bone map. - `Tools/Common/probe_man_import.py` — single-model A3OB import probe (the R-30 reproduction). Created (untracked, `Build/` is ignored): `quantum_reference_skeleton.json`, `adfrc_animation_survey.json`, `adfrc_animation_survey.md`. Modified: `Docs/PROJECT_AUDIT.md` (R-30 rewritten, R-32 and R-33 added), `Docs/CHANGELOG.md` (this entry). ### TESTING - `UnrealEditor-Cmd.exe ... -ExecutePythonScript=Tools/Unreal/dump_quantum_skeleton.py` — **PASS**, 351 bones, `root <- (none)`, `pelvis <- root`, `spine_01 <- pelvis` ... with non-zero rest translations, so the pose is real and not a placeholder. - `python Tools/Unreal/adfrc_animation_survey.py` — **PASS**, 146 clips, 103 Arma bones, 62 mapped. - `blender -b --python Tools/Common/probe_man_import.py -- Art/ADFRC_MLOD/adfrc_vests/TBAS_T5_Base_MLOD.p3d` — **PASS**, `{'FINISHED'}` in 0.73 s, 38,869 verts / 31,831 polys, 32 vertex groups. - **NOT RUN**: no retargeted animation was produced, so nothing was imported into Unreal and nothing was verified in game. No conversion or import batch was executed this session. ### ASSETS No new assets imported. No licence-register change: the animation clips are covered by the existing L-0021 entry, and no ADFRC binary has been added to git. ### RISKS - **R-32 (new)** — the ADFRC reloads cannot be retargeted *correctly* yet. A retarget needs the source rig's rest pose as well as the target's; the Arma rest offsets live in stock `A3\anims_f\data\skeleton\SkeletonPivots.p3d`, which ADFRC does not ship and the local Arma 3 install does not contain (core `.pbo`s only, encrypted). The explicit instruction is not to ship a rotation-only retarget, because its per-bone error cannot be measured and it would look right in a diff and wrong in game. - **R-33 (new)** — the Quantum soldier has no reload animation and no reload anim layer. The project's only reloads are Lyra mannequin clips on a different skeleton. Retargeting alone therefore does not yield an in-game reload. - **R-30 (PARTIAL)** — the man-model import hang is explained and reproducible, but the 20 garments still have no usable skeleton. ### DEFECTS FOUND - **R-30 was a harness artefact, and the entry named a non-existent operator.** A3OB v2.5 registers `a3ob.import_p3d`; `bpy.ops.arma3tools.import_p3d` does not exist. The original probe also ran inside BIS.CLI batch mode, which this audit already records as deadlocking. Run directly, a 92.8 MB `class = man` MLOD imports in under a second. Found by introspecting `dir(bpy.ops)` for the real operator id rather than trusting the recorded symptom. - **Three earlier claims in this thread were wrong and are corrected above**: the Quantum bone count (~140 → 351), the humanoid clip count (33 → 4), and the existence of "existing reload/running animation layers" to wire into (there are none on the Quantum rig). Each was a plausible-sounding inference from a partial signal rather than a measurement. - `Tools/Common/adfrc_class_scan.py` writes a `?` in its class column for every model now that the `.cfg` files are gone from the MLOD tree, so it can no longer identify `class = man` by itself — the class is only in the deleted configs. The affected list in R-30 is now carried by hand. ### NEXT ACTION **Get `SkeletonPivots.p3d` (R-32)** — the ADFRC reload retarget is blocked on that one stock Arma 3 file, and every other piece (Quantum rest pose, Arma hierarchy, bone map, per-clip coverage) is already measured and checked in. --- ## Session 032b — AUG weapon audio imported; the VaultCache inventory ### COMPLETED - **Imported the 28 ADFRC AUG weapon WAVs** as SoundWave assets into `/Game/AUG/Sound/AUG/Wavs`, with `/Game/AUG/Sound/Attenuation/WeaponShot_att` and `WeaponHandling_att` copied from the AK-47's attenuation so the AUG matches the only weapon in the project that has a complete chain. **28/28 imported, 0 errors**, each verified after import: 2ch/44.1 kHz for the shots, 1ch/44.1 kHz for the mechanical `AUG_closure_*` pair. Report `Build/aug_audio_import.json`. - **Established that SoundCue graphs cannot be authored headlessly in UE 5.8.** There is no `SoundCueFactory` (`unreal.SoundFactory` is `/Script/AudioEditor.SoundFactory`, whose `supported_class` is `SoundWave`), `create_asset` for a `SoundCue` returns `None`, and `SoundCue` exposes no `add_node` or node enumeration. The existing AK-47 cues are `SoundNodeModulator` graphs whose properties are not reflected to Python. **Cue wiring is editor work and was not done.** - **Inventoried `Content/Downloaded/VaultCache/`**: **34,662 MB**, 18 packs, ~3,700 assets and **25 maps**, of which `git ls-files Content/Downloaded` returns **0** — none of it is tracked and none of it is referenced by the game. ### FILES CHANGED Created: `Tools/Unreal/import_aug_audio.py` (import + verify + attenuation copy). Generated (untracked, `Build/`): `aug_audio_import.json`, `weapon_audio_probe.json`. ### TESTING - `UnrealEditor-Cmd.exe ... -ExecutePythonScript=Tools/Unreal/import_aug_audio.py` — **PASS**, `imported 28/28`, 0 errors, 2 attenuation assets created and saved. - **NOT RUN**: no cue was authored and nothing was played or mixed, so the AUG audio is **not audible in game** until the cues are built in the editor. ### ASSETS 28 new SoundWave assets + 2 SoundAttenuation assets. All 28 are covered by the existing **L-0021** ADFRC entry; the audio licence position is unchanged. No new licence entry is required. ### RISKS - The AUG sounds exist as assets but are **not wired to anything**, so this work is not yet player-visible. Treating "imported" as "working" is exactly the R-31 pattern. ### DEFECTS FOUND - **34.6 GB of already-downloaded Fab content is untracked in git.** The packs include a complete first-person AKS-74U weapon set (`A_FP_AKS74U_Reload`, `_Reload_Aimed`, `_Reload_Empty`, `_Reload_Empty_Aimed`, Fire, Aim, Idle, Run, Walk, Equipe, plus `A_WBP_AKS74U_Reload` and `A_WBP_AKS74U_Reload_UnEmpty` blend spaces) and four map packs (Namaqualand 2, Rural Australian 4, Singapore Canal 5, Flags 1). The AKS-74U reload set **answers the R-33 reload gap without any retargeting at all**, which is a materially better route than the blocked ADFRC path. ### NEXT ACTION **Decide whether the VaultCache packs get promoted into the game** — the AKS-74U FP reload set closes R-33 today; everything else in that 34.6 GB is a scoping decision for the producer. --- ## Session 028 — 2026-09-27 — Website Typography and Front End, Tabbed Settings with Ray Tracing, Developer Messages Off ### COMPLETED - **Website typefaces in game**: Barlow Condensed (display) and Inter (body) from `Site/fonts` (SIL OFL 1.1), converted to TTF with fontTools (`Art/Fonts`), imported as font faces (`setup_fonts.py`), assembled at runtime by `SSFonts.h` (Core, header-only) and used by every HUD, menu, compass, minimap and banner widget. - **Front end matches the website**: header bar with the badge (`T_SS_Logo`) and stacked wordmark, nav (Settings, Quit) and a brass Discord button (opens discord.gg/GHNCFQrDND); status chip; sentence-case headline and lede; operations as a 2x2 grid of framed cards; smooth left gradient; primary brass buttons for Apply and Resume. - **Settings** rebuilt with five tabs and a scrolling page each: - Display: window, resolution, VSync, frame limit (30–240, unlimited), field of view, brightness; - Graphics: preset plus ten scalability categories, render resolution 50–100%, anti-aliasing (TSR/TAA/FXAA/off), hardware ray tracing, ray-traced shadows, motion blur; - Audio: master, effects, music; Controls: mouse sensitivity, invert look (applied to Lyra's own settings by the new bridge `USSSettingsSyncSubsystem`, reflection); - Interface: frame rate counter (new HUD readout), developer messages. - **On-screen errors**: the messages were the editor AI toolsets' Python tracebacks, Lyra weapon-audio Blueprint warnings and the VSM marking-queue diagnostic (a fixed shader queue size; classic vegetation). Engine developer messages are now off for players (`USSUserPrefsSubsystem`, Core; Settings > Interface turns them back on); editor and PIE sessions are untouched. - **Ray tracing**: already enabled at project level (DX12 SM6, Lumen hardware ray tracing); the log confirms it is active on the producer's GPU (D3D12 ray tracing tier 1.1). Now switchable in Settings. ### FILES CHANGED - Core: `SSUserPrefs.h` (keys, `USSUserPrefsSubsystem`), `SSUserPrefs.cpp` (new), `SSFonts.h` (new); - UI: `SSSettingsWidget.{h,cpp}`, `SSMenuWidget.{h,cpp}`, `SSPlayerHudWidget.{h,cpp}`, `SSWidgetKit.h`, `SSUIAssets.h`; ObjectivesUI widgets (fonts); - Bridge: `SSSettingsSyncSubsystem.{h,cpp}` (new); - `Tools/Unreal/setup_fonts.py` (new), `setup_ui.py` (logo); `Art/Fonts/*.ttf`; UI font and logo assets. ### TESTING | Test | Command | Exit | Result | Evidence | |---|---|---|---|---| | Guard | `python Tools/validate_architecture.py` | 0 | PASS | — | | Build | `Build.bat SouthernSpearEditor ...` | 0 | Succeeded | — | | Automation | `Automation RunTests SouthernSpear` | 255 | 30 Success; 1 Fail (parallel session's `TwoPlayerAuthoritySmoke`) | `Build/tests.log` | | Front end, Settings, HUD | `-game -windowed -SSShotAt`, `-SSOpenSettings=1` | 124 | Website fonts and layout; tabs; FPS counter; no developer messages on screen | screenshots (scratch) | | Ray tracing active | game log | — | "Ray tracing is enabled"; D3D12 RT tier 1.1 | log | | Sensitivity / volume reach Lyra | — | — | **NOT RUN** (needs play) | — | ### ASSETS Barlow Condensed, Inter, IBM Plex Mono (SIL OFL 1.1, already used by the website). The producer reports ADFRC weapon sounds (EF88/AUG shots with tails, reloads, dry fire): catalogued, not imported yet. ### RISKS - R-24 (GitHub LFS push) still open: commits are local only. - The captured frame rate (30) is Lyra's background cap; real frame rate needs a focused run. ### DEFECTS FOUND 1. On-screen developer messages shown to players (producer). ### NEXT ACTION **Weapon audio**: play the ADFRC EF88 shot/tail/reload sounds for the A-series weapons in place of Lyra's rifle cue. --- ## Session 032c — Blanket ADFRC permission recorded; map documentation written ### COMPLETED - **Recorded the producer's blanket 100% permission from the ADFRC mod team** in all three places that gate asset use, so they cannot disagree: - `CLAUDE.md` — the ADFRC rule is no longer a hold. It now states that ADFRC models, textures, animations, audio, configs and scripts are **cleared for free use in Southern Spear**, including converted and derived work and as visual/design reference, and that using them as game art is the expected case rather than the exception. - `Docs/LICENCE_REGISTER.md` (L-0021) — the **multi-author scope limit is resolved**. It previously held that "Tonnie" was not among the authors credited in the pack (Brucey, Exer, Growlor, Louetta, Quiggs, ADFU Team, ADF Re-Cut Team), so his grant could only ever have covered his own components and per-component confirmation was required. A blanket permission from the team as a whole closes it. - `Docs/PROJECT_AUDIT.md` (R-24) — **CLOSED** (branding tracked separately as R-27), from "OPEN (partially cleared)". - **Two limits are recorded as still standing**, because the mod team cannot lift them: **third-party and service marks** (Crye Precision G3, Ops-Core, PASGT, "Team Wendy", ADF camouflage and insignia belong to those companies and to the ADF), and **redistribution** (use in the project is cleared; pushing the assets through the repository is not, which is why `Art/ADFRC/*` stays git-ignored). APL-SA (Bohemia) is also untouched. - **Wrote the two missing map design documents** to the Dry River standard: `Docs/MAPS_SALTBUSH.md` and `Docs/MAPS_SELATCANAL.md`. Both are recorded honestly as **documented but not signed off**, because the evidence does not support signing them off — see DEFECTS FOUND. - **Diagnosed the in-game verification blocker.** The recorded claim that "windowed `-game` runs stall during module load" is **no longer true**: `Saved/Logs/SouthernSpear.log` holds a successful windowed `-game` run that loaded `/Game/Maps/L_SS_FrontEnd` in 0.42 s and wrote a 1600x900 screenshot. Added `Tools/run_map_capture.sh` to make that repeatable, and found two causes of silent failure along the way (see DEFECTS FOUND). ### FILES CHANGED Created: `Docs/MAPS_SALTBUSH.md`, `Docs/MAPS_SELATCANAL.md`, `Tools/run_map_capture.sh`. Modified: `CLAUDE.md`, `Docs/LICENCE_REGISTER.md` (L-0021), `Docs/PROJECT_AUDIT.md` (R-24), `Docs/CHANGELOG.md` (this entry). ### TESTING - **`bash Tools/run_map_capture.sh /Game/Maps/L_DryRiver_01` — NOT RUN to a passing result.** The script is written and two real bugs are fixed in it, but the run was interrupted before it completed. **No gameplay map has been captured and no in-game verification is claimed.** - `python Tools/Unreal/adfrc_animation_survey.py`, the VaultCache inventory and the `armis_f_data.pbo` header analysis — PASS (recorded in Sessions 032 / 032b). - All map figures in the two new documents are read from tool reports (`Build/objective_map_*`, `Build/deployment_tags.json`) and `Config/DefaultGame.ini`. None are estimated. ### ASSETS No new assets. No redistribution implication: the licence change authorises **use**, and `Art/ADFRC/*` and `Art/ADFRC_Player/*` remain git-ignored with only their `.md` files tracked. The two new documents are documentation and are safe to track. ### RISKS - **R-27 is now the only thing between the ADFRC material and release.** With R-24 closed, the branding substitution (Crye / Ops-Core / PASGT / Team Wendy / ADF camo) is the single remaining gate, and it is a build task rather than a pending approval. It must not be lost now that the surrounding block has lifted. - The captured map docs describe two maps that **should not be played as balanced content yet** (below). Documenting them is not endorsing them. ### DEFECTS FOUND - **Selat Canal has a 7/8 deployment split.** `Build/deployment_tags.json` records TeamOne 7, TeamTwo 8, against 8/8 on Dry River, Red Gum and Saltbush. Where a dead team rotates back to a start, the team with the extra start has a compounding advantage. Needs a producer decision: add a start, or record the acceptance. - **Selat Canal is the worst map in the set for navigation: 35 of 154 sampled grid points reachable (23%)**, against Saltbush's 37% and Dry River's verified full rebuild. All four round legs are walkable only because the nav pass **relocates** any objective it cannot reach — Saltbush's were moved up to 22.5 m, so "Stock Yards" may no longer sit on the stock yards. The objective positions on both maps are an artefact of navigation, not a design decision. - **Selat Canal deploys teams 70 m apart**, less than half Saltbush's 162 m and a quarter of Red Gum's 560 m, and unlike Dry River's deliberately equal 86 m opening this is documented nowhere. - **The `-game` harness had two silent-failure modes, both now fixed in `Tools/run_map_capture.sh`.** `-unattended` is a commandlet flag: a `-game` instance given it initialises the engine and then exits, which reads as a stall in the log. And Git Bash rewrites `/Game/Maps/...` into `C:/Program Files/Git/Game/Maps/...`, so the game loads nothing **while still writing a 2.87 MB screenshot of an empty frame** — a capture that looks like success and is not. The script now requires both a map-load line and an image before reporting PASS. - **Pre-existing:** `Docs/CHANGELOG.md` contains **two different entries both numbered Session 028** in committed history (line 1798 "Written ADFRC Authorisation Recorded", line 2137 "Website Typography and Front End"). Not introduced here and not renumbered, because renumbering committed history is riskier than the collision. - **A false alarm worth recording.** Midway through this session `Docs/CHANGELOG.md` was seen with 65 deletions and 0 additions against HEAD, which matched the R-26 silent-truncation signature, and it was restored from HEAD. On inspection this was a **transient mid-write state from the concurrent session** that committed the full entry moments later (`44e073d1`); nothing was lost. Recorded because the temptation in a shared checkout is to "fix" a sibling's in-progress edit, and here the right move was to check the log before acting. ### NEXT ACTION **Run `Tools/run_map_capture.sh` to completion on a gameplay map** — it has never produced a passing result, and until it does, every claim in this log about how the game looks remains an inference. --- ## Session 029 — 2026-09-27 — Fair Map Layouts: Spawn Tool, Cross-Map Objectives, Map Documents ### COMPLETED - **Reviewed the other agent's `MAPS_SALTBUSH.md` and `MAPS_SELATCANAL.md`** (sound, evidence-based) and acted on their findings; both are rewritten with the new measurements; `MAPS_REDGUM.md` is new. - **`layout_spawns.py`** (new, final authority on starts, all four maps): 8 starts per team on clear walkable ground (70 cm wall clearance, 2 m headroom, reachable from the deployment **and** to the centre objective), farthest-point spread, hidden from the enemy deployment first, facing the centre objective; deployments more than 200 m on foot from the centre objective slide in. Reports spacing, spawn exposure and each team's walk to every objective. - Selat Canal 7/8 → **8/8**; spawn exposure 0/64 on Red Gum, Dry River, Saltbush (Canal 11/64). - **`layout_objectives.py`** (new): objectives placed across the map at even-walk points (Red Gum flanks; Saltbush spread). The opening objective is now even on Saltbush (105 vs 104 m; was 46 vs 134 m) and within 11% on Red Gum (213 vs 240 m; was ~170 vs ~390 m). Red Gum's flank objectives are renamed North Paddock / South Paddock (they no longer stand on the bore pump and shearing shed). - **`build_objective_map.py`**: deployments chosen by `min(walk, 1.6 x straight line)` and objectives along the walking route (straight line gave a 70 m walled street; walk alone gave two banks 20 m apart). - Front-end Red Gum card text follows the renamed objectives. ### FILES CHANGED `Tools/Unreal/layout_spawns.py`, `layout_objectives.py` (new), `build_objective_map.py`; `Docs/MAPS_SALTBUSH.md`, `MAPS_SELATCANAL.md` (rewritten), `MAPS_REDGUM.md` (new); `SSMenuWidget.cpp`; maps `L_RedGum_01`, `L_DryRiver_01`, `L_Saltbush_01`, `L_SelatCanal_01`. ### TESTING | Test | Command | Exit | Result | Evidence | |---|---|---|---|---| | Guard | `python Tools/validate_architecture.py` | 0 | PASS | — | | Build | `Build.bat SouthernSpearEditor ...` | 0 | Succeeded | — | | Automation | `Automation RunTests SouthernSpear` | 255 | 30 Success; 1 Fail (parallel session's `TwoPlayerAuthoritySmoke`) | `Build/tests.log` | | Spawn layout | `layout_spawns.py` | 0 | ok=true, 8/8 on all maps; figures above | `Build/spawn_layout.json` | | Objective layout | `layout_objectives.py` | 0 | Red Gum flanks 3–4% at placement; Saltbush spread; Canal rejected (objectives within 28 m) | `Build/objective_layout.json` | | Saltbush match | 8 bots, 180 s, `-nullrhi` | 124 | Windmill captured by Team One (first capture on this map) | log | | Red Gum match | 8 bots, 180 s, before and after pull-in | 124 | **No capture** (contested stalemate) | log | | Canal / Dry River matches | 8 bots, 150 s | 124 | **No capture** | log | | Nav step-height experiment | canal reach probe | 0 | 36 → 41 of 164 reachable at 45 cm; reverted | log | ### ASSETS None new. ### RISKS - R-25 (new): **capture stalemate.** ADR-018 freezes a contested objective; respawning bots keep both teams present, so small or long maps can go a whole round without a capture. Needs a rules or bot decision (producer). - Selat Canal cannot host a fair three-objective sequence on its connected footprint (22% of samples). - R-24 (GitHub LFS push) still open. ### DEFECTS FOUND 1. Selat Canal 7/8 starts (other agent's review). 2. Opening objective much nearer Team One on the builder maps (fairness measurement). 3. Canal Team One starts on a disconnected nav island (layout debug; fixed: starts must reach the objectives). ### NEXT ACTION **Producer decides the capture stalemate** (R-25): majority capture, respawn waves, or attack/defend bot roles. --- ## Session 030 — 2026-09-27 — Native Resolution, Ray Tracing Default Off, Scoreboard with Ping, Dry River Real Cover; Locomotion Audit ### COMPLETED - **"Textures are horrible" — two causes, both fixed:** - the game rendered at **60.6% (1552x873)** of the producer's 2560x1440 display and upscaled: `sg.ResolutionQuality=0` means "project default", which UE 5.8 scales down on large displays. An unset value now becomes native 100% (`USSUserPrefsSubsystem`); Settings > Render resolution still lowers it. Verified: `stat unit` shows 100.0% (2560x1440). - hardware ray tracing rendered Nanite-converted rocks and the weapons **black** (Saltbush capture, A/B with the setting off). It now defaults off (Settings: "Hardware ray tracing (experimental)"). - Settings console variables now apply at game-override priority: ray tracing, ray-traced shadows and anti-aliasing were silently ignored below the project's own defaults (log: "SetByGameSetting ... ignored"). - A startup crash introduced and fixed in session (the settings object was created before Lyra's settings class loaded; now only once a world exists). - **Scoreboard (hold Tab)**: viewer-relative (own side "3 ACR · Friendly" first, "MAF · Opposing"), eliminations / deaths / assists / ping per player, team kill totals, local row highlighted. `USSScoreboardState` (Core) filled by `USSScoreboardSubsystem` (bridge, reflection); Lyra's base per-player scorer `B_ShooterGameScoring_Base` granted by the experience (not the team-deathmatch scorer, whose kill limit would end rounds). Verified live: kills and deaths count. - **Dry River real cover**: `dryriver_blockout.py` exports the designed cover (40 rocks, 20 trees, 30 scrub, fence) as `SS_MAP_DryRiver_01_Cover.csv` instead of baking low-poly cones and boulders into the terrain; `dress_dryriver_cover.py` reimports the terrain and places Rural Australia rocks, trees and grass trees at the same positions and sizes, and post-and-wire fences (119 posts) for the greybox and dressing fences. Nav rebuilt (`build_dryriver_nav.py` ok, path verified); spawns re-laid (exposure 3/64). - Dev tools: `ss.Debug.FollowBot` (third-person look at the nearest bot), `-SSShowScoreboard`, `-SSScoreDebug`. - **Locomotion audit** (`Docs/LOCOMOTION_AUDIT.md`) written before any movement change (producer brief). ### TESTING | Test | Command | Exit | Result | Evidence | |---|---|---|---|---| | Guard | `python Tools/validate_architecture.py` | 0 | PASS | — | | Build | `Build.bat SouthernSpearEditor ...` | 0 | Succeeded | — | | Automation | `Automation RunTests SouthernSpear` | 255 | 30 Success; 1 Fail (parallel session's `TwoPlayerAuthoritySmoke`) | `Build/tests.log` | | Render resolution | default launch, `stat unit` | 124 | 60.6% before, 100.0% after | screenshots | | Ray tracing A/B | Saltbush spawn, HWRT on/off | 124 | Black cliff and weapon with HWRT on; correct with it off | screenshots | | Scoreboard | Dry River 10 bots, `-SSShowScoreboard`, `-SSScoreDebug` | 124 | Rows, teams, K/D, local row | screenshot, log | | Dry River cover | `dress_dryriver_cover.py`, `build_dryriver_nav.py`, `layout_spawns.py` | 0 | 90 props + 119 posts; nav ok; exposure 3/64 | reports | ### RISKS - R-26 (new): third-person weapons are rotated ~90° in the soldiers' hands (bot-follow capture); covered by the locomotion audit. - R-24, R-25 open. ### DEFECTS FOUND 1. Render resolution 60% (producer: textures; stat capture). 2. Hardware ray tracing black meshes (A/B capture). 3. Settings console variables ignored (log). 4. Startup crash (introduced and fixed here). 5. Third-person weapon rotation (bot-follow capture). ### NEXT ACTION **Producer approves the locomotion roadmap** (`Docs/LOCOMOTION_AUDIT.md` §7), including adding Epic's Game Animation Sample. --- ## Session 031 — 2026-09-27 — ADF Soldiers from ADFRC (ADR-025), Locomotion Decisions (ADR-024), Audit Correction ### COMPLETED - **Producer decisions recorded:** ADR-024 (locomotion rebuild per `LOCOMOTION_AUDIT.md`) and **ADR-025** (real ADF look from the ADFRC set, overriding ADR-016's pattern rule for the AMCU textures; patches and flags stripped; release needs Defence permission or the CMECU swap: R-27). CLAUDE.md content rule amended. - **Audit F2 withdrawn after measurement**: with Lyra's rifle drawn over ours (`ss.Debug.ShowLyraWeapon`) the meshes overlap exactly; socket data agrees. The "rifle pointing up" is Lyra's jog pose (F5). - **3 ACR soldiers rebuilt from ADFRC gear** (producer: "very low quality, don't replicate Australian soldiers"): - `Tools/Blender/adfrc_gear_rig.py`: fits Arma gear to the UE5 mannequin. It places the gear with the Memory LOD joints, re-poses the limbs onto the mannequin's joints with a segment-distance rig, transfers skin weights from the mannequin body, and encodes texture and rvmat names in the material slots. Rigid mode is for helmets. Arma helper faces, BIS skin, flags and patches are dropped. - Found: the other converter's `.blend` files carry no bone weights, and the gear files use Arma's true character space, which is offset from the uniform file (`SS_GEAR_SPACE`, measured). - Kit: Crye G3 combat uniform in **AMCU** with gloves and boots, **Ops-Core helmet** (AMCU cover, Peltor headset), **TBAS T5 plate carrier** (AMCU carrier and pouches, belt, holster). Multicam pouches use AMCU or coyote variants, or flat coyote. - `Tools/Unreal/setup_adf_soldier.py`: imports onto Lyra's `SK_Mannequin`; `M_SS_GearPBR` (colour, normal, SMDI; skeletal); matte non-metallic fabric (the SMDI gloss rendered chrome-white). `setup_soldiers.py`: 3 ACR = head + uniform + carrier + helmet on the animated skeleton (audit F3 for the friendly side). - Bugs found on the way: a rigid helmet group named "head.001" (an Arma selection shared the name), and `save_loaded_asset` skipping Python-set materials (not dirty). - The parallel session's MAF materials (`Characters/Materials`, `Characters/Textures` and their scripts) are committed with the soldier Blueprint that references them. ### TESTING | Test | Command | Exit | Result | Evidence | |---|---|---|---|---| | Gear fit | Blender renders (mannequin overlay, front and side) | 0 | Uniform, helmet and carrier on the mannequin | renders (scratch) | | Textures | Blender textured render | 0 | AMCU sleeves and trousers, plain torso (as authored); UVs correct | render | | Import | `setup_adf_soldier.py` | 0 | ok=true; every slot has colour, normal and SMDI; materials persisted (verified) | `Build/adf_soldier_setup.json` | | In game | Dry River, `ss.Debug.FollowBot` | 124 | ADF soldier animated: helmet on head, AMCU uniform, carrier, matte fabric | captures | | Weapon alignment | `ss.Debug.ShowLyraWeapon` | 124 | Lyra's and our meshes coincide | captures | ### RISKS - **R-27 (new):** AMCU and ADF kit are Commonwealth designs; a commercial release needs Defence permission or the CMECU swap (ADR-025). - The uniform's upper sleeves are slightly puffy after the re-pose; the MAF side still uses the Fab parts. ### NEXT ACTION **S1: `ASSCharacter` + `USSCharacterMovementComponent`** (tactical speeds, momentum, replicated stances; ADR-024). --- ## Session 033 — Map playability measured for the first time; the reference map fails its own rules Run in parallel with the other session's soldier and locomotion work. Nothing outside `Tools/Unreal/`, `Docs/MAPS_PLAYABILITY_AUDIT.md` and this entry was touched, and no map asset was written. ### COMPLETED - **Wrote `Tools/Unreal/audit_map_playability.py`**, a read-only audit that measures the Dry River design rules against every map. It loads a map, builds nothing, places nothing and saves nothing, so it is safe to run while another tool is dressing the same level. It reports walkable ground, open-crossing distance, hard:soft cover ratio, cover density, close-quarters and long-range sightlines, per-objective cover and overwatch, walk parity between the two teams, and spawn exposure. - **Ran it on all four maps**: `ok: true`, 0 errors, report at `Build/map_playability.json`. - **Wrote `Docs/MAPS_PLAYABILITY_AUDIT.md`** with the scoreboard, the per-map findings and the method. Headline results, all measured: - **Dry River fails its own design rules.** 13% of the ground has no cover within 30 m (the document promises under 20 m), 44 blocking cover props on the whole map against a documented ~120, 0.09 props per walkable cell, a 291.5 m maximum sightline against a 220 m limit, the Farmstead 58% walk-imbalanced between teams, and **35 of 64 spawn pairs can see each other**. - **Red Gum is not playable as it stands.** 1020 x 1020 m of paddock whose navigation volume covers **17%** of it, **12 hard and 0 soft cover objects in the entire map**, 50% of ground with no cover within 30 m, 310 m median sightline, and 0-1 cover positions within 20 m of any objective. - **Saltbush is the best map in the project** (7 rules pass): p90 open crossing 16.0 m, 0 of 64 spawn pairs exposed, 95% nav coverage. Its two failures are the 35% and 22% walk imbalance on Stock Yards and Dry Dam, and a 1:9.97 hard:soft cover ratio where the design wants 1:3. - **Selat Canal has the best geometry and the worst fairness.** 2.0 m p50 open crossing and 51.7 m median sightline make it the tightest map in the project, but all three objectives fail walk parity at **63%, 42% and 75%**, on the project's only Special Forces map. ### FILES CHANGED Created: `Tools/Unreal/audit_map_playability.py`, `Docs/MAPS_PLAYABILITY_AUDIT.md`. Modified: `Docs/CHANGELOG.md` (this entry). **Left uncommitted on purpose** — this file also carries another session's uncommitted edits, and staging it would stage theirs with it. ### TESTING - `UnrealEditor-Cmd.exe SouthernSpear.uproject -nullrhi -unattended -ExecutePythonScript=Tools/Unreal/audit_map_playability.py` — **PASS**, `ok: true`, 0 errors, all four maps, `Build/map_playability.json`. Three maps in one run (~9 min); Selat Canal separately with `SS_MAPS=L_SelatCanal_01 SS_OUT=mp_canal.json` after a loop bug was fixed, then merged. - The script is **not** a game run. No map was played, no screenshot captured, and `run_map_capture.sh` is still unproven. ### ASSETS None. No asset was created, imported, modified or licensed. The audit is measurement only. ### RISKS - **R-34 — the map set is documented as finished and is not.** `MAPS_DRYRIVER.md`, `MAPS_SALTBUSH.md` and `MAPS_SELATCANAL.md` read as design intent. Measured, Dry River misses its own cover and sightline rules, Red Gum has no cover and a navmesh on a sixth of the map, and Selat Canal's objectives are badly lopsided. The three "documented, not signed off" notes are correct and must not be relaxed until the maps measure clean. - **R-35 — `Tools/Unreal/layout_spawns.py` under-reports spawn exposure.** It traces each candidate start to the enemy **centroid**, one point, rather than to the enemy starts. It reported 3/64 for Dry River; tracing all 64 pairs gives 35/64. Every `exposed_pairs` figure in `Build/spawn_layout.json` is optimistic. Not fixed here because the file belongs to the concurrent session's work. ### DEFECTS FOUND Four defects in the audit tool itself, all found by running it, and all the kind that would have produced confident nonsense: 1. **`get_actor_bounds` returns (origin, extent), not (min, max).** Treating the origin as a corner classified every prop as a kerb and reported **0 cover objects on a map with 290 dressing props**. 2. **Projecting to navigation from the middle of the map's bounding box** silently misses on any map with a tall z range. It reported Dry River at 31% nav coverage; measured from the traced ground height the figure is 100%. The first version of this audit published that 31%. 3. **`find_path_to_location_synchronously` floods the navmesh** when the goal is on an island the start cannot reach. On one Dry River objective that took **nine minutes** and nothing can interrupt it. Replaced with a projected polyline, which reports unreachable by failing rather than by stalling. 4. **A `while len(pairs) < 3000` loop cannot terminate on a small map.** Selat Canal has 31 walkable points, so 465 distinct pairs exist and the loop spun forever. Now bounded by the pair count. Also corrected in the tool's own method: the "close quarters" rule was unmeasurable as written, because a 10 m sample grid means two sampled points are never within 5 m of each other. It now probes 3, 5, 8 and 15 m in eight directions from every sample point. ### NEXT ACTION **Fix the Selat Canal objective placement** — three objectives, all three between 42% and 75% walk imbalanced, on the only Special Forces map — and make the placement refuse to save a lopsided result rather than silently relocating it. --- ## Session 034 — 2026-09-28 — Damage Model, Blood, Bullet Penetration (ADR-026) ### COMPLETED - **Damage model** (`Tools/Unreal/setup_damage_model.py`): hit zones PM_SS_Head/Torso/Limb (tags SS.Zone.*), set per physics body by bone name in `ASSCharacter::BeginPlay` (head 3, torso 7, limb 12 bodies); per-weapon `B_SS_WeaponInstance_` with zone multipliers and damage flat to 300 m. Rifles: head one hit, torso 3, limbs 5; A25 torso 2. Crash on first death (GC freed the zone materials) fixed: `ZoneMaterials` UPROPERTY. - **Blood**: `ASSCharacter::HandleGameplayCue` on `GameplayCue.Character.DamageTaken` spawns the VFX-pack blood burst at the hit point along the shot (clients only). - **Hero class (ADR-026, D-08)**: Lyra's `B_Hero_Default` reparented to `ASSCharacter`; `B_SS_Hero*` and `HeroData_SS` deleted (copies broke Lyra's class-identity casts: bots never fired). - **Bullet penetration (ADR-026, D-09)**: a marked hook in `ULyraGameplayAbility_RangedWeapon::TraceBulletsInCartridge` (one penetration; hits beyond carry the damage lost in `PenetrationDepth`, which replicates with target data) and in `LyraDamageExecution` (applies it). `USSBallisticsSubsystem` (bridge) measures thickness with a reverse trace on the blocking component; `FSSPenetrationRules` (Core): up to 20 cm, 25-75% damage lost; never terrain or pawns. - `Docs/LYRA_ADOPTION.md` D-08, D-09; `Docs/DECISION_LOG.md` ADR-026. - Numbered 034: the parallel session's uncommitted changelog already uses 032, 032b, 032c and 033. ### FILES CHANGED Core `SSBallistics.*`, `Tests/SSBallisticsTests.cpp`, `SSNativeGameplayTags.*`; bridge `SSCharacter.*`, `SSCharacterMovementComponent.*`, `SSBallisticsSubsystem.*`, `Build.cs`; Lyra `LyraGameplayAbility_RangedWeapon.*`, `LyraDamageExecution.cpp`; `Tools/Unreal/setup_damage_model.py`, `setup_tactical_movement.py`; content listed in ASSETS. ### TESTING | Test | Command | Exit | Result | Evidence | |---|---|---|---|---| | Guard | `python Tools/validate_architecture.py` | 0 | PASS | console | | Build | `Build.bat SouthernSpearEditor Win64 Development` | 0 | Succeeded | console | | Automation | `Automation RunTests SouthernSpear` | 255 | 33 Success, 1 Fail (`TwoPlayerAuthoritySmoke`, the parallel session's test, failing before this session) | `Docs/evidence/S034_tests.txt` | | Live | Dry River, 8 bots, 150 s round, `LogTemp=Verbose` (timeout kill) | 124 | 3 penetrations (1.0 and 9.3 cm surfaces), 175 blood spawns, no crash | `Docs/evidence/S034_penetration_live.txt` | | Earlier live | Dry River, 8 bots | 124 | 16 kills in about 2 min after the GC fix; 42 blood spawns in 90 s | session log | NOT RUN: direct measurement of reduced damage through a wall (Lyra's execution has no per-hit damage log); visual confirmation of blood (two rendered captures did not show a splat clearly); multiplayer client-to-server penetration check. ### ASSETS PM_SS_Head/Torso/Limb, B_SS_WeaponInstance_* (copies of Lyra's weapon instances), WID_SS_* InstanceType, Lyra `B_Hero_Default` (reparented), experience and IMC updates; blood uses the already-registered VFX pack. ### RISKS - **R-36 (new; first numbered R-28, which was already taken):** two Lyra departures (D-08, D-09) must be re-applied on any Lyra update. - Penetration thickness is geometric only (no per-material table); thin rock edges can be shot through. ### DEFECTS FOUND - Ragdoll crash at first death: zone physical materials garbage-collected (live bot run, fatal assert). - Bots never fired with a copied hero class (live bot run: zero kills). ### NEXT ACTION **Measure through-cover damage**: a scripted test that fires through a 5 cm board at a target and checks the health lost. --- ## Session 035 — 2026-09-28 — Full Test Suite Green; Penetration Hook Measured; Parallel Session's Work Committed ### COMPLETED - **All 35 SouthernSpear tests pass (exit 0)**, the first fully green run in several sessions. The parallel session's `TwoPlayerAuthoritySmoke` had two causes, found from its callstacks: - **ours**: `USSSettingsSyncSubsystem` pushed volumes into Lyra's settings in editor test worlds, where Lyra cannot load its audio control-bus mix (it needs `GEngine->GetCurrentPlayWorld()`): three `bSoundControlBusMixLoaded` ensures. Volumes are now pushed only with an audio device and a play world; - **Lyra's loading screen** adding a widget to a headless test viewport (`ViewportOverlayWidget.IsValid()`). Test runs now pass `-NoLoadingScreen` (Lyra's own switch): CLAUDE.md, README, CI. - **CI ran almost no tests**: it filtered on `SouthernSpear.Unit/Integration/Network/Leak`, of which only `Network` exists. Now `RunTests SouthernSpear`, with `-nosound -NoLoadingScreen`. - **Through-cover measurement** (Session 034 NEXT ACTION): `SouthernSpear.Bridge.Ballistics.PenetrationHook` calls the hook Lyra's weapon actually uses (so it also proves the bridge registered it) against engine-cube boards: a 5 cm board is penetrated with 37.5% damage lost (an A88 torso hit 38 → 23.75) and the trace resumes just beyond it; a 40 cm block stops the bullet. `LyraBulletPenetration::GetHook` exported for the test. - Changelog repaired: the parallel session's uncommitted copy had dropped Sessions 029–031; restored, and its Session 033 kept. Risk renumbered: Session 034's "R-28" was taken, now **R-36** (added to PROJECT_AUDIT). - Committed the parallel session's finished, uncommitted work: Session 033 entry, ASSET_REGISTER, SOURCED_ASSET_REVIEW, DECISION_LOG execution note, TEST_PLAN, the smoke test and its `EngineSettings` dependency, README. Untracked art folders (`Art/Weapons/{A88/New,AKM,C4A1,PKM,_Optics}`, `Content/AUG`, `Content/SouthernSpear`, `Docs/images/conceptart*.png`) handled below. - **Untracked folders resolved:** the blocked or reference-only weapon sources (`Art/Weapons/{A88/New,AKM,C4A1,PKM}`), the Fab optic export `Art/Weapons/_Optics` (no republishing) and the raw CC BY scan sources `Content/SouthernSpear/Vendor` are git-ignored; `Content/AUG` (ADFRC audio) committed as LFS pointers. ASSET_REGISTER 4.9g added (AUD-AUG-001, CH-ADF-001, GP-DMG-001). `Docs/images/conceptart{1,2}.png` left untracked: their source is unknown. ### FILES CHANGED `SSSettingsSyncSubsystem.cpp`, `Tests/SSPenetrationHookTests.cpp` (new), Lyra `LyraGameplayAbility_RangedWeapon.h` (export), `.github/workflows/build.yml`, `CLAUDE.md`, `README.md`, `Docs/TEST_PLAN.md`, `Docs/PROJECT_AUDIT.md`, this file. ### TESTING | Test | Command | Exit | Result | Evidence | |---|---|---|---|---| | Guard | `python Tools/validate_architecture.py` | 0 | PASS | console | | Build | `Build.bat SouthernSpearEditor Win64 Development` | 0 | Succeeded (first try failed to link: `GetHook` not exported; fixed) | console | | Network smoke alone, before the flag | `RunTests SouthernSpear.Network` | 255 | Fail: only the loading-screen ensure left after the audio fix | `Build/smoke.log` (not retained) | | All tests | `UnrealEditor-Cmd ... -nosound -NoLoadingScreen ... "Automation RunTests SouthernSpear;Quit"` | 0 | 35/35 Success | `Docs/evidence/S035_tests.txt` | NOT RUN: the CI workflow itself (self-hosted runner); damage through a board in a live match (the execution's `1 - PenetrationDepth` factor is verified by reading, not by a measured health change); rendered blood check. ### ASSETS None. ### RISKS - R-36 (from Session 034) recorded in PROJECT_AUDIT. - The concept art images have no recorded source. ### DEFECTS FOUND - Settings sync triggered Lyra audio ensures in test worlds (callstack in the smoke-test log). - CI test filter matched only one of four test groups (reading the workflow). - Risk ID collision R-28 (grep of PROJECT_AUDIT). ### NEXT ACTION **Wire the AUG audio (`/Game/AUG/Sound`) into the A88 family's fire, tail and reload cues**, replacing Lyra's rifle sounds. --- ## Session 036 — 2026-09-28 — A-Series Rifles Fire with the AUG Recordings ### COMPLETED - **Rifle fire audio from the ADFRC AUG recordings** (AUD-AUG-001, L-0021), in C++ because SoundCues cannot be authored headlessly (Session 032b). `ASSCharacter` handles `GameplayCue.Weapon.Rifle.Fire` on every client: - a random close shot (3 variants, ±3% pitch); unspatialised for the human shooter, 60 m falloff for everyone else; - a distant shot layer (3 variants) heard to 600 m, and the `tailMeadows` outdoor tail; - Lyra's `MSS_Weapons_Rifle2_Fire` muted. Found by logging: Lyra's weapon Blueprint spawns that MetaSound **owned by the pawn**, not the weapon actor; it is matched by sound name so footsteps stay. - Pistol and shotgun cues keep Lyra's sounds. All rifles (A88 family, A4, A416, A89, A25) share the AUG set for now. - Defect caught before shipping: AI pawns count as locally controlled on the server, so bot shots would have played unspatialised to the player. The 2D close shot is now for `IsLocallyControlled() && IsPlayerControlled()` only. ### FILES CHANGED `SSCharacter.h/.cpp`, this file, `Docs/evidence/S035_rifle_audio_live.txt`. ### TESTING | Test | Command | Exit | Result | Evidence | |---|---|---|---|---| | Guard | `python Tools/validate_architecture.py` | 0 | PASS | console | | Build | `Build.bat SouthernSpearEditor Win64 Development` | 0 | Succeeded | console | | All tests | `UnrealEditor-Cmd ... -nosound -NoLoadingScreen ... "Automation RunTests SouthernSpear;Quit"` | 0 | 35/35 Success | `Build/tests_036.log` (not retained) | | Live, audio device on | Dry River, 8 bots, `LogTemp=Verbose`, no `-nosound` (timeout kill) | 124 | 257 rifle shots: AUG layers played, Lyra's MetaSound muted on 257/257, all spatialised (bots), no crash | `Docs/evidence/S035_rifle_audio_live.txt` | NOT RUN: **listening**. No one has heard the result; the mix levels (close 1.0, distant 0.7, tail 0.3–0.45) and whether Lyra's first-shot trigger leaks before the mute are unverified by ear. Reload and dry-fire sounds not wired. ### ASSETS Uses AUD-AUG-001; nothing new. ### RISKS - Mix levels unverified by ear (see NOT RUN). ### DEFECTS FOUND - Bot shots would have played as the player's own (code reading after the first live log showed `local=1`). - Mute target was wrong at first: the MetaSound is on the pawn, not the weapon actor (live log, `muted=0` on 319 shots). ### NEXT ACTION **Play a match with sound and judge the rifle audio by ear** (levels, first-shot leak, distance), then tune the three volumes. --- ## Session 037 — 2026-09-28 — Pre-Test Pass: Minimap Edge, First-Person Rifle Placement, Game Icon ### COMPLETED - Rendered review of the current build (first person, a followed bot, front end) before the next playtest. - **Minimap showed a black band** (about 30% of the corner map) near the Dry River spawns: the view reached past the edge of the ground. The corner map's centre is now clamped to the playable area (objectives + player starts); measured on Dry River, a +15 m margin still left a strip, so there is no margin. The player arrow is clamped inside the frame. - **First-person rifle too large**: hip position `ss.FP.Hip` 38 13 -16 → **48 16 -20** (further forward, right and lower); aiming is unchanged (it is computed from the sight socket). - **Game icon**: the producer's `Docs/images/SouthernSpear.ico` (10 sizes, 16–256 px) is installed as `Build/Windows/Application.ico` by `Tools/build_game_icon.py` (the tool still generates one from the logo if the file is absent). - Committed the producer and parallel-agent images: `SouthernSpear.ico`, `conceptart1.png`, `conceptart2.png`. ### FILES CHANGED `SSMinimapWidget.h/.cpp`, `SSFirstPersonSubsystem.cpp`, `Tools/build_game_icon.py`, `Docs/images/*`, evidence, this file. ### TESTING | Test | Command | Exit | Result | Evidence | |---|---|---|---|---| | Guard | `python Tools/validate_architecture.py` | 0 | PASS | console | | Build | `Build.bat SouthernSpearEditor Win64 Development` | 0 | Succeeded | console | | All tests | `UnrealEditor-Cmd ... -nosound -NoLoadingScreen ... "Automation RunTests SouthernSpear;Quit"` | 0 | 35/35 Success | `Build/tests_037.log` (not retained) | | Rendered | Dry River `-game -windowed -SSShotAt=30`, before and after | 124 | Minimap fully filled; rifle smaller | `Docs/evidence/S037_before_fp.jpg`, `S037_after_fp.jpg` | | Icon | `python Tools/build_game_icon.py` then `cmp` | 0 | identical to the supplied file | console | NOT RUN: minimap on Red Gum, Saltbush and Selat Canal; a packaged exe showing the icon. ### ASSETS `Docs/images/SouthernSpear.ico` (producer-supplied). Concept art: source not recorded (committed at the producer's request). ### RISKS - The rifle optic texture carries a maker's mark (visible in first person): third-party branding to strip before release (R-27). ### DEFECTS FOUND - Minimap black band (rendered review). Oversized first-person rifle (rendered review). ### NEXT ACTION **Playtest with sound**: judge the rifle audio, the new first-person placement and the minimap on all four maps. --- ## Session 038 — 2026-09-28 — Window Icon, Game Splash, Hit Direction, Minimap Fixes; GPT-6 Findings ### COMPLETED - **Icon (producer: "still the Unreal logo")**: development runs are `UnrealEditor.exe -game`, whose embedded icon is Unreal's; `Application.ico` only reaches a packaged `SouthernSpear.exe`. New `USSWindowIconSubsystem` (SouthernSpearUI) sets the window and class icons from `Build/Windows/Application.ico` (fallback `Docs/images/SouthernSpear.ico`) once the game window exists (a first attempt at map load ran before the window existed and did nothing). - **Splash**: `Content/Splash/Splash.bmp` (960×240, from `Docs/images/header.png`) replaces Unreal's game splash (engine lookup: project `Content/Splash/Splash.*` first). Editor splash unchanged. - **Hit direction**: `USSLocalHudState` gains `LastHitTime`/`LastHitFrom` and `HitBearing()`; `ASSCharacter` fills them from Lyra's damage cue for the local player; the HUD shows a clay triangle on a 150 px ring pointing at the shooter, fading over 1.5 s (alongside the existing clay flash). - **Defect found: the UI fonts have no ▲ glyph** (Barlow Condensed, Inter: checked with fontTools), so the minimap's player arrow had always rendered as nothing. `SSGlyphTextures::Triangle()` (Core, runtime texture) now draws both arrows; the minimap arrow has an ink outline (pale brass on sand was invisible). - **Minimap view**: Session 037's clamp to objectives + player starts pinned the arrow to the edge, because deployment spawns lie outside that box. Replaced with downward ground traces at the view edges on each re-render, pulling the view inward (at most half a view) where there is no ground; re-render is keyed to pawn movement, not the view centre. - **Blood** no longer spawns on the local player's own body (it filled the first-person view with pale sprites, seen in a capture). - Dev tool: `-SSExecAt= -SSExec="cmd1|cmd2"` runs console commands as the local player after the pawn exists. - **GPT-6 (producer's parallel agent) reported, no git writes**: 16 local commits ahead of origin; `git lfs push --dry-run` lists 258 candidate objects; `git lfs fsck` over the range passes (local objects valid); first LFS-introducing unpushed commits 44e073d1, 0a20459c, 7c9b0210. The exact GH008 object list is unknown without the full rejection output. Branding audit: 207 candidate PNGs by filename, not inspected visually, **no edits made**. ### FILES CHANGED Core `SSLocalHudState.h`, `SSGlyphTextures.h` (new), `Tests/SSHudTests.cpp` (new); bridge `SSCharacter.cpp`; UI `SSWindowIconSubsystem.*` (new), `SSPlayerHudWidget.*`; ObjectivesUI `SSMinimapWidget.*`, `SSObjectiveHudSubsystem.*`; `Content/Splash/Splash.bmp`; `CLAUDE.md`; evidence; this file. ### TESTING | Test | Command | Exit | Result | Evidence | |---|---|---|---|---| | Guard | `python Tools/validate_architecture.py` | 0 | PASS | console | | Build | `Build.bat SouthernSpearEditor Win64 Development` | 0 | Succeeded (one shadowing error fixed on the way) | console | | All tests | `UnrealEditor-Cmd ... -nosound -NoLoadingScreen ... "Automation RunTests SouthernSpear;Quit"` | 0 | 36/36 Success (new `Core.Hud.HitBearing`) | `Build/tests_038.log` (not retained) | | Icon | windowed `-game`, front end | 124 | `SSWindowIcon applied=1` (WM_GETICON returns the set icon) | log | | Hit arrow | Dry River, `-SSExecAt=14 -SSExec="EnableCheats|DamageSelf 20" -SSShotAt=14.2` | 124 | health 100→80, `SSHitDir` logged, arrow drawn (points behind: self-damage source is the pawn) | `Docs/evidence/S038_hit_arrow.jpg` | | Minimap | Dry River, Red Gum, Selat Canal, `-SSShotAt=20` | 124 | Dry River and Red Gum filled, arrow visible; Selat Canal mostly black (see RISKS) | `Docs/evidence/S038_minimaps.jpg` | NOT RUN: the taskbar button and the splash window were not seen (no desktop capture by rule); a real shooter's bearing in a firefight (the idle player was not hit in 75 s); Saltbush minimap. ### ASSETS `Content/Splash/Splash.bmp` from the producer's `header.png` (`Docs/evidence/S038_splash.jpg`). ### RISKS - Selat Canal's minimap is mostly black: the map is built over void and water, and the minimap pass skips translucency. - GH008 push block still open; decision needed (upload LFS objects after a rights check, rewrite history, or pointer-only remote). ### DEFECTS FOUND - Missing ▲ glyph: minimap player arrow never rendered (screenshot zoom, then fontTools cmap check). - Session 037 minimap clamp hid the player at the edge (rendered capture). - Own-body blood in the first-person view (rendered capture). ### NEXT ACTION **Producer playtest**: confirm the taskbar icon and splash on launch, and judge hit arrow, rifle audio and first-person placement. --- ## Session 039 — 2026-09-28 — Class Selection with a 3D Soldier and Weapon Preview; Smart App Control Block ### COMPLETED - **Class screen redesigned (producer request)**: classes down the left (number, name, role, kit), a DEPLOY button, and on the right a live 3D preview of the selected soldier and weapon as they appear in game, with the class name and kit. Clicking a class now selects and previews it; DEPLOY confirms (it used to deploy on click). - **Preview stage** (`USSClassSelectWidget`, SouthernSpearUI; no new dependencies): spawned 2.5 km above the map while the screen is open, destroyed on close. Lyra's invisible mannequin body plays the rifle hip-fire idle; the friendly soldier parts are read from `B_SS_Soldier`'s `FriendlyParts` by reflection (SouthernSpearTeam is not a UI dependency) and follow it; the class weapon (`B_SS__Weapon`, standard or Special Forces kit) is attached as Lyra attaches it (`weapon_r`, yaw -90). Scene capture with a show-only list, sky and clouds off, three studio point lights, a lit backdrop wall and a floor disc (engine shapes, palette tints); slow sway around a three-quarter view. - Found on the way (rendered captures): the sky rendered behind the show-only capture; point lights over-exposed the model; dark rifles vanished against black (backdrop added); the backdrop plane faced away. ### FILES CHANGED `SSClassSelectWidget.h/.cpp`; evidence; this file. ### TESTING | Test | Command | Exit | Result | Evidence | |---|---|---|---|---| | Guard | `python Tools/validate_architecture.py` | 0 | PASS | console | | Build | `Build.bat SouthernSpearEditor Win64 Development` | 0 | Succeeded | console | | Rendered | Dry River windowed `-game -SSShotAt=12` (class screen opens on deploy) | 124 | Layout as requested; 3 ACR soldier in AMCU, Ops-Core, TBAS; A88 held (seen in a brightened zoom) | `Docs/evidence/S039_class_select.jpg`, `S039_class_preview_weapon.jpg` | NOT RUN: automation tests after the final build, and any rendered check of the last four builds: **Windows Smart App Control blocked `UnrealEditor-SouthernSpearUI.dll`** (GetLastError 4551; Code Integrity events 3077/3118, `Docs/evidence/S039_smart_app_control.txt`). Two earlier builds this session were blocked, then allowed after a code change; from 10:31 every rebuild was blocked, including one of the exact source that had loaded. The committed source is that last-loaded version. Selecting other classes (weapon swap) was not exercised on screen. ### ASSETS None new (engine BasicShapes; existing soldier, weapon and Lyra animation assets). ### RISKS - **R-37 (new): Smart App Control blocks locally built, unsigned module DLLs**, so the game does not start (SouthernSpearUI fails to load). Not changeable by an agent (security setting). Producer decision: turn Smart App Control off (Windows Security → App & browser control; Windows only allows turning it back on after a reset) or sign the binaries with a trusted certificate. ### DEFECTS FOUND - Sky behind the show-only capture; over-exposed studio lights; rifle invisible against black; backdrop facing away (all from rendered captures). ### NEXT ACTION **Producer: resolve the Smart App Control block (R-37)**, then rebuild, run the tests, and check the class screen with each class. --- ## Session 040 — 2026-09-28 — Soldiers Slimmed (Uniform Fit Cap), Fabric Material, Crisp Uniform Textures; Class Preview Polish ### COMPLETED - **"Player models way too bulky"** (producer). Measured before changing: `Tools/Blender/probe_gear_fit.py` (uniform vertex distance to the UE5 mannequin's skin, per region) gave medians of 1.3–1.6 cm (torso, forearm, thigh) but 3.0 cm on the upper arm; a Blender silhouette render (`render_gear_silhouette.py`) showed ~12 cm cloth "wings" at both shoulders and baggy thighs, a re-pose artifact of `adfrc_gear_rig.py` (Arma shoulder cloth left behind when the arms move onto the mannequin). New optional step 2b, `SS_GEAR_CAP` (cm): cloth further than the cap from the skin is pulled in, keeping 15% of the excess so pockets and folds stand proud. Compared 3.5 and 2.5 cm; **2.5 cm adopted** (8538 of 23187 vertices moved; worst 12.3 cm before). Plate carriers are not capped (their offset is real). The MAF side uses the same uniform mesh. - **"Textures don't look good"**: resolution was not the cause (uniform colour 4096², normal/SMDI 2048²). Two causes found: 1. fabric used one flat roughness (0.85) with the Arma SMDI switched off (it rendered chrome under the weapon mapping): new **`M_SS_FabricPBR`** (`setup_adf_soldier.py`): roughness 0.95→0.62 from the SMDI gloss, specular 0.15 + 0.35 × SMDI specular, adjustable normal strength; hard items keep the weapon material; 2. the first capture after the change showed the AMCU pattern smeared: script-built materials carry no texture-streaming data, so textures stayed at low mips. Soldier textures are now resident (never stream), capped at 2048 px (the fix the weapons already use). - Class preview: the soldier (and his weapon) turns, the backdrop and lights stay put (the wall edge had swung into view); right-hand three-quarter view where the rifle is carried; larger backdrop. Window icon confirmed in the producer's screenshot. ### FILES CHANGED `Tools/Blender/adfrc_gear_rig.py` (fit cap; rebuild command uses `SS_GEAR_CAP=2.5`), `probe_gear_fit.py` and `render_gear_silhouette.py` (new); `Tools/Unreal/setup_adf_soldier.py` (fabric master, resident textures); `Art/Characters/ADF/SK_ADF_Uniform_G3.fbx`; soldier content under `SSExp_ObjectiveAssault/Characters/ADF`; `SSClassSelectWidget.cpp`; evidence; this file. ### TESTING | Test | Command | Exit | Result | Evidence | |---|---|---|---|---| | Fit, before | `blender -b --factory-startup -P Tools/Blender/probe_gear_fit.py -- ` | 0 | uniform upper arm median 3.0 cm, p90 5.6 cm; others 1.3–2.4 cm | console | | Silhouette | `render_gear_silhouette.py`: original, cap 3.5, cap 2.5 | 0 | shoulder wings largely removed, legs slimmer | `Docs/evidence/S040_uniform_fit_before_cap35_cap25.jpg` | | Re-import | `UnrealEditor-Cmd ... setup_adf_soldier.py` | 0 | `ok: true` | `Build/adf_soldier_setup.json` | | Rendered | class screen `-SSShotAt=10`; bot follow `-SSShotAt=30` | 124 | slimmer soldier, crisp AMCU, fabric sheen | `Docs/evidence/S040_soldier_before_after.jpg`, `S040_soldier_ingame.jpg` | | Guard, build, tests | as in CLAUDE.md | 0 | PASS; Succeeded; 36/36 | `Build/tests_040.log` (not retained) | NOT RUN: Arma normal-map green-channel convention check (DirectX vs OpenGL is still unverified); texture memory measurement of the resident soldier set; the MAF look re-checked in a capture. ### ASSETS `M_SS_FabricPBR` (new, original material); new SMDI texture imports for existing ADFRC items (L-0021, CH-ADF-001). ### RISKS - Resident soldier textures raise memory use (bounded by the 2048 px cap); not measured yet. ### DEFECTS FOUND - Uniform shoulder "wings" from the gear re-pose (silhouette render). Flat fabric shading (material review). Low-mip uniform textures from script-built materials (rendered capture). Class-preview wall edge (producer screenshot). ### NEXT ACTION **Verify the Arma normal-map green channel** (render a seam-heavy fabric patch with and without the flip under a low light) and set the import flag accordingly for soldier and weapon textures. --- ## Session 041 — 2026-09-28 — Weapon Optics Re-Placed Onto The Sight Line; Loadout Renders Rebuilt ### COMPLETED - **Optics moved off the buffer tubes.** `Tools/Blender/adfrc_weapon.py` placed a mounted optic at `eye.x / 2` — the midpoint of the trigger (the mesh origin) and the REAR sight — so every scope on the A4, A416 and A25 sat over the buffer tube or the stock, and the EF88's Spectr sat off the back of its rail. The rule now uses the model's own memory points: the midpoint of `front_sight_axis` and `rear_sight_axis` at eye height, falling back to the bullpup `op_axis` proxy when a weapon has no iron sights. - **The four exported meshes corrected** by `Tools/Blender/fix_weapon_optics.py` (the optic source blends under `Art/ADFRC_BLEND/adfrc_optics_ss/` are no longer on disk, so the meshes were edited rather than re-exported; the FBX settings match the exporter's exactly and every non-optic part is byte-identical afterwards): - A88 Spectr +0.131 m · A4 TA31 +0.305 m · A416 TA31 +0.245 m · A25 TA648 +0.309 m - `manifest.json` for each weapon records the new `optic.centre_m` and a `placement` note. - **Loadout renders rebuilt** (`Tools/Blender/render_weapons.py`). The previous pipeline guessed textures by filename across the whole ADFRC tree, which picked up other weapons' maps, and it wired Arma's transparent `_CA` overlay maps to Base Color — that is why the scopes came out as black holes and white reticle blobs. Materials now come from each weapon's export manifest (colour + NOHQ normal + SMDI gloss/grime), with lens and reticle slots built as what they are. - **Lighting and framing fixed**: the area lights were placed with hand-written Euler angles that aimed them away from the weapon; they are now derived from the offset like the camera. Power is scaled to each model's size, exposure is set per material, and the camera fits the real silhouette (mesh vertices, iterated) instead of the bounding box, which was clipping muzzle and stock. - **AKM textures found.** The AKM's maps are packed in its .blend, so `has_data` was False and the lookup silently returned nothing — the rifle rendered as default white plastic. The AKM now renders with its packed base colour, normal, roughness, metallic and AO. - **A89 shown as the ADFRC F89** (producer decision). The in-build A89 mesh came from `ADFRC_F89_Minimi_MLOD`, which is assembled from Minimi parts plus Maximi (M249) and Mag58 parts and reads as an M249. The site now renders `ADFRC_F89_Minimi_Mod_MLOD` — F89_Base_01/02, F89_MK3_01, MK3_Handguard, with their own textures — and the card, alt text and ASSET_REGISTER all say so. - Website rebuilt and published; `Docs/Website/*` updated. ### TESTING | Check | Command | Result | |---|---|---| | Renders | `blender --background --factory-startup -P Tools/Blender/render_weapons.py` | 6/6 rendered; texture report shows every slot resolved to a real map | | Render statistics | `python Build/audit/render_check.py` | opaque-pixel means 102–150, crushed blacks ≤ 2.1%, blown highlights ≤ 0.2% (AKM 0.7%) | | Optic placement | `blender -P Build/audit/probe_slots_geom.py` | scopes now between the sights; every other slot's bounds unchanged | | Sight derivation | `blender -P Build/audit/sight_points.py` | centres from the Arma memory points, not hand-tuned | | Optic-only renders | `blender -P Build/audit/scope_check.py` | lenses and reticles render as glass, no black holes or white blobs | | Website (local) | `node interaction_test.js`, `node responsive_audit.js`, `node text_audit.js` | all checks passed; 9 viewports × 2 pages clean | | Website (live) | `node live_verify.js` | sections incl. `loadout`; no console errors, no failed requests | ### ASSETS - No new third-party assets. `Docs/images/weapons/*.png` re-rendered (studio renders of existing models). The A89 render changes provenance from `SM_A89.fbx` to `ADFRC_F89_Minimi_Mod_MLOD`, both L-0021 material under the recorded website promotion exception. - ASSET_REGISTER website-promotion exception updated to name the ADFRC F89 stand-in. ### RISKS - **R-38 — the game FBX files were edited in place.** The optics were moved without re-running the exporter (its optic sources are gone), so `adfrc_weapon.py`'s corrected rule and the meshes agree only because `fix_weapon_optics.py` applied the same deltas. Re-exporting a weapon from source in future will overwrite the mesh; check the optic position afterwards. - The A89 shown on the site is not the mesh in the build. The build still carries the Minimi/Maximi variant; swapping it to the F89 is a separate change to `SM_A89.fbx`. ### DEFECTS FOUND - Optic mounted over the stock on all three scoped rifles (producer review of the renders; root cause found in `adfrc_weapon.py`). - Renders showed no texture: filename-guessed texture lookup, overlay maps wired as diffuse. - Lights aimed away from the subject; bounding-box framing clipping the weapons. - Packed-texture lookup keyed on `has_data` (AKM rendered untextured). - ADFRC blend collector left the six-triangle LOD proxy boxes in the scene, so they rendered as multi-metre grey planes around the A89. ### NEXT ACTION **Swap the build's A89 mesh to the ADFRC F89** so the game and the site show the same weapon, or record a decision to keep the Minimi variant in the build and label it as such. --- ## Session 042 — 2026-09-28 — Player-Model Renders For Both Sides: BUILT, PUBLISHED, THEN WITHDRAWN ### OUTCOME **The Soldiers section was published and then withdrawn in the same session.** The producer reviewed the renders on the live site and rejected them: the body's proportions read as a mannequin, not a soldier. The section, its nav link, its styles and its published image derivatives have all been removed. The renders and the tooling remain in the repository for internal use. This entry records what was built, what was wrong with it, and what was learned, because the tooling work is sound and will be reused when there is a body worth rendering. ### BUILT - **`Tools/Blender/render_soldiers.py`**, producing matched studio renders of both sides: 3 ACR in CMECU carrying the A88, MAF in the red-earth set carrying the A4. - **Camouflage retuned to the producer's reference photography** (Australian Disruptive Pattern-style uniform shots). `Build/audit/tune_camo.py` measures the fabric in the reference and searches the palette for the closest match instead of eyeballing it: reference lum p10/p50/p90 = 40/129/235, median saturation 0.46, oxide-red population 8.7%. The old four-tone set had no pale ground and no red and rendered as a khaki wash (sat 0.23, nothing above 175). The new seven-tone set scores lum 43/107/205, sat 0.47, red 10.2%. Still noise-generated, still original, still not AMCU/Auscam (ADR-016). **This part is kept** — it is the game's own texture, not a website asset. - Renderer bugs found and fixed along the way, all by measuring rather than looking: - Each kit FBX carries its own armature scaled 0.01 to match the body's centimetre rig; deleting it stripped the scale and blew the kit up 100x, which put a 1.8 m figure inside a 183 m camera frame and rendered the MAF as a torso with no legs. - The camera was fitted to rest-pose vertices (the A-pose) rather than the evaluated mesh. - `pbr()` looked for `T_SS_` where the files are `T_SS__BC`, so **no base colour was ever linked** and both soldiers rendered as untextured default grey. - The weapon was anchored to the midpoint of the two wrists, which put it in mid-air beside the figure. It is now fitted by construction: the grip point is placed on the right wrist and the bore swung at the left, bringing each wrist to within 44–59 mm of the rifle (a hand's width) from several metres. - `bpy.ops.object.mode_set` returns CANCELLED with no exception unless the object is *also* selected, so pose-bone rotations written from OBJECT mode were silently discarded. ### WHY IT WAS WITHDRAWN - **The body is the blocker.** It is the L-0016 Fab mannequin — which ASSET_REGISTER already tracks as C-001, "Manny is a placeholder; original body required" — wearing L-0021 kit fitted to a *different* skeleton. The proportions read as a dummy and the carrier sits on the torso as flat slabs. Neither lighting nor texture work fixes a fitting and silhouette problem, and the producer's read was that the body is the most obviously wrong thing in the image. - **The rig cannot be posed properly.** The bone chain is disconnected: the upperarm's tail sits 16 cm from the lowerarm's head while the mesh spans the gap, so rotating the upperarm does not carry the wrist. IK is impossible on it. `Build/audit/solve_carry.py` instead searches the four arm angles against carry targets and converges to a 1.4 mm residual, which is good enough to put the hands on the weapon but cannot make the pose read as natural. - **The preview tooling failed for the whole session** ("produced no frames"), so the renders were never visually checked before publishing, and two changes made while guessing at numbers made the result worse rather than better (see DEFECTS). ### DEFECTS FOUND - The weapon floated beside the figure instead of being held (midpoint-of-wrists anchoring). - No base colour was ever linked; both soldiers were untextured grey. - The gear rendered 100x oversized and the figure was framed out of shot. - The MAF render had no legs and the camera fitted the rest pose. - **UV tiling was the wrong fix and was reverted.** The FBX lays UVs 0..1 over the whole 1.8 m figure, so the camo is genuinely coarse. Scaling the UVs 7x (camo) and 16x (fabric) was tried on the reasoning that finer blobs would read better; at render resolution it turned the pattern into high-frequency noise and looked worse. Reverted to 1.0, with a note that the correct fix is UVs authored at the right scale in the source meshes, not a global multiplier. - **Flat ambient was the wrong fix and was reverted.** Raising the world background from 0.35 to 0.85 to open up crushed blacks removed the form-shaping entirely and made the figures read as cardboard. Reverted to 0.35. - Both of the above were made without ever seeing the render, because the preview capture tool was unavailable. **Process lesson: publish nothing visual that has not been looked at.** ### ASSETS - No new third-party assets. `Docs/images/soldiers/*.png` are studio renders of existing models, now internal-only. Nothing from this session is published. - The CMECU and MAF texture retune is kept and applies to the game, not the website. - ASSET_REGISTER and LICENCE_REGISTER L-0022 record the camouflage work and the withdrawal. ### RISKS - **R-39 (closed).** The soldier renders were two restricted layers in one image (L-0016 body, L-0021 kit) with no single material whose clearance carried the picture. Withdrawn, so the exposure no longer exists. ### NEXT ACTION **Do not render the player model for the website until C-001/C-002 has an original body.** The renders, the pose solver and the camera/lighting work in `render_soldiers.py` are ready to reuse the moment there is a mesh worth rendering. The gap the section was filling is better filled by a frozen in-engine capture, which is the site's top outstanding asset. --- ## Session 043 — 2026-09-28 — Map Section Corrected: Four Invented Names Replaced With The Six Real Maps ### OUTCOME **The website's map section was factually wrong and has been corrected.** It listed five maps — **Red Ridge, Ironbark, Port Wakefield and Wattle Creek** — that do not exist and never have. No level, no design document, no `.umap` file. They were carried over from the Phase 0 brief and GDD and were never reconciled with the maps actually built. The producer caught it. They are replaced with the **six real maps**, each at its true documented status, taken from `Docs/MAPS_*.md` and the ADR log. Published and verified live. ### THE SIX MAPS, AS PUBLISHED Each map, its level name, and the status the website now publishes. Every one of these is taken from the map's own design document, not from the brief. - **Red Gum Station** (`L_RedGum_01`) — **in the game.** The first playable map (ADR-022). Its objectives were re-laid and its deployments pulled in for fairness. Not signed off, and it has not yet produced a capture. Source: `MAPS_REDGUM.md`. - **Dry River** (`L_DryRiver_01`) — **in the game, and in production.** The Phase 1 greybox vertical slice, and the design standard the other maps are measured against. Source: `MAPS_DRYRIVER.md`. - **Selat Canal** (`L_SelatCanal_01`) — **in production.** Built; the only urban map and the only Special Forces map. Its walkable area is too small for a fair three-objective sequence, so it needs a redesign pass. Source: `MAPS_SELATCANAL.md`. - **Saltbush** (`L_Saltbush_01`) — **in production.** Built; deployment and objectives rebuilt for fairness, and it has produced a capture in a bot match. Navigation, sightlines and cover are still unmeasured. Source: `MAPS_SALTBUSH.md`. - **Bluestone** (`L_Bluestone_01`) — **early build.** The flooded slate-pit level, built 2026-09-28 from the quarry diorama. Not playtested, and it has no design document yet. - **Ravenshoe Crossing** (no level) — **design proposal only. Nothing has been built.** Source: `MAPS_RAVENSHOE.md`, ADR-027. Each card states what the map is for and what is *not* finished about it. Ravenshoe says in bold that nothing has been built, because that is the single most likely thing to be misread as implied work. ### THE DEFECT WAS NOT ONLY IN THE HTML The invented names were in four places, and three of them were still live: - `Site/index.html` — the map grid, plus a count in the Adapt pillar ("Five original map concepts") and the section title. Fixed. - `Docs/DEVELOPMENT_ROADMAP.md` §7 — the Phase 4 task table listed MP-02…MP-05 as *Red Ridge, Ironbark, Port Wakefield, Wattle Creek*. **This file is published to the site as `data/DEVELOPMENT_ROADMAP.md` and rendered in the live Roadmap section**, so the invented maps were visible to any visitor who scrolled there. Rewritten against the real six, with the remaining IDs renumbered to MP-08…MP-11. - `Docs/ASSET_REGISTER.md` §4.9 — the map table tracked Dry River, the four invented maps and Ravenshoe, and did not track Red Gum, Selat Canal, Saltbush or Bluestone at all. M-002…M-005 now carry the real maps with their real statuses; Ravenshoe stays M-008. - `Docs/GAME_DESIGN_DOCUMENT.md` §4.7 — "Five map concepts" listing the invented set. Replaced with the six real maps and a pointer to `Docs/MAPS_*.md` as the source of truth. `Docs/ORIGINAL_BRIEF.md` still lists the invented names. **Deliberately left alone**: it is the Phase 0 brief, a record of what was asked for at the time, and rewriting history in it would be worse than the staleness. ### STATUS VOCABULARY The cards use three badges, and each is a text label rather than a colour, as the design system requires: `In the game` (new `.badge--live` modifier, sage), `In production` and `Early build` (`.badge--wip`, brass), `Design proposal` (default badge). The section is now titled **Maps in the build** rather than "Map concepts", because two of the six are not concepts. ### A RESPONSIVE DEFECT FOUND ON THE WAY The first draft of this entry carried the six maps as a four-column table, and the changelog page immediately measured 418 px wide inside a 390 px phone viewport. The cause was not the table: `.sessions` was a grid with the **default `auto` column**, so a single wide descendant sized the whole column to max-content, and `body { overflow-x: clip }` turned the resulting 68 px overhang into **clipped, unreachable content on every session on the page** — not just the one with the table. `responsive_audit` cannot see this: it compares `scrollWidth` to `clientWidth`, and the clip defeats the measurement. Fixed in `Site/styles.css` with `grid-template-columns: minmax(0, 1fr)` on `.sessions` and `min-width: 0` on `.day-group`, matching the convention already used by `.pillars`, `.kv` and `.phase`. The table would then have scrolled inside its own box, but a 308 px-wide scrollable table is a poor way to read six sentences, so the table was replaced with a list anyway. The CSS fix stands: it is latent for every other session until the next one carries a wide block. ### FILES CHANGED - `Site/index.html` — map grid, section title, section lede, Adapt pillar bullet and link. - `Site/styles.css` — added `.badge--live`; fixed the `.sessions` grid column (above). - `Docs/DEVELOPMENT_ROADMAP.md` — Phase 4 map task table (**published**). - `Docs/ASSET_REGISTER.md` — §4.9 map table. - `Docs/GAME_DESIGN_DOCUMENT.md` — §4.7. - `Docs/CHANGELOG.md`, `Docs/Website/WEBSITE_TEST_REPORT.md` — this entry and the defect note. ### TESTING - `python Tools/publish_site.py` — built and published. - `responsive_audit` — all 12 viewport/page combinations clean. - `text_audit` — no text under 12 px, no tap target under 44 px, no viewport spill on either page (the single reported spill on the home page is the intentional full-bleed hero, `scrollW == docW`). - `interaction_test` — all checks passed, no console errors, no failed requests. - `phase_measure` — shell gaps symmetric at 1920 and 390. - `faq_check` — 11 entries, Steam answer intact. - `live_verify.js` — console errors none, failed requests none. - Live content check: all six real names present, all four invented names and the phrase "map concepts" return **0 hits** on the live page and in the live `data/DEVELOPMENT_ROADMAP.md`. ### RISKS - **R-40 (open, low).** The website's copy is hand-maintained and is only ever as accurate as the last time someone reconciled it with `Docs/`. A guard that fails the build when a published document names a map with no matching `Docs/MAPS_*.md` entry would catch this class of error automatically. Not written. ### NEXT ACTION **Add the map-name cross-check to the audit suite** so a published map name without a design document behind it fails the build, the way a broken image link does today. --- ## Session 044 — 2026-09-28 — Dry River Expanded And Dressed From Packs, Kill Feed, Lyra Pops Removed, Bluestone Quarry, Gloved First-Person Arms ### COMPLETED - **Dry River ground and size** (producer screenshots: grey checker, "red dirt slightly off"). The Rural Australia `_NA` textures are packed masks; sampled as normals they failed the material and Unreal drew its checker. Terrain and outer skirt now both use the pack's `MI_Ground_Dirt_01` (world-space UVs), so there is no seam; Nanite is off on the terrain mesh (it showed flat low-mip texture). Playable area 340 x 240 m (`dryriver_world.PLAY_HALF_X/Y`), with boundary volumes and a matching nav volume. - **Dry River dressing from packs, not blocks** ("assets added are very poor quality, just a block"): - textured corrugated shelters built in Blender (`dryriver_shelters.py`: lean-to, three-bay shed, tank); - Megascans corrugated iron (`setup_fab_materials.py`: `M_SS_ScanPBR`, `MI_SS_CorrugatedIron`), replacing Singapore Canal metal that carried Asian ornament; - yard clutter, sandbag sangars and supply dumps from Singapore Canal *generic* props; - Rural Australia rocks, logs and trees, with hidden trunk colliders; - tinted quarry ledges, rock clusters and 560 creek stones from the Scene Quarry pack. Old blockout crates, barrels and scrub were swapped in place; puddles and pale ground patches were removed. - **Red Gum homestead textured** from pack materials (edits to the other agent's `import_redgum_homestead.py` and `redgum_homestead.py`; those files and the map stay uncommitted with that agent's work). - **Kill feed** (producer): top-right "killer · WEAPON · victim" in viewer-relative colours, plus "ELIMINATED " below the crosshair for the viewer's own kills. Server binds each pawn's `ULyraHealthSet::OnOutOfHealth` and sends per-viewer entries through `USSKillFeedRelay` (a client RPC on each PlayerController). Rules (`FSSKillFeedRules`) are in Core, with a new test. Lyra's message path could not be used: `FLyraVerbMessage` is not exported, and the game state multicast only broadcasts on clients. - **Lyra presentation removed** (producer: "weird lyra blocks when damaged", "hit points above the enemy"): the bridge destroys Lyra number-pop and nameplate components client-side; the log showed 11 removed in a live match. Lyra's red/blue Tab scoreboard is collapsed while ours shows. - **Bluestone Quarry** (`L_Bluestone_01`, producer: "activate that african map"). The African Slate Quarry is a studio-lit ~70 x 80 m diorama, not a level, so the generic builder (`build_objective_map.py`, key `quarry`) now does the following: - strips the showroom and the light bars above the pit; - gives the pack meshes complex collision; - adds outdoor daylight and fog; - rings the pit with a rim at each side's measured edge height, dressed with the pack's own rocks and bushes and textured with a new world-aligned material (`M_SS_WorldGroundVT`); - adds boundary walls and three objectives. All legs connect. It is on the operations menu. Paused for fine-tuning on the producer's instruction. - **First person: gloved arms view model** (producer decision on ADR-024 S2: "arms view model", recorded here). The Fab M4 and G17 FPS packs, with their real-weapon models removed (`Tools/Blender/fp_arms.py`), give draw, fire, reload, empty reload and holster. The A-series weapon rides the pack's weapon bone at a grip measured from the idle pose. Pistols switch to the G17 arms automatically (held mesh shorter than 35 cm). Reloads are time-scaled to Lyra's montage; walk bob, sway, sprint lowering and fire kick are procedural. Sleeves use the soldiers' AMCU G3 shirt fabric and the hands a coyote glove, via a per-polygon UV mask (`make_fp_arms_texture.py`). Weapons moved forward twice on producer feedback (`ss.FP.ArmsOffset 17 0 -2`, `ss.FP.WeaponOffset 9 0 -3`). - **Own-body shadow** ("da Vinci shadow"): hidden meshes stopped refreshing bones, so the shadow stayed in the bind pose. The local body and gear now keep posing while hidden. - **VibeUE** (MIT, git-ignored clone in `Plugins/VibeUE`) builds and loads. Its services are Python-callable from our commandlets, verified: 800+ functions including AnimGraph, AnimSequence and Skeleton; CLAUDE.md section added. The Unreal MCP and EditorToolset plugins are enabled in the `.uproject`, with VibeUE marked `Optional`. Auto-starting the editor MCP server was **not** configured (blocked by the permission classifier; the producer's call). - **ADR-028** (producer): every Fab asset is cleared for our use; no per-asset licence lookups. - Asset review for the producer: the Animation Starter Pack (deaths, prone, hit reacts) is the next animation source; Vibe3D (scripted collision, LODs, UVs) and Universal PCG Scatter are useful. See NEXT ACTION. ### FILES CHANGED - **Code:** - `Plugins/SouthernSpearCore`: `SSKillFeedState.h/.cpp`, `Tests/SSKillFeedTests.cpp`; - `SouthernSpearLyraBridge`: `SSKillFeedSubsystem.h/.cpp`, `SSLyraReflection.h`, `SSHudStateSubsystem.h/.cpp`, `SSFirstPersonSubsystem.h/.cpp`; - `SouthernSpearUI`: `SSKillFeedWidget.h/.cpp`, `SSPlayerHudSubsystem.h/.cpp`, `SSMenuWidget.h/.cpp`. - **Tools:** `Tools/Unreal/expand_dryriver.py`, `setup_fab_materials.py`, `setup_fp_arms.py`, `build_objective_map.py`; `Tools/Blender/dryriver_shelters.py`, `dryriver_skirt.py`, `fp_arms.py`; `Tools/Common/dryriver_world.py`; `Tools/Textures/make_fp_arms_texture.py`. - **Content:** `Content/Maps/L_DryRiver_01.umap`, `L_Bluestone_01.umap`; `Content/Art/Blockout/SS_MAP_DryRiver_01`, `SS_MAP_DryRiver_Skirt`; `Content/Art/Environment/DryRiver/*`, `Fab/*`; `SSExp_ObjectiveAssault/Content/FirstPerson/*`. - **Config and docs:** `SouthernSpear.uproject`, `.gitignore`, `CLAUDE.md`; `Docs/DECISION_LOG.md` (ADR-028); `Docs/LICENCE_REGISTER.md`; this file; `Docs/evidence/S044_*.jpg`. ### TESTING | Test | Command | Exit | Result | Evidence | |---|---|---|---|---| | Guard | `python Tools/validate_architecture.py` | 0 | PASS | console | | Build | `Build.bat SouthernSpearEditor Win64 Development` | 0 | Succeeded (after every C++ change) | console | | Automation | as in CLAUDE.md | 0 | 37/37 `Result={Success}` | `Build/tests_044.log` (not retained) | | Kill feed live | bot match, `-FORCELOGFLUSH` | 0 | kill lines with weapons and "(you)"; producer: "kill feed looks good" | `Docs/evidence/S044_kill_feed.jpg` | | Bluestone passes | `SS_MAP=quarry SS_PASS=level|nav ... build_objective_map.py` | 0 | ok true; legs [T,T,T,T]; deployments 120 m apart, 158 m walk | `Build/objective_map_quarry_*.json`, `Docs/evidence/S044_bluestone_quarry.jpg` | | First person | `-game -SSShotAt ... -SSExec="ss.FP.DebugSlot 0/1"` | killed after shot | rifle and pistol held in gloved hands, AMCU sleeves | `Docs/evidence/S044_fp_arms_rifle_pistol.jpg` | | Dry River | `expand_dryriver.py` | 0 | ok true, nav 3 path points; the two known dressing checks fail as before | `Docs/evidence/S044_dryriver_*.jpg` | NOT RUN: - the own-body shadow is not visually confirmed (no capture caught it in frame); - the Lyra Tab scoreboard hide was not seen in play (no scripted Tab); - a damage hit was not captured after the number-pop removal; - ADS with the new arms was not tuned; - `verify_dressing.py` was not re-run after expansion; - nav was not rebuilt for Red Gum. ### ASSETS - **Fab, cleared under ADR-028:** M4 and G17 FPS animation packs (arms and clips only; weapon models dropped); African Slate Quarry (the Bluestone base, as ADR-022 did for Red Gum, on producer direction); Megascans Military Trenches Corrugated Wall (texture set). - **Other packs:** Singapore Canal generic props and wood materials on Dry River (no architecture or ornament); Rural Australia ground, rocks, logs and trees. - **Not used:** the gloves pack (CC BY 4.0, "Bobeer"; credit if used). - **Tooling:** VibeUE (MIT). - **Original:** `M_SS_WorldGroundVT`; the arms sleeve/glove textures are derived from the pack's shading and the ADFRC G3 fabric (L-0021). ### RISKS - **R-40:** first-person arms reuse Fab pack skeletons that differ from the body; third-person and first-person reloads are two separate animations kept in step by time scaling, not by shared data. - **R-41:** `L_Bluestone_01` references the git-ignored Scene Quarry pack, so a clone needs the pack from Fab (as R-19). ### DEFECTS FOUND - Checker ground from packed mask textures sampled as normals (producer screenshot). - Studio lights, showroom and void edges in the quarry scene (probe and captures). - Arms texture UV split wrong twice: forearms and hands overlap in U, so a per-polygon mask was needed (in-game capture). - Pistol arms culled after a mesh swap (bounds; capture). - The asset-rename step double-prefixed clips on re-import (import report). ### NEXT ACTION **Deaths and hit reactions:** retarget the Animation Starter Pack deaths and hit reacts to the UE5 mannequin, play them on death, then blend into ragdoll. --- ## Session 046 — 2026-09-28 — Ragdoll Deaths, No Lyra Cubes, KILLED IN ACTION, Re-Deploy, Quieter HUD, Working Scopes, Dry River Farm Props ### COMPLETED - **Build unbroken** (producer: "SouthernSpear could not be compiled"): the death work had three compile errors (a non-existent collision-profile constant, a missing ability-system include, and a `TeamOf` clash between the scoreboard's local helper and `SSLyraReflection::TeamOf` under the unity build; the helper is now `ScoreboardTeamOf`). - **Deaths** (producer: "ragdoll / death animations?"): `ASSCharacter` ragdolls the body 0.12 s after death with a push along the killing shot (from the damage cue), keeps the corpse 15 s instead of Lyra's instant hide-and-destroy, and hides the first-person arms and weapon once the body goes limp. 17 ragdolls in a 10-bot match, no new errors. - **Lyra cubes gone** (producer: "the spawn 'weird cubes' needs to go"): `ShouldAcceptGameplayCue` refuses `GameplayCue.Character.Death` (NS_DeathCubes) and the spawn cue (`GameplayCue.Character.Spawn`, GCNL_Spawning's cube materialise). The log showed 50 spawn cues refused in one match. - **KILLED IN ACTION** (producer): on the viewer's own death the screen darkens and "KILLED IN ACTION / BY · " shows for 4 s (`FSSKillFeedRules::RecentLocalDeath`, tested); the class selection opens after it. - **RE-DEPLOY** in the match menu (producer): the menu sets a request in `USSLocalHudState`; the bridge sends it through the player controller's relay (`USSKillFeedRelay::ServerRedeploy`) and the server applies Lyra's own self-destruct, so the player respawns with the selected class. The banner then reads "RE-DEPLOYING". - **HUD re-laid, after America's Army 2** (producer: "a lot at the top of the screen", "kill feed does push under the mini map"; AA2 kept objectives small in the top-right corner): - minimap 200 px at the top right, with a compact objective block the same width directly under it; - the kill feed moved to the top left; - the compass is a slim strip alone at the top centre. - **Scopes work** (producer: "scopes don't work at all"; the eye looked at the back of a solid optic). Aiming a magnified optic hides the view model, narrows the view by the optic's power and draws an eyepiece (black surround, round mask, stadia, post, aim point). The power comes from `USSLocalHudState::OpticMagnificationFor`, tested: A25 6x, the A88 family and A89 4x, others 0 (iron or red dot, aimed over the view model as before). The producer confirmed "scope on the A88 is working". - **Dry River farm props** (producer: "replace the 'cars' that were just boxes with actual cars ... add barns, the windmill, wells"). New re-runnable pass `Tools/Unreal/farm_dryriver.py`: - the three box wrecks become the Fab car wreck at the dressing plan's wreck positions; - a windmill, a water tower, a hand pump and the StoneWell well at the Water Point; - at the Farmstead, the buildings stand where the blockout designed them (`MAPS_DRYRIVER.md` 4.4): the open pole barn as the objective's shed, the enclosed barn as the residence, and a timber rail fence on the stock-pen line; - also a pump, a well, a caravan, an outhouse, and 14 fuel drums by the sheds and barns. - **Prop textures** (producer: "all of these have no textures on them and look strange"). Four Fab downloads (windmill, barn, old barn, fuel barrel) shipped their FBX without the texture files it references. Their slots now take textured materials: timber and roofing iron from the Dry River shelters, and rusted or galvanised metal from the Modular Rural Cabin pack. The Cabin pack's own wood and roof came out green and blotchy on these meshes, so they were replaced. The drums became the Cabin pack's textured drum. The pump's textures were embedded in its FBX (`Tools/Blender/extract_fab_textures.py`), and the water tower ships its PNGs: both now have their own PBR instances. - **Greybox out of the terrain** (producer screenshot: red wall and slab at the objective): the farm shed, residence, pen rails and 2 m layout-marker slabs no longer export into the terrain FBX (`dryriver_blockout.py`; CSVs byte-identical), re-imported alone by `Tools/Unreal/reimport_dryriver_terrain.py` (120 fewer triangles). - Map work then halted on the producer's instruction ("we will get there and improve them down the track"). ### FILES CHANGED - **Core:** - `SSKillFeedState.h/.cpp` (`RecentLocalDeath`); - `SSLocalHudState.h` (re-deploy request, optic power); - `SSGlyphTextures.h` (`ScopeMask`); - tests `SSKillFeedTests.cpp`, `SSHudTests.cpp`. - **Bridge:** - `SSCharacter.h/.cpp` (ragdoll, corpse, cue refusal); - `SSFirstPersonSubsystem.cpp` (dead or scoped: hide the view model); - `SSFirstPersonCameraMode.cpp` (scope FOV); - `SSHudStateSubsystem.cpp`; - `SSKillFeedSubsystem.h/.cpp` (`ServerRedeploy`); - `SSScoreboardSubsystem.cpp`. - **UI:** `SSKillFeedWidget.h/.cpp`, `SSPlayerHudWidget.h/.cpp`, `SSPlayerHudSubsystem.h/.cpp`, `SSMenuWidget.h/.cpp`. - **ObjectivesUI:** `SSObjectiveStatusWidget.cpp`, `SSMinimapWidget.cpp`, `SSCompassWidget.cpp`. - **Tools:** `Tools/Unreal/farm_dryriver.py`, `reimport_dryriver_terrain.py`; `Tools/Blender/extract_fab_textures.py`, `dryriver_blockout.py`. - **Content:** `Content/Maps/L_DryRiver_01.umap`; `Content/Art/Blockout/SS_MAP_DryRiver_01` (fbx, uasset, blend); `Content/Art/Environment/DryRiver/Farm/*`. - **Docs and evidence:** this file; `Docs/evidence/S046_*.jpg`. ### TESTING | Test | Command | Exit | Result | Evidence | |---|---|---|---|---| | Guard | `python Tools/validate_architecture.py` | 0 | PASS | console | | Build | `Build.bat SouthernSpearEditor Win64 Development` | 0 | Succeeded after every C++ change | console | | Automation | as in CLAUDE.md | 0 | 37/37 `Result={Success}` | `Build/tests_046.log` (not retained) | | Ragdoll | 10-bot Dry River match, `-FORCELOGFLUSH` | killed | 17 `SSRagdoll` lines, no new errors | log | | Cubes | 8-bot match | killed | 50 `SSCue refused GameplayCue.Character.Spawn`; death cue refused in code | log | | KIA | `-SSExec="EnableCheats\|DamageSelf 500"` + capture | killed | banner, darkened view, no arms | `Docs/evidence/S046_killed_in_action.jpg` | | HUD layout | 8-bot match capture | killed | compass top centre, minimap and objectives top right, feed top left | `Docs/evidence/S046_hud_layout.jpg` | | Scope | `-SSExec="ss.FP.ForceAim 1"` + capture | killed | eyepiece, reticle, 4x view, view model hidden | `Docs/evidence/S046_scope_view.jpg` | | Farm pass | `farm_dryriver.py` | 0 | ok true (wrecks 3, windmill, tower, pumps, wells, barns, 13 fence segments, drums) | `Build/farm_dryriver_report.json` | | Terrain | `reimport_dryriver_terrain.py` | 0 | ok true, 15096 → 14976 triangles, complex-as-simple | `Build/reimport_dryriver_terrain.json` | | Nav | `build_dryriver_nav.py` | 0 | ok true, 6 path points | `Build/dryriver_nav_report.json` | NOT RUN: - **RE-DEPLOY** was not exercised in play (no scripted menu click); - the **final Farmstead layout** (barns in their designed places, pen fence, greybox gone) was not captured: the producer halted map work first. `S046_dryriver_farmstead.jpg` shows the textured barns before the move; - scopes with the A25 (6x) and A89 were not captured; - the pistol still has no iron-sight alignment (producer, end of session). ### ASSETS - **Fab, cleared under ADR-028:** - Old Rustic Hand Water Pump and Water Tower, with their own textures; - Red car wreck, American Old Windmill, Barn and Old Barn, via the Ravenshoe agent's prepared meshes. - Seller AI-use flags on some of these are recorded by that agent (ADR-029). - **Packs in `Content/`, referenced in place, not committed (ADR-021):** - Modular Rural Cabin: drums, caravan, outhouse, fence, rust and metal materials; - StoneWell: the well; - Singapore Canal timber; Megascans corrugated iron (already in use). - Both new packs had complex-as-simple collision set on the used meshes. ### RISKS - **R-42:** Dry River now references the Ravenshoe agent's prop meshes and two further git-ignored packs (Modular Rural Cabin, StoneWell): a clone needs those packs, as R-19/R-41. - **R-43:** the Farmstead's barns replace the designed shed and residence by footprint, not by exact cover shape (the open barn is 12 × 8 m against the 18 × 10 m shed); sight lines through the objective were not re-audited. ### DEFECTS FOUND - Death code broke the build: missing include, a wrong constant, and a unity-build name clash (build log). - First-person arms stayed on screen after death (capture). - Scoped aim looked at the back of a solid optic (capture). - Fab downloads missing their textures, so props rendered flat (producer). - The Cabin materials rendered green on foreign meshes (producer screenshot). - The farm pass deleted the box wrecks and then its own replacements on a re-run. Fixed: it now places from the dressing CSV. - Greybox marker slabs and pen rails were still in the terrain mesh (producer screenshot). ### NEXT ACTION**First-person weapon handling:** - iron-sight alignment for the pistol; - a better grip fit in the gloved hands, trying the Fab gloves pack (CC BY 4.0, "Bobeer": credit line); - ADFRC weapon animations, textures, models and sounds, using the ADFRC agent's `ASSET_MANIFEST.json` and integration guide. --- ## Session 047 — 2026-09-28 — Map Section Brought Up To Date, And The Playability Audit Published ### OUTCOME **A second agent worked in this repository while Session 043 was still being published, and moved the maps a long way.** Sessions 044, 045 and 046 landed: Dry River expanded and dressed from asset packs, Bluestone activated into a playable level, and Ravenshoe Crossing went from a written proposal to a built, dressed map of 467 actors. A new `Docs/MAPS_PLAYABILITY_AUDIT.md` appeared as well, and it measures the four older maps against the design rules written for Dry River. Most of them fail. The site's map section was therefore wrong again within a day of being corrected, in the most absolute way available: it said, in bold, that **nothing had been built** for Ravenshoe Crossing, and 467 actors now exist. The section has been rewritten against the current state of the repository, and the audit's verdicts are now published on the site. ### THE RAVENSHOE CORRECTION The card used to read: *"A bridge-crossing map written up as a full design proposal. **Nothing has been built** — there is no level for it yet."* It now reads as **in early work**, and says what is true: the gorge, a 68 m iron lattice-girder road bridge, a stone road-gate house and a playable creek bed are all built and dressed; 467 actors; 32 of 32 structural audit checks pass; **navigation is not baked, so it is not yet playable with bots**; and it is early enough that all of it will change. The bridge, gatehouse and gorge are original geometry (ADR-027, ADR-029) dressed by already-cleared Class A packs referenced in place (ADR-021) — the site's originality rules still hold, and nothing about the map's provenance needed softening. ### THE OTHER FIVE CARDS, AGAINST THE REPOSITORY - **Red Gum Station** — in the game, the first playable map, and now carries the audit's verdict: 17% nav coverage, twelve hard cover objects and no soft cover on a kilometre map. The card says it is **not playable as it stands**. - **Dry River** — in the game and in production; playable area expanded to 340 × 240 m and dressed from cleared packs. The card now records that the map the others are measured against **fails six of its own nine rules**. - **Selat Canal** — in production; the best close-quarters geometry measured in the project and the worst-placed objectives, all three between 42% and 75% walk-imbalanced. - **Saltbush** — in production, the only map with a capture in a bot match, and measured the strongest of the four: 0 of 64 start pairs see each other, longest sightline 148.5 m. - **Bluestone** — **built, being tuned**: showroom and light bars stripped, outdoor daylight and fog, complex collision, boundary rim, three objectives, all legs connecting, on the operations menu, paused for fine-tuning on the producer's instruction. ### THE AUDIT, PUBLISHED A new panel sits under the grid: **"Measured against their own rules, and not signed off."** It explains that `Tools/Unreal/audit_map_playability.py` is read-only — it loads a map, measures it and writes a report, and never places, moves or saves anything — and gives the scoreboard: Saltbush 7 pass / 3 fail, Selat Canal 5 / 5, Dry River 3 / 6, Red Gum 3 / 6 / 1 n-a. The panel also states what the numbers *do not* say: none of them measures whether a map is fun, and none can tell whether a piece of cover is fairly placed. That is recorded as the producer's decision to publish the failures rather than the passes. **Every figure was traced back to `MAPS_PLAYABILITY_AUDIT.md` before publication.** Two were adjusted for precision rather than left as prose: "a sixth of the ground" became the measured 17%, and Saltbush's sightline is quoted as 148.5 m rather than rounded to 148. `Docs/DEVELOPMENT_ROADMAP.md` §7 (published to the site) now carries the audit's own priority order — Selat Canal objectives first, then Dry River cover and spawn exposure, then Red Gum — and `ASSET_REGISTER.md` §4.9 rows for the five maps record their measured status. ### COORDINATION NOTE The other agent is working in this repository concurrently and its files are changing under this session. Only website files and the four documentation files above were touched here; nothing belonging to Sessions 044–046 was edited, staged or committed. **Session 045 has no entry in this changelog** — the Ravenshoe work was committed with a descriptive commit message (7cda16a4) and the session body, unlike 044 and 046, was never written up. Left alone rather than reconstructed. ### FILES CHANGED - `Site/index.html` — six map cards, section lede, the new audit panel, Adapt pillar bullet. - `Site/styles.css` — `.maps__audit` and `.map-card__meta dd strong`. - `Docs/DEVELOPMENT_ROADMAP.md` — Phase 4 map tasks rewritten to the audit's priority order (**published**). - `Docs/ASSET_REGISTER.md` — §4.9 rows M-001 to M-005. - `Docs/CHANGELOG.md`, `Docs/Website/WEBSITE_TEST_REPORT.md`, `Docs/Website/WEBSITE_DESIGN_SYSTEM.md` — this entry and the content/design rules behind it. ### TESTING - `python Tools/publish_site.py` — built and published. - `responsive_audit` — all 12 viewport/page combinations clean. - `text_audit` — no text under 12 px, no tap target under 44 px, no viewport spill on either page. - `interaction_test` — all checks passed, no console errors, no failed requests. - `faq_check` — 11 entries, Steam answer intact. - `maps_check.js` (new, `Build/audit/`) — reads the rendered DOM at 390 and 1440: six cards with the expected names, badges and states, four audit rows with the expected scores, heading order H2→H3 with no skipped level, no element overflowing the viewport at either width, document scroll width equal to the viewport. Console and request errors: none. ### RISKS - **R-41 (open).** **Published map claims decay in hours when another agent is building in parallel.** Ravenshoe went from "nothing has been built" to 467 actors in a day. The Ravenshoe card says so in as many words, but the same staleness applies to the other five. The site copies status prose by hand; nothing detects that `Docs/MAPS_*.md` and the HTML have diverged. - **R-42 (open, low).** The audit's cover counts only `StaticMeshActor`s, so cover inside a blueprint or instanced foliage is not counted and the hard:soft ratios are approximate — the audit document says so for Selat Canal. The site publishes the ratios as "measured" without that caveat in the scoreboard row; the caveat is in the panel's note. ### NEXT ACTION **Generate the map cards from the map documents' status lines, the way the roadmap, status and changelog sections already are**, so a status that changes in `Docs/MAPS_*.md` cannot drift from what the site says about it. That closes R-41 and the R-40 class of defect from the same mechanism. --- ## Session 048 — 2026-09-28 — Section Assault (One Life, Attack And Defend), Service Record And Ranks ### OUTCOME **Two producer requests, both code, written in a cloud container without Unreal Engine.** The producer asked for (1) a main game mode modelled on the design philosophy of *America's Army 2*'s round-based mode ("single spawn", Counter-Strike-like), and (2) the ranking and player profile systems wired up. Both are written, reviewed and tested as far as this environment allows. **Nothing in this session has been compiled or run in the engine.** The container is Linux with no UE 5.8 install, so every C++ and automation-test result below is **NOT RUN** and the first Windows build is the real test (R-50). Research (public sources: Wikipedia, GameFAQs/Neoseeker guides, AA2Reborn) on AA2's main mode: rounds where one team attacks and one defends; **one life per round**, and the dead watch until the round ends; win on the objective or by eliminating the other team; honour lost for friendly fire; points for your fireteam surviving. Only the *rules* were taken. No map, mission name, UI, audio or asset (L-0008, ADR-013). ### COMPLETED **ADR-031 — Section Assault** (`Docs/DECISION_LOG.md`). A second rule set on the existing director, maps and objectives, chosen per match (`RulesMode`, map URL `?Rules=Section`, or the new **RULES** choice on the front end). One life per round. Team One attacks the first half and Team Two the second (symmetric over the match, ADR-017). Only attackers capture, and defenders on the point contest or clear it. Attackers win on the final objective or by wiping the defenders. Defenders win by wiping the attackers or **on time**. Halves of 4, first to 5, 4–4 is a drawn match. 300 s rounds. - `FSSSectionAssaultRules` (pure): `StepAttackCapture`, `ResolveRound`, `ApplyRoundResult`, `StepRound`, `NewMatch`. - `FSSMatchState` replicated on `ASSObjectiveAssaultDirector` (round score, attacker, half, alive counts, last round's reason). The director branches on `RulesMode`, and Objective Assault is unchanged by default. - **Core `USSRespawnGate`** (new): the round roster and who is eliminated. The director locks it when play starts and unlocks it at the reset. The bridge reports eliminations from Lyra's `OnOutOfHealth`. - **Holding the dead out without modifying Lyra:** Lyra's `ControllerCanRestart` is private and non-virtual. `USSDeploymentSpawningComponent` overrides the supported virtual `OnFinishRestartPlayer` and unpossesses and destroys the pawn Lyra just spawned for a held-out controller, in the same server frame. RE-DEPLOY is refused while the gate is locked. At the reset the director restarts every pawnless controller, bots included. - HUD: `FSSObjectiveHudModel::ApplySectionAssault`. Round score replaces captures, and the label reads "Attack 4 v 3" / "Defend 4 v 3" from the viewer's side (neutral vantage without a team). The post-round header gives the reason. The round banner says "Take/Hold objective A · one life". **ADR-032 — `Plugins/SouthernSpearProgression`** (new plugin, Core-only dependency). This is the TDD §6.2–6.4 design: - `FSSServiceRecord` schema v1 (XP, statistics, qualifications, commendations, callsign) with a migration step. A newer-schema record is refused and left untouched. - `ISSPersistenceProvider` plus a **dev-only** `FSSLocalDevPersistence`: JSON in `Saved/SouthernSpear/Profiles/`, written to a temporary file then moved over the record. A corrupt file is quarantined as `.corrupt-.json`. `IsAuthoritative() == false`. - Ranks and awards are **data** (`Config/DefaultGame.ini` `[/Script/SouthernSpearProgression.SSProgressionSettings]`): the nine GDD §6.2 enlisted ranks (Recruit 0 … WO1 80 000) and six award rules. - **Core `USSServiceEventSubsystem`** (new) carries service events from the director and bridge to progression: objective captured, round won, round won alive, match completed, match won, friendly kill. **No kill event**, enforced by new guard rule **SS009**. - Server `USSProgressionServerSubsystem` keeps per-player match tallies and applies caps. Positive awards are capped per match and the friendly-kill penalty (−150) never is. It sends results through the `USSServiceRelay` client RPC, and the client `USSPlayerProfileSubsystem` applies them, saves and publishes to **Core `USSLocalProfileState`** (new). Bots earn nothing. - Front end: a profile line (rank abbreviation · callsign, rank · XP / next) and the RULES choice. **Docs:** ADR-031 and ADR-032. GDD §4.6 "as built" note. Roadmap §13 status ("written, not yet compiled"). CLAUDE.md module table, export macro `SSPROG_API`, SS009, and the Core cross-module subsystems. ### FILES CHANGED - New plugin `Plugins/SouthernSpearProgression/` (`.uplugin`, `Build.cs`, `SSServiceRecord.h`, `SSProgressionSettings.h`, `SSProgressionRules.h/.cpp`, `SSPersistence.h/.cpp`, `SSProgressionSubsystems.h/.cpp`, `Private/Tests/SSProgressionTests.cpp`). Registered in `SouthernSpear.uproject`. - Core: `SSServiceEvents.h/.cpp`, `SSRespawnGate.h/.cpp`, `SSLocalProfileState.h` (new). - Objectives: `SSObjectiveTypes.h` (rules/reason enums, `FSSSectionRules`, `FSSMatchState`, `FSSSectionInputs`, `FSSRoundEvents::bMatchEnded`), `SSSectionAssaultRules.h/.cpp` (new), `SSObjectiveActor.h/.cpp` (attacker-only step, presence list), `SSObjectiveAssaultDirector.h/.cpp`, `Tests/SSSectionAssaultTests.cpp` (new). - ObjectivesUI: `SSObjectiveHudModel.h/.cpp`, `SSObjectiveStatusWidget.cpp`, `SSRoundBannerWidget.cpp`, `Tests/SSObjectiveHudTests.cpp`. - LyraBridge: `SSDeploymentSpawningComponent.h/.cpp`, `SSKillFeedSubsystem.cpp`. - UI: `SSMenuWidget.h/.cpp`. - `Config/DefaultGame.ini`, `Tools/validate_architecture.py` (SS009), `CLAUDE.md`, `Docs/DECISION_LOG.md`, `Docs/GAME_DESIGN_DOCUMENT.md`, `Docs/DEVELOPMENT_ROADMAP.md`, `Docs/CHANGELOG.md`. - Evidence: `Docs/evidence/G059_guard_positive.txt`, `Docs/evidence/G059_guard_negative.txt`. ### TESTING - `python Tools/validate_architecture.py` (real tree): **exit 0**, "PASS - no architecture violations found". Evidence `Docs/evidence/G059_guard_positive.txt`. - Guard negative test on a scratchpad copy of `Tools/` + SS plugins. SouthernSpearProgression → LyraGame **and** an `EnemyKilled` service event: **exit 1**, SS002 + SS009 reported (`Docs/evidence/G059_guard_negative.txt`). A separate run with Progression → SouthernSpearObjectives: **exit 1**, SS001. Real source was never left broken. - `SouthernSpear.uproject` re-parsed as JSON after the edit: valid. - **NOT RUN — no Unreal Engine in this environment:** - `Build.bat SouthernSpearEditor Win64 Development`: the first compile of every file above. - Automation: new `SouthernSpear.Objectives.Section.*` (8: CaptureAttackersOnly, CaptureMirrored, ResolveRound, MatchScoringAndSwap, RoundLoop, ReplicationContract, RespawnGate, World.EliminationAndEvents), `SouthernSpear.Objectives.Hud.SectionAssault`, and `SouthernSpear.Progression.*` (6: ShippedTablesAreValid, NoKillReward, ValidationCatchesBadTables, RanksAndCaps, IdsAndCallsigns, PersistenceRoundTrip), plus every existing suite for regressions. - Live Section Assault: `L_DryRiver_01?NumBots=8?Rules=Section?RoundSeconds=90`. Held-out players stay out, bots come back at the reset, sides swap after round 4, and the match ends at 5. - Front end RULES row and profile line, rendered. `Saved/SouthernSpear/Profiles/ServiceRecord_local.json` created and growing across two matches. - `python Tools/publish_site.py`: not run. The public site repository is outside this session's access, and nothing here is ready to publish while it is uncompiled. ### ASSETS None. No asset imported, created or modified. The rank names are GDD §6.2's structure. Insignia remain unmade placeholders (L-0003). ### RISKS - **R-50 (open, high until built).** All Session 048 C++ was written without a compiler or the engine. Care was taken over UE 5.8 specifics (unity-build name collisions, shadowing-as-error, the Lyra private restart path), but expect first-build fixes. Nothing here is claimed to work until the Windows build and the automation run pass. - **R-51 (open, medium).** Holding a player out destroys the pawn Lyra just spawned, in the same server frame. It should never replicate, but the client still receives `ClientRestart` for a pawn that no longer exists. Watch for camera or input glitches on the held-out client. The fallback is a C++ game mode overriding `ControllerCanRestart` (ADR-031 alternative 1). - **R-52 (open, low).** Eliminations come from `USSKillFeedSubsystem`, which binds each pawn's health set on a 0.5 s scan. A pawn killed within 0.5 s of spawning is never reported, so its team can't be eliminated that round and the round runs to time. - **R-53 (open, low, dev-only).** The local service record is editable by its owner. PIE clients on one machine share one file. Acceptable for development only; online persistence needs its own ADR (Phase 5). - **R-54 (open, low).** Objective Assault never ends a match, so its per-match award caps span the whole map session. Section Assault closes a tally on `MatchCompleted`. - **Design gaps (not defects):** a dead player's view stays where they fell (no spectating a teammate yet), there's no movement freeze in pre-round, no HUD toast for awards (`USSLocalProfileState::LastAward` is filled but not drawn), and no callsign entry UI (`SetCallsign` exists). ### DEFECTS FOUND - **Unity-build name collisions caught in review, before any build.** `IsValidStep` existed in both `SSObjectiveRules.cpp` and the new `SSSectionAssaultRules.cpp` anonymous namespaces, and would collide when UBT merges the module into one translation unit. Renamed. A progression helper `Settings()` would have been shadowed by a test local `Settings` (C4459, an error under UE's shadow-variable policy). Renamed. - **Award-order bug caught in review.** The director first posted `MatchCompleted` before `MatchWon`. The progression tally closes on `MatchCompleted`, so `MatchWon` would have opened the next match's tally and used up its cap. The order is now fixed, with a comment. - **Client HUD wouldn't know the rule set, caught in review.** `RulesMode` is set from the URL on the server only. Unreplicated, a remote client's HUD would have shown Objective Assault during a Section Assault match. It now replicates, and `Section.ReplicationContract` asserts it. ### NEXT ACTION **Build `SouthernSpearEditor` on the Windows machine and run `Automation RunTests SouthernSpear`.** Fix whatever the first compile of Session 048 reports, until all suites pass (including the 15 new tests). Then run one live Section Assault match on Dry River with 8 bots and `?RoundSeconds=90`, and record the round/match log lines as evidence. That closes R-50, or turns it into specific defects. --- ## Session 049 — 2026-09-28 — Australian Army Ranks 1–100 With Insignia, Kill XP, HUD Clipping And Scoreboard Title Fixed ### OUTCOME **Session 048 was verified on the producer's machine** (reported by the producer's local agent, not run here): `SouthernSpearEditor` built (20 actions, the new `SouthernSpearProgression` module included), and **52/52 automation tests passed** (37 before plus the 15 new). A live Section Assault run logged `Section Assault round 1: TeamOne attacked, TeamOneWon (DefendersEliminated). Score T1=1 T2=0.` with 8 bots and no spawn failures. R-50 (Session 048 unverified) is **closed** for what those tests cover. The playtest screenshots showed three defects, fixed here. Two producer requests are also done: realistic Australian Army ranks with insignia, and kill XP. ### COMPLETED - **ADR-034.** Service levels 1–100, as America's Army honour worked. Seventeen Australian Army ranks, PTE (1–4), LCPL (5–9) … LTGEN (96–99), GEN (100), shown with insignia on the scoreboard and front end. Kills earn 10 XP, capped at 50 per match (an objective is 100). SS009 retired. - **Insignia drawn in code**, `SSInsigniaRaster.h` (engine-free), following the Army's devices: point-down chevrons; crown (St Edward's pattern) over three chevrons for SSGT; crown for WO2 and MAJ; the Coat of Arms (shield, kangaroo, emu, star) for WO1; Order of the Bath pips in a column; crown and pips for field officers, with Brigadier's three in a triangle; crossed sword and baton for generals. `FSSServiceRanks::InsigniaTexture` turns them into cached 64 px textures. No image files, so nothing to import or register. - **Staff Sergeant:** first dropped (a search said it was being phased out), then **restored**. The producer's source (army.gov.au/about-us/ranks) and Defence's *Badges of Rank* list it. RSM-A (one appointment) and Field Marshal (honorary) are not levels. - **Core:** `USSRankSettings` (ladder + curve, data), `FSSServiceRanks`, and `USSServiceRankComponent` (replicated level on the player state). `FSSScoreRow::ServiceLevel`, `USSScoreboardState::ModeTitle`, `ESSServiceEvent::EnemyKill`, and profile level fields. - **Progression:** ranks moved to Core. Record **schema v2** (`Statistics.EnemyKills`) with a v1→v2 migration. `USSServiceRelay::ServerReportProfile` (level + callsign → server → player state and scoreboard name). `ss.Callsign ` console command. - **Bridge:** `EnemyKill` posted for kills of the other side, and the scoreboard reads the level component. - **UI:** scoreboard RANK column (insignia + level; "—" for bots). Front end: level, rank, XP and the insignia badge. - **Fix — objective panel clipping under the minimap** (screenshots: "ROUND 2 · HALF 1FRIENDLY WIN", "ROUND 2: FRIENDLY WIN, ("). Section Assault strings are now short: "Round 2" + "Attack"/"Defend"; the alive count ("5 V 4") is on its own right-aligned slot on the status line; the post-round line is only the reason ("defenders eliminated"); match result "Match won/lost/drawn". A test bounds the lengths. - **Fix — scoreboard said OBJECTIVE ASSAULT during Section Assault.** The kicker is now a member, filled from `USSScoreboardState::ModeTitle`, which the objective HUD subsystem sets from the director's replicated `RulesMode`. The other agent also flagged `SSMenuWidget.cpp:307`. That line is the front-end **RULES choice button**, which offers both modes by name, so it is correct and unchanged. - **Tool:** `Tools/Progression/rank_preview.py` compiles `insignia_sheet.cpp` with g++ against the game's own engine-free headers, prints the level/XP bands, checks the maths at all 100 thresholds and that every rank draws, and writes the insignia sheet. ### FILES CHANGED - Core: `SSServiceRanks.h/.cpp`, `SSInsigniaRaster.h`, `SSServiceLevelMath.h`, `Tests/SSServiceRanksTests.cpp` (new). Also `SSServiceEvents.h/.cpp`, `SSScoreboardState.h`, `SSLocalProfileState.h`, and `SouthernSpearCore.Build.cs` (private `NetCore`, an engine module). - Progression: `SSServiceRecord.h`, `SSProgressionSettings.h`, `SSProgressionRules.h/.cpp`, `SSProgressionSubsystems.h/.cpp`, `Tests/SSProgressionTests.cpp`. - Objectives UI: `SSObjectiveHudModel.h/.cpp`, `SSObjectiveStatusWidget.h/.cpp`, `SSObjectiveHudSubsystem.cpp`, `Tests/SSObjectiveHudTests.cpp`. - Bridge: `SSKillFeedSubsystem.cpp`, `SSScoreboardSubsystem.cpp`. UI: `SSScoreboardWidget.h/.cpp`, `SSMenuWidget.cpp`. - `Config/DefaultGame.ini` (rank ladder, awards). `Tools/validate_architecture.py` (SS009 retired). `Tools/Progression/` (new). - `CLAUDE.md`, `Docs/DECISION_LOG.md` (ADR-034), `Docs/GAME_DESIGN_DOCUMENT.md` §6.2/§6.4, `Docs/CHANGELOG.md`. - Evidence: `Docs/evidence/G060_rank_table.txt`. The sheet PNG is not committed (no git-lfs in the authoring container); regenerate with `python Tools/Progression/rank_preview.py`. ### TESTING - `python Tools/Progression/rank_preview.py Docs/evidence/G060_rank_insignia_sheet.png`: **exit 0**. g++ `-std=c++20 -Wall -Wextra -Werror` against `SSInsigniaRaster.h` and `SSServiceLevelMath.h`. All 17 ranks draw (Private draws nothing, as intended). The level curve inverts exactly at all 100 thresholds. Level 100 needs 371,414 XP. Evidence `Docs/evidence/G060_rank_table.txt`. The sheet was inspected visually and revised twice: pips too small, the crown too mitre-like. - `python Tools/validate_architecture.py`: **exit 0**. - **NOT RUN (no Unreal Engine here):** the editor build, and automation for the new `SouthernSpear.Core.Ranks.*` (3), the changed `Progression.*` (KillAwards replaces NoKillReward; Caps replaces RanksAndCaps; PersistenceRoundTrip now covers v1→v2), `Objectives.Hud.SectionAssault`, and every suite for regressions. Also not run: the scoreboard insignia column and front-end badge rendered in game, and `ss.Callsign` renaming a player. ### ASSETS No asset files. The insignia are procedural textures created at runtime, after the Australian Army's devices (ADR-034). They are recorded under release gate R-55 rather than as a licence-register item, because no third-party file is used. ### RISKS - **R-50 closed** by the producer's build and 52/52 run for Session 048. **R-56 (open, high until built):** Session 049 C++ is again uncompiled. - **R-55 (open, release blocker).** The crown and the Coat of Arms (and the Army insignia set) need Defence and PM&C permission before release, or the WO1/crown devices must be swapped for fictional ones. - R-51, R-52, R-53, R-54 as Session 048. R-53 now also covers the scoreboard level: it is self-reported by each client from a local record. - **Lyra noise (not ours):** `W_SB_TeamStat:Construct` "Accessed None ... GetComponentByClass" ×4 comes from ShooterCore's own scoreboard widget. It is logged here for the known-noise list. ### DEFECTS FOUND - HUD clipping and the wrong scoreboard title: from the producer's playtest screenshots. - Pips drawn too small, and a crown that read as a mitre: from the rendered insignia sheet (`rank_preview.py` failed the coverage check on 2LT/LT/CAPT before the fix). - The Medic kit equals the Rifleman kit (A88 + A9; healing not in the game): reported by the producer's agent. It is a design gap, noted and not changed here. ### ADDENDUM — ADF Re-Cut registry reviewed for weapons and animation - `Tools/Weapons/adfrc_weapon_data.py` (new, stdlib) reads the committed `config_registry.json` and writes `Docs/WEAPON_SOURCE_DATA.md/.json`. For each A-series weapon it gives the source class, rpm and dispersion per fire mode, magazine, handling, muzzle/ejection memory points, grip-pose clip, reload gesture (shipped or vanilla) and resolved audio. Run: `python Tools/Weapons/adfrc_weapon_data.py` → **exit 0**, 8 weapons, 0 missing. - `Docs/WEAPONS_ANIMATION_PLAN.md` (new): **W1** per-weapon stats (today every weapon is a copy of Lyra's rifle or pistol), **W2** grip sockets from the `handAnim` poses for left-hand IK (not blocked by R-32), **W3** ejection and muzzle effects, **W4** reload audio, **W5** bullpup reload as IK trajectories, **W6** bipod and prone holds, **W7** locomotion (existing audit). Recommendation: W1 then W2. ### NEXT ACTION **Build and run `Automation RunTests SouthernSpear`, then play one Section Assault match with Tab held.** Confirm the RANK column (insignia + level), the front-end badge, the unclipped objective panel, and that `ss.Callsign ` renames you on the scoreboard. Screenshot the scoreboard for `Docs/evidence/`. (After that, the weapons plan's W1 is the next build task.) --- ## Session 050 — 2026-09-28 — Licensing Settled (ADR-035), And W1: The A-Series Stop Being One Rifle ### COMPLETED - **ADR-035 (producer):** Southern Spear is free to play. Every asset the project holds is cleared: Fab, other free sources, and ADFRC from its creators. Real names are allowed everywhere: weapons, the Australian Army, its ranks and insignia, ADF equipment and camouflage. `LICENCE_REGISTER.md` gets a §0 current-position section and in-place amendments (L-0003 hold and L-0004 prohibition lifted; L-0007 narrowed to ripped/CAD sources; release gate revised; L-0021 cleared). CLAUDE.md's content rules are rewritten to match, and the GDD, ADFRC readmes and code comments are updated. These still stand, as non-licence rules: no endorsement claim, MAF portrayal, no *America's Army* content, no ripped assets. **R-57** records the accepted residual trademark/emblem risk. Pushed as 15bc062a. - **W1 per-weapon stats** (`Docs/WEAPONS_ANIMATION_PLAN.md`): - **Core:** `SSWeaponStats.h/.cpp` (`FSSWeaponStats`, `USSWeaponStatsSettings`, `FSSWeaponStatsRules`) and `Tests/SSWeaponStatsTests.cpp` (2 tests). - **Config:** eight rows from `Docs/WEAPON_SOURCE_DATA.md`. A88/A88G 682 rpm, 30 rounds; A89 750 rpm, a 200-round belt, spread ×1.74; A4/A416 857 rpm; A417 600 rpm, 20 rounds, ×1.18; A25 semi-auto, 20 rounds, ×0.75; A9 semi-auto, 15 rounds. - **Bridge:** `USSWeaponStatsSubsystem`. On the server, each new `ID_SS_*` item gets its magazine size, a full magazine and its spare rounds (Lyra's stat-tag functions, by reflection). On every machine, each new ranged weapon instance has its own copy of `HeatToSpreadCurve` scaled. Lyra is unmodified. - **Rate of fire and semi/full-auto are not applied yet.** Lyra keeps them in its fire-ability Blueprint. `Tools/Unreal/probe_weapon_fire.py` (new, read-only) dumps the Blueprint-declared variables along WID → ability sets → abilities, using the new `USSObjectivesEditorLibrary::ListPropertiesAsText`. ### FILES CHANGED - `Docs/DECISION_LOG.md` (ADR-035), `Docs/LICENCE_REGISTER.md`, `CLAUDE.md`, `Docs/GAME_DESIGN_DOCUMENT.md`, `Docs/Sourced/ADFRC/README.md`, `Docs/Sourced/ADFRC/ADFRC_CONFIG_REGISTRY.md`, `Docs/WEAPONS_ANIMATION_PLAN.md`, `Docs/WEAPON_SOURCE_DATA.md`, `Tools/Weapons/adfrc_weapon_data.py`, `SSInsigniaRaster.h`, `Config/DefaultGame.ini`. - Core: `SSWeaponStats.h/.cpp`, `Tests/SSWeaponStatsTests.cpp`. Bridge: `SSWeaponStatsSubsystem.h/.cpp`. Objectives editor: `SSObjectivesEditorLibrary.h/.cpp` (`ListPropertiesAsText`). `Tools/Unreal/probe_weapon_fire.py`. ### TESTING - `python Tools/validate_architecture.py`: **exit 0**. `python Tools/Weapons/adfrc_weapon_data.py`: **exit 0**, 8/8. `probe_weapon_fire.py` parses (`ast`). - **NOT RUN (no Unreal Engine here):** the build; `SouthernSpear.Core.Weapons.*` (2); a live check that the A89 shows 200/200 and the A25 20/80 on the HUD, and that the `LogSSWeaponStats` lines print; the probe. ### RISKS - **R-57 (accepted, ADR-035):** third-party trademarks and Commonwealth emblems; remedy if ever needed is a rename or a swap. - **R-60 (open, medium; was R-58, renumbered: Session 051 also used R-58/R-59):** W1 reaches Lyra by reflection (`GetStatTagStackCount`/`Add`/`RemoveStatTagStack`, `HeatToSpreadCurve`). A signature change fails soft, with a log line and Lyra's numbers kept. The first build and a live check confirm it. - **R-61 (open, low; was R-59):** a weapon whose ammo was granted before the subsystem's first pass (0.25 s) could fire one magazine at Lyra's size. Items are adjusted once, on first sight. ### DEFECTS FOUND - **Access violation in `USSPlayerProfileSubsystem::Deinitialize` (Session 049 code)**, found by the producer's test run after this session's build (22 actions, OK). 57 tests found, 20 completed (all Success, including both new `Core.Weapons.*` tests), then the run died in `SouthernSpear.Network.Gameplay.TwoPlayerAuthoritySmoke`. The cause: `ss.Callsign` was registered per game instance. The smoke test runs two, which share one console object; the first `Deinitialize` deleted it and the second unregistered the dangling pointer. **Fixed:** the command is now registered once per process (`FAutoConsoleCommand`) and applies to every local profile, and the profile subsystem no longer overrides `Deinitialize`. Also hardened: the server subsystem unbinds from the bus through a weak pointer kept at `Initialize`, not a lookup during world teardown. NOT RUN here; the re-run should report 57/57. ### ADDENDUM — rate of fire wired from the probe - The producer's local agent ran `probe_weapon_fire.py`. Lyra's fire ability keeps its interval in the Blueprint variable **`FireDelayTimeSecs`**; the A25 (semi-auto) needs a different ability, not a delay. Its report and `Docs/PLAYER_MODEL_PLAN.md` are on the producer's machine, not yet pushed. - `USSWeaponStatsSubsystem::ApplyFireRate`: for each pawn, the ability specs whose `SourceObject` is a ranged weapon instance (Lyra grants a weapon's abilities with the weapon as source, `LyraAbilitySet.cpp:117`) have each ability instance's `FireDelayTimeSecs` (double or float) set to 60 / rpm, once. It runs on every machine, because firing is predicted: A88 0.088 s, A89 0.080, A4/A416 0.070, A417/A25/A9 0.100. - Still pending: semi-auto for the A25 and A9 (grant the semi-automatic fire ability). NOT RUN: build, and a live check of the `LogSSWeaponStats` "rpm" lines. ### ADDENDUM 2 — semi-automatic A25, and IDs - `Tools/Unreal/setup_weapons.py`: the weapons whose stats row has `bFullAuto=False` (read from `DefaultGame.ini`: A25, A9) and that copy Lyra's rifle (the A25; the A9 is already pistol-based) get `AbilitySet_SS__Semi`. That is a copy of the rifle's ability set with `GA_Weapon_Fire_Rifle_Auto` swapped for `GA_Weapon_Fire_Pistol` (one shot per press, per the producer's probe). `WID_SS_A25` points at the copy. Paths are found from Lyra's own sets, not hard-coded. No Lyra asset is modified. Pure helpers checked here against the real config (semi = A25, A9) and sample property text. **NOT RUN:** the script in the editor, and whether the pistol ability's fire montage looks right on a rifle (third person). - Risk IDs renumbered: this session's R-58/R-59 became **R-60/R-61**, because Session 051 (the other agent, concurrently) also used R-58/R-59 for the soldier skeleton and vertex budget. ### NEXT ACTION **Build, run the tests, and run `Tools/Unreal/probe_weapon_fire.py`.** Then send me `Build/probe_weapon_fire.json` so I can wire rate of fire and semi/full-auto (the rest of W1) before W2. --- ## Session 051 — 2026-09-28 — The Player Model Diagnosed, And The Work That Actually Matters Written Down ### COMPLETED **Pulled** `f2c8870b` (the other agent's `ss.Callsign` fix) — clean fast-forward from `62f37cc6`. **Diagnosed the producer's model complaint** ("the shoulders look weird, the patch between the webbing and the waist is just a weird green") from the assets, the source scripts and the render. It is **not** a material-binding fault. `Crye_G3_Shirt_AMC_co.png` paints the G3 as a camo jacket over a plain olive under-shirt, and that under-shirt owns the whole torso, both forearms and the shoulder caps: - 53% of the sheet has no local detail, and the flat fill occupies luminance 73–86 where the real camo on the same sheet spans 14–124. - The band in the capture (`Build/zoom_torso.png`) measures `rgb(60,57,35)` with a local std of **1.7** — flat, not dark. Not a missing texture, not the 64×64 fallback, not the mannequin, not lighting. - The green patch, the smooth forearms and the shoulder seam are **the same pixels**: one cause, three symptoms. No shirt variant avoids it; AMP/DPC/DPD are the same Arma base mesh. - The bones are **fine**: all 28 uniform arm bones and 18 vest arm bones resolve against Manny's 164-bone skeleton. Leader pose is not the fault. The real structural issue is that every part is parented onto Lyra's mannequin, so the soldier can never own a bone. **Published `Docs/PLAYER_MODEL_PLAN.md`** — what the model is, the evidence for both complaints, the skeleton findings, an ordered plan (texture pipeline → own the skeleton → physics asset and LODs → the under-shirt decision), the producer decision on the body, six silent-failure traps, and the W1 hand-off with the exact `FireDelayTimeSecs` values the other agent needs (A88 0.088, A89 0.080, A4/A416 0.070, A417 0.100; A25 needs a different ability, not a delay). **Cheapest certain win identified:** `max_texture_size = 2048` in `setup_adf_soldier.py:texture()` is halving every 4096² ADFRC sheet. That is most of "the textures look bad" on its own. ### FILES CHANGED Created (all committed): `Docs/PLAYER_MODEL_PLAN.md`, `Tools/Textures/make_g3_shirt_camo.py`, `Tools/Unreal/probe_manny_bones.py`, `Tools/Unreal/probe_soldier_render.py`, this entry. Modified: `Docs/CHANGELOG.md` only. **Deliberately not committed:** `Art/Characters/Textures/T_SS_ADF_G3_Shirt_co.png` (13.8 MB) — the camo tool's output, wrong colours, bound to no material, regenerated in ~4 s. The project's other generated textures are committed because they are in use; this one is not. ### TESTING `git pull --ff-only origin main` — clean fast-forward, no conflicts. `Tools/Unreal/probe_manny_bones.py` — run headless, 164 bones read from `Skeleton.get_reference_pose().get_bone_names()`. Confirmed all 28 uniform and 18 vest arm bones present. `Build/probe_manny_bones.json`. `python Tools/Textures/make_g3_shirt_camo.py` — run, panel mask correct (36.6% of the sheet isolated), **pattern generator output rejected by the producer** on colour. Output not bound, no game change. **Build and `Automation RunTests SouthernSpear`: NOT RUN** — `f2c8870b` is pulled but not yet compiled on this machine. ### ASSETS No new licensed material. The camo tool samples its palette from the ADFRC sheet already held under L-0021 and invents no new pattern (ADR-033). Nothing added to the asset register. ### DEFECTS FOUND - `unreal.load_asset("/SSExp_ObjectiveAssault/…")` returns **None** in a `-run=pythonscript` commandlet — the game-feature plugin is unmounted, so probes over the ADF parts read nothing and still report success. A skeleton probe built on this reported 26 "bones" that were the characters of an error string. Removed rather than left in the tree. - `Skeleton.get_reference_skeleton()` returns a **`str`** in UE 5.8 Python, not a bone-name array. Working route recorded in `Docs/PLAYER_MODEL_PLAN.md` §6. - `verify_character_materials.py` calls `get_material_property_input_expression`, which does not exist in 5.8, and reads textures with `get_material_default_texture_parameter_value`, which cannot see graph-wired textures. It cannot gate anything. **NOT FIXED** — queued as P1.3 of the plan. ### RISKS - **R-58 (new):** the soldier is welded to Lyra's mannequin skeleton. Every ADF part is imported onto `/Game/Characters/Heroes/Mannequin/Meshes/SK_Mannequin`, so replacing Lyra's mannequin breaks all five parts, and the soldier cannot take its own clavicle/pec/corrective bones. Not a present fault — every bone resolves today — but it is the reason the model is hard to improve. - **R-59 (new):** the five soldier parts total 89,996 verts (23,187 uniform + 42,311 vest + 24,498 helmet) with **no LODs** and `create_physics_asset = False` on every import. The largest character cost in the game, and invisible in a single front-end screenshot. ### NEXT ACTION **Producer decision: what is the body?** The soldier is a Fab `Modern_Insurgent_7` head on an ADFRC G3 body under ADFRC gear, with the arms belonging to neither — there is no single source of truth, which is why it reads as assembled rather than worn. Recommended: take the head from ADFRC too, so head, arms, torso and gear are one source. Then P1 of `Docs/PLAYER_MODEL_PLAN.md` (the 2048 texture cap, the 64×64 fallback, and the broken material verifier).### ADDENDUM — build verified, P1.1 done, and a latent bug in the gear script **Build:** `Build.bat SouthernSpearEditor Win64 Development` → **Result: Succeeded**, 10 actions, covering the other agent's callsign fix and the new `USSWeaponStatsSubsystem` fire-rate code. **Tests:** `Automation RunTests SouthernSpear` → **57 found, 57 Success, 0 Fail, 0 NotRun, exit code 0**, no crash. `SouthernSpear.Network.Gameplay.TwoPlayerAuthoritySmoke`, which killed the run at 20 last time, **passes**. Re-run a second time after the content change below: still 57/57. **P1.1 — texture cap 2048 → 4096, applied.** `max_texture_size` in `setup_adf_soldier.py:texture()` was halving every ADFRC colour sheet, so `T_ADF_crye_g3_shirt_amc_co` and `T_ADF_crye_g3_pants_amc_co` (4096²) reached the screen at 1024². The cap can only downscale, so 4096 leaves the 1024 gloves and 2048 normals alone. The report now proves it rather than claiming it: a `textures` block lists every texture's size beside the cap set on the asset, and `report["ok"]` fails if `textures_halved != 0`. **89 textures, 0 downscaled.** ~140 `.uasset` files re-saved, which is the fix, not churn — the cap lives on the texture assets. **DEFECT FOUND — `setup_adf_soldier.py` was never idempotent.** `EditorAssetLibrary.does_asset_exist()` returns **False** for `/SSExp_ObjectiveAssault/…` paths (the game-feature plugin is not mounted in a commandlet) while `unreal.load_asset()` and `find_asset_data()` both resolve them. So `load_asset(p) if does_asset_exist(p) else create_asset(p)` took the **create** branch for assets that already existed; `create_asset` returned `None` and the run died on the MAF uniform slots, leaving `Build/adf_soldier_setup.json` with an empty `maf_uniform_slots` — which `setup_soldiers.py` reads, so a re-run of that script would have stripped the MAF soldier's green uniform. Fixed in `material_for()` and `fabric_master()` by branching on `load_asset()` returning `None`. The script now runs to completion: `ok: true`, 0 errors, all 31 material slots bound, `maf_uniform_slots` back to 4. `Build/` is gitignored, so the report could not be restored from git — it was regenerated. **Measured for P1.2:** 4 of the soldier's 31 material slots are flat 64×64 colour slabs — `safariland` (multicam → flat coyote) and the MAF `belt`, `tacgear` and `pasgt` (flat olive). The report's `note` field names them, so the evidence for deleting the fallback already exists. ### NEXT ACTION (Session 051 addendum) **P1.2 — delete the 64×64 flat fallback** in `setup_adf_soldier.py` and make an unbound `BaseColorMap` an error, so a failed texture lookup can never again look like a finished garment. Then P1.3, the broken `verify_character_materials.py`. The producer's body decision (§5 of the plan) is still open and gates P2. --- ## Session 052 — 2026-09-28 — W1 Finished In Code (Fire Rate, Semi-Auto), W2 Grip Sockets From The ADFRC Poses ### COMPLETED - **W1 rate of fire** (commit 72b374ad): `USSWeaponStatsSubsystem::ApplyFireRate` sets each weapon's own fire-ability instance `FireDelayTimeSecs` = 60 / rpm. The variable and its shared default (0.120 s on `GA_Weapon_Fire_C`) come from the producer's probe run (`PLAYER_MODEL_PLAN.md` §7). - **W1 semi-auto** (commit fe27ff23): `setup_weapons.py` gives the A25 an `AbilitySet_SS_A25_Semi` with Lyra's pistol fire ability in place of the rifle's auto one. The A9 is already pistol-based. - **W2 grip sockets:** `Tools/Common/adfrc_grip.py` (pure Python) rebuilds each handAnim pose's bone transforms and expresses both wrists in the `weapon` bone's space. It calibrates the axis map to the converted MLOD from the weapon's own `trigger_axis` and `muzzle_pos` (48 candidates, A3OB's (x, z, y) preferred on a tie), and refuses a pose that doesn't fit. `Tools/Blender/adfrc_weapon.py` carries the two points through the same transforms as the mesh and exports `SOCKET_LeftHandGrip` / `SOCKET_RightHandGrip`, recording the fit in `manifest.json → grip`. `Tools/build_adfrc_weapons.py` maps each weapon to its pose: A88 EF88_Vg_static, A88G AUG_GL, A4 ar15_8in_cgrip_static, A416 hk416_cgrip_static, A25 ar15_10in_cgrip_static, A89 Minimi_Standard. ### FILES CHANGED `Tools/Common/adfrc_grip.py` (new), `Tools/Common/test_adfrc_grip.py` (new), `Tools/Blender/adfrc_weapon.py`, `Tools/build_adfrc_weapons.py`, `Docs/WEAPONS_ANIMATION_PLAN.md`, `Docs/CHANGELOG.md`. (Also this session: `SSWeaponStatsSubsystem.h/.cpp`, `SSWeaponStats.h`, `Config/DefaultGame.ini`, `Tools/Unreal/setup_weapons.py`, already pushed in 72b374ad and fe27ff23.) ### TESTING - `python Tools/Common/test_adfrc_grip.py`: **exit 0**, 16/16 PASS. A synthetic rig (weapon bone rotated 35° and offset) recovers both wrists exactly in weapon space and picks the A3OB map (right hand 0.022 m from the trigger, left hand 0.048 m off the bore). It refuses a pose 1.7 m out, and every mapped clip is in the committed `ASSET_MANIFEST.json`. - `python Tools/validate_architecture.py`: **exit 0**. Both Python scripts parse. - **NOT RUN (needs the producer's machine):** `python Tools/build_adfrc_weapons.py` on the real MLODs and clips, so `grip.fit`, the right-hand-to-trigger distances and the sockets are unmeasured. Also not run: `setup_weapons.py` (sockets import with the FBX; semi-auto A25), the build, and the in-game checks of W1. ### ASSETS No asset files; the weapon FBXs are regenerated on the producer's machine (ADFRC, L-0021, ADR-035). ### RISKS - **R-62 (open, medium):** W2 assumes the weapon model's origin sits on the Arma `weapon` bone. If Arma places the proxy with an extra offset or rotation, the calibration won't fit and the weapon gets no sockets, and says so. It fails loud, not wrong. The first real run tells. - **R-63 (open, low):** the A25's semi-auto swap uses Lyra's pistol fire ability, whose third-person fire montage was made for a pistol. Check it looks right on a rifle. ### DEFECTS FOUND - My probe script listed only four weapons (A88, A89, A25, A9), which is why the producer's probe report said A4/A416/A417 were "absent". It was not a data gap. Noted rather than re-run: the rows exist and the subsystem reads them by item name. ### NEXT ACTION **On the producer's machine:** run `python Tools/build_adfrc_weapons.py`, then `Tools/Unreal/setup_weapons.py`, then build and test. Read each `Art/Weapons//ADFRC/manifest.json → grip` (`fit`, `right_hand_to_trigger_m`) and send them to me. Also check in game that the A89 shows 200/200, the A25 fires one shot per press, and the `LogSSWeaponStats` rpm lines print. --- ## Session 053 — 2026-09-28 — The "Asset Does Not Exist" Trap Swept Out Of Every Plugin Script ### COMPLETED - The producer's agent (fec930c1) found that `EditorAssetLibrary.does_asset_exist()` answers False for `/SSExp_ObjectiveAssault/...` assets in a commandlet while `load_asset()` resolves them. So a load-if-exists-else-create script tries to create an existing asset, and the run dies part-way. It fixed `setup_adf_soldier.py` and asked for a sweep of the rest. - New `Tools/Unreal/ss_assets.py`: `asset_exists(path)` = `does_asset_exist(path)`, then `load_asset(path) is not None` (the check proven on the producer's machine). Every `does_asset_exist` call in the 15 scripts that write under a plugin mount (`/SSExp_ObjectiveAssault`, `/SouthernSpearUI`) now uses it: 31 calls. Scripts that only touch `/Game` are unchanged. - This matters now because `setup_weapons.py`, the next script the producer runs (Session 052), had 7 such checks. Among them are the new `AbilitySet_SS_A25_Semi` copy and the `B_SS_*_Weapon` delete-then-create. ### FILES CHANGED `Tools/Unreal/ss_assets.py` (new); `setup_weapons.py`, `setup_ui.py`, `setup_maf_weapons.py`, `setup_flags.py`, `setup_objective_assault.py`, `setup_damage_model.py`, `setup_tactical_movement.py`, `setup_fp_arms.py`, `setup_fonts.py`, `setup_soldiers.py`, `setup_character_textures.py`, `upgrade_weapon_materials.py`, `build_objective_map.py`, `fix_visual_regressions.py`, `setup_adf_soldier.py` (all `Tools/Unreal/`); `Docs/CHANGELOG.md`. ### TESTING - `python3 -m py_compile` on every `Tools/Unreal/*.py`: **exit 0**, all parse. - `asset_exists` with a stub `unreal` module in which the registry misses an on-disk plugin asset: **True** for that asset, **False** for a missing one. - `python Tools/validate_architecture.py`: **exit 0**, PASS. - **NOT RUN (needs the producer's machine):** any of the 15 scripts in Unreal. The first real check is the Session 052 run of `setup_weapons.py`. ### ASSETS None. ### RISKS No new risks. A missing asset now also costs one `load_asset` miss, which may log a "failed to find" warning on a first run. That is harmless. ### DEFECTS FOUND - Latent in 14 more scripts: the same misleading existence check the producer's agent found in `setup_adf_soldier.py`. Found by grep after its report. None had failed yet, because each had only ever run once, or the registry happened to know the asset. ### NEXT ACTION Unchanged from Session 052. **On the producer's machine:** run `python Tools/build_adfrc_weapons.py`, then `Tools/Unreal/setup_weapons.py`, then build and test. Send each `Art/Weapons//ADFRC/manifest.json → grip`, and check in game: A89 200/200, A25 one shot per press, and the `LogSSWeaponStats` rpm lines. ### ADDENDUM — weapons pipeline run: the grip fit produced nothing, and setup_weapons.py has its own blocker **Ran `python Tools/build_adfrc_weapons.py`. All seven weapons built, but the grip fit failed on every one of them**, so there are no grip numbers to send: | Weapon | `fit` | `reason` | clip | |---|---|---|---| | A88, A4, A416, A25, A89 | `false` | no axis map puts the left hand on the barrel ahead of the trigger | per-weapon | | A88G | `false` | pose has no bone(s): lefthand, righthand | `AUG_GL` | | A9 | *(no grip key at all)* | — | — | The same reason on five different rifles means the axis-map heuristic never produces a mapping, not that five weapons are awkward. A88G's pose clip has no hand bones at all. The A9 was not attempted. **W2 is blocked at its first step: the manifests carry no grip data.** `fit` and `right_hand_to_trigger_m` cannot be sent because they do not exist. **Ran `Tools/Unreal/setup_weapons.py`: it fails on the first weapon**, before any grip code, and this is independent of the above. `visual_blueprint()` deletes `B_SS_A88_Weapon` and immediately recreates it; `eal.delete_asset()` does **not** remove the blueprint in this commandlet — the file is still on disk after the delete — so `create_asset()` returns `None` for a name that is still taken. The `None` then reached `k2_gather_subobject_data_for_blueprint()`, which returned an empty array, and `handles[0]` died with an opaque `IndexError`. Ruled out by direct probe (`Build/probe_weapon_bp_parent.json`), so the next attempt need not repeat them: it is **not** the game-feature plugin being unmounted (a blueprint created in the plugin path and one in `/Game` both report 2 subobjects), **not** delete-then-recreate (2 subobjects after), and **not** the parent class (it resolves: `/ShooterCore/Weapons/Rifle/B_Rifle.B_Rifle_C`, 2 subobjects from a fresh blueprint). The one difference is that the real run deletes an asset that is genuinely there. Fix applied here, minimally: compile the blueprint before gathering subobjects, check the delete actually took, check `create_asset` returned something, and never index `handles[0]` unchecked. **The script still cannot complete** — the delete behaviour needs a real fix (build in place, or delete through a route that works headlessly). The half-finished weapon assets were restored with `git checkout`; only the intended FBX and manifest output of `build_adfrc_weapons.py` is left modified. **Build and tests on `a99a8692`:** `Target is up to date`. `validate_architecture.py` → PASS, no violations. `Automation RunTests SouthernSpear` → **57 found, 57 Success, 0 Fail, exit code 0, no crash.** **Body decided — ADR-036.** The ADFRC G3 is the body: head, arms, torso and gear from one pack, Fab `Modern_Insurgent_7/SK_Head` off the friendly soldier, `SKM_QuantumCharacter` retired as a candidate. This unblocks P2. NOT DONE: the head swap itself, which needs an ADFRC head in `Art/Characters/ADF/` first. ### NEXT ACTION (Session 054 addendum) **On the other agent: the grip fit.** The axis map never fits on any rifle, so that is a code fix, not a data problem. **Still needs a human: the live match** (A89 shows 200/200, the A25 fires one shot per press with 20 rounds, `LogSSWeaponStats` prints the rpm lines) — none of that can be checked headlessly. ## Session 055 — 2026-09-28 — Both Session 054 Blockers Fixed: Weapon Blueprint Built In Place, Grip Fit Tries Arma's Skeleton ### COMPLETED - **`setup_weapons.py` no longer deletes the weapon blueprint.** `visual_blueprint()` loads an existing `B_SS__Weapon` and updates its `SSVisual` component (mesh, +90° yaw, no collision). It adds the component only when it is missing, and creates the blueprint only when there is none. The Session 054 guards stay in place: compile before the gather, and fail loudly on a missing parent class, a failed create or empty handles. The report says `created: True/False` per weapon. - **Grip fit, with three causes found by reading the decoded pose data in `ASSET_MANIFEST.json` and `rtm_rigs.py`:** 1. **A88G:** `AUG_GL`'s rig uses capitalised names (`LeftHand`, `RightHand`) and the lookup was case-sensitive. Bones are now matched case-insensitively. 2. **Every rifle (probable):** `rtm_rigs.py` parents `weapon` to `righthand` by a naming rule. The clips list their bones in Arma's own skeleton order (`pelvis, spine…spine3, camera, weapon, launcher, neck…`). In Arma's `OFP2_ManSkeleton` (from memory, not from a file in the repo), `weapon` and `launcher` are children of `Spine1` and `Camera` of `Pelvis`. Composing a Spine1-relative weapon bone under the hand puts it somewhere meaningless, and no axis map can fit that. `adfrc_grip.py` now tries both hierarchies (`arma` first, then `decoded`) and reports which one fits. 3. **No numbers on failure:** a refusal said only why. It now carries the hand span, the trigger-to-muzzle length, the nearest failing axis map with its distances, and each hierarchy's hands in weapon space. The trigger and muzzle memory points are included too. One failed run now diagnoses itself. - `python Tools/Common/adfrc_grip.py --export Docs/evidence/w2_grip_clips` copies each grip clip's first frame and its rig into small JSONs. It prints both hierarchies' hand positions. Committed, these let the maths be checked without the git-ignored `Animations/` tree. ### FILES CHANGED `Tools/Unreal/setup_weapons.py`, `Tools/Common/adfrc_grip.py`, `Tools/Common/test_adfrc_grip.py`, `Docs/CHANGELOG.md` (also fixes the Session 054 addendum heading, which had been joined onto the previous line). ### TESTING - `python Tools/Common/test_adfrc_grip.py`: **exit 0, 23/23 PASS** (16 before). New checks: - the `arma` hierarchy re-parents `weapon` to `Spine1`; - capitalised bone names resolve; - on a synthetic rig stored Spine1-relative with the decoder's `weapon -> RightHand` parents, `grip_points` picks `arma` and reports the failed `decoded` attempt with its numbers; - a refusal carries the hand span and barrel length. - `python3 -m py_compile` on `setup_weapons.py`, `adfrc_weapon.py` and `adfrc_grip.py`: exit 0. `python Tools/validate_architecture.py`: exit 0, PASS. - **NOT RUN (the producer's machine):** `build_adfrc_weapons.py` on the real clips, and whether the Arma hierarchy is the actual cause. It is the probable one, not a proven one. Also not run: `setup_weapons.py` in Unreal, including `SubobjectDataBlueprintFunctionLibrary.get_variable_name`, which is used here for the first time. ### ASSETS None. The `--export` JSONs are ADFRC pose data (L-0021, ADR-035) once committed. ### RISKS - R-62 (W2 assumes the weapon model's origin sits on the `weapon` bone) still stands. If both hierarchies fail with the full report, R-62 is the next suspect, and the exported clips let it be checked here. ### DEFECTS FOUND - Case-sensitive bone lookup (A88G). Found by reading the rig bone lists in `ASSET_MANIFEST.json`. - The pose maths trusted `rtm_rigs.py`'s rule hierarchy for the attachment bones. Found by comparing each clip's bone order with Arma's skeleton order. - A refused fit reported no numbers, so a 7/7 failure could not be diagnosed remotely. Found by the producer's run. - `setup_weapons.py` deleted and re-created the blueprint, which cannot work when the delete silently fails. Found by the producer's agent (Session 054 addendum). Fixed by building in place. ### NEXT ACTION **On the producer's machine:** 1. Pull. 2. Run `python Tools/build_adfrc_weapons.py`, then `python Tools/Common/adfrc_grip.py --export Docs/evidence/w2_grip_clips`, and commit that folder along with the seven `manifest.json` files. 3. Run `Tools/Unreal/setup_weapons.py`, then build and test. 4. Report each `manifest.json → grip`: `fit`, `hierarchy`, `right_hand_to_trigger_m` or `attempts`. Also report whether `setup_weapons.py` completes with `ok: true`. ### ADDENDUM — Quantum modular character assessed (ADR-037): good modules, blocked toolchain The producer asked whether the main player model could pivot to the Quantum **modular** character. **ADR-036 was wrong on the fact it decided on**, and ADR-037 withdraws that point. `Content/QuantumCharacter/Mesh/Modules/` is a genuinely modular set — eight separate meshes, one or two clean slots each: `SKM_Head` (5), `SKM_Arms`, `SKM_Shirt_RolledUp_Blue`, `SKM_Jeans`, `SKM_Bulletproof_Bege`, `SKM_Holster_Hard_Bege` (2), `SKM_Drops_1_Bege`, `SKM_Patch_Back`. Each ships a physics asset, the eight total 27 MB against 22 MB for the three ADF parts, and the pack has its own locomotion set. That is Lyra's character-parts shape, and it gives the head from the same pack as the body — which is what ADR-036 wanted. ADR-036 retired the *assembled* mesh on its 14 tangled slots and nobody had looked at `Modules/`. **The catch is the skeleton.** `SK_Military_Character_Skeleton` has **351 bones**: 159 shared with Manny's 164, **192 Quantum-only** (fingers, toes, face). Not leader-pose compatible. **The prototype the producer authorised — modules on Manny, ADF camo on shirt and jeans — is blocked on tooling.** All three routes fail: - **No source meshes.** The pack ships **230 `.uasset`, zero FBX/OBJ**, so the Blender re-rig route that produced the ADFRC gear cannot be repeated. - **No reparameterise API** in UE 5.8 Python. Checked `SkeletalMeshTools`, `SkeletalMeshEditorSubsystem`, `AnimationLibrary`, `EditorSkeletalMeshLibrary`, `SkeletalMeshUtilitiesLibrary` — none exposes it. - **`SkeletalMeshExporterFBX` crashes the editor**: hard assert `MeshObject` (`SkinnedMeshComponent.cpp:4987`) inside `MeshMergeUtilities`, killing the commandlet with no Python traceback. Reproduced **with and without `-nullrhi`**, so not a headless artefact. **The only route that needs no re-export is to adopt the 351-bone skeleton itself** and use the modules unmodified with our own ABP — the larger risk §P2 already flagged, and not what was authorised. Left undecided for the producer. No committed file changed by this investigation; the probe scripts were removed and the tree is clean. ### NEXT ACTION (Session 055 addendum) **Producer decision: re-scope the Quantum pivot to the 351-bone skeleton, or leave it held open.** It is the only path that works with today's toolchain. If held open, the fallback is a UE bug report for `SkeletalMeshExporterFBX` on high-bone-count meshes, since a working export is what unblocks the cheap Manny prototype.### ADDENDUM — weapons pipeline run on `63bd2089`: setup_weapons now completes, but the grip fit is still 0 of 7 — and the exported clips disprove the Spine1 theory **`Tools/Unreal/setup_weapons.py`: `ok: true`, 36 steps, 0 errors.** The in-place blueprint fix works; the delete that could never take is gone. (The first attempt reported `could not create B_SS_A88_Weapon` — that was my own fault, a 30-second timeout killing the commandlet mid-run. Re-run properly, it completes.) **Grip fit: still 0 of 7.** The improved diagnostics are what make this useful, and the eight exported clips (`Docs/evidence/w2_grip_clips/`, committed) settle the cause: 1. **The `arma`/`Spine1` hierarchy is contradicted by the clips themselves.** Every one of the eight exported clips carries a `parents` array, and in **all eight** `weapon`'s parent is the right hand (`righthand`, or `RightHand` in the two AUG clips). `spine1` is index 2. The attachment does not need recalling from memory — each file states it. The "arma" attempt parents the weapon to a bone these files say it is not on. 2. **The right hand carries no per-weapon information, which is why the search rejects everything.** The `decoded` right hand is byte-identical across all five rifles at `[-0.9651, 0.2163, -0.44]`. In the raw clip data the cause is visible: `righthand`'s frame is identical across the five clips that share a stance — `ar15_10in`, `ar15_8in`, `hk416`, `hk417` all carry `q = [-0.179498, -0.05286, -0.018…]`, and the root is identical too. Only `lefthand` varies per weapon. Arma's rifle clips share one authored stance and vary only the support hand, so any test leaning on the right hand accepts or rejects all weapons identically — and parenting the weapon to that same hand makes its weapon-space position degenerate by construction. 3. **The left hand, the only per-weapon signal, is ~1.3–1.5 m forward of the weapon origin** in every clip — beyond the muzzle on weapons 0.4–0.7 m long (`barrel_m` 0.399–0.6855). Whatever the attachment question, the pose frame or its units are wrong independently of it. 4. `A88G` is a different rig: hand span **1.13 m**, and a distinct right hand. It is not a two-handed rifle hold in the same sense as the others. **Build and tests on `7a086366`:** `Result: Succeeded`. `Automation RunTests SouthernSpear` → **57 found, 57 Success, 0 Fail, exit 0, no crash.** `python Tools/Common/test_adfrc_grip.py` → **exit 0, 23/23 PASS.** **NOT DONE, still needs a person playing:** the live match — A89 shows 200/200, the A25 fires one shot per press with 20 rounds, `LogSSWeaponStats` prints the rpm lines. ### NEXT ACTION (Session 056 addendum) **On the other agent: the grip maths, from the committed clips.** The Spine1 hierarchy is refuted by `parents[weapon]` in every file; drop it or re-derive it from the data. And key the fit off the **left** hand and the weapon transform, because the right hand is identical across clips sharing a stance and cannot discriminate. The left hand's ~1.4 m forward offset is the next thing to explain. ## Session 057 — 2026-09-28 — W2 Grip Fit: 6 Of 6, Once The Pose Data Is Read The Way It Was Written ### COMPLETED - **The committed clips (`Docs/evidence/w2_grip_clips`, Session 056 addendum) showed what the decoded pose data holds.** Neither my Spine1 theory nor the addendum's "constant right hand" reading was the root cause: 1. **The stored transforms are not bone offsets.** In a parent-relative skeleton a child's offset is its bone length, pose-independent. `lefthand`'s ranges 0.08–0.42 m across clips of one rig. 2. **Each bone's transform is a rotation about its own rest joint, relative to its parent.** Every arm bone's transform has one fixed point, the same in every clip: its rest joint. Solving `(I − R)·x = p` over the eight clips gives: - wrists at x = ±0.587 m, mirror-symmetric; - elbows at ±0.37 m and shoulders at ±0.064 m; - feet about 0.9 m below the origin. The residual is under 1 mm. That's Arma's rest skeleton, recovered from the animation data. 3. **The quaternion handedness is wrong in the decoded files.** Rotations must be read as `(−x, −y, z, w)`. A search over all 96 sign and order variants found this as the only reading that makes the fixed points consistent. The summed residual over ten arm joints falls from 1.04 m as stored to under 0.001 m. 4. **`weapon` hangs off the body.** Its transform is identical in every rifle clip, because the arms move to it. Under the body, the right wrist lands at the same place in weapon space in every clip (1–4 cm) and the left wrist moves forward along one axis. The addendum's point 1 (the file's `parents` array says `righthand`) reads the decoder's own naming rule back (`rtm_rigs.py` `link("weapon", "righthand")`), so it isn't Arma's data. - **The posed skeleton now looks right.** Shoulders are at 1.35 m and the neck at 1.40 m, with both wrists in front at chest height. Wrist-to-wrist spans run from the EF88 bullpup at 0.24 m to the AR-15 10-inch at 0.35 m. The 8-inch and 10-inch rails differ by 3.7 cm, and the rails differ by 2 inches (5 cm). - **`Tools/Common/adfrc_grip.py` rewritten on that model.** - Rotations are read as `(−x, −y, z, w)`. - Transforms are composed down the hierarchy with `weapon` under `Spine1`. - Each wrist is placed as its rest joint (`REST_JOINTS`, reproducible with `--solve-rest`) carried by its posed transform. - The hands are expressed in the weapon's frame and mapped to the MLOD by the one proper rotation between rest space (x left, −y forward, z up) and the model (barrel axis, z up). The 48-map search, which could choose reflections, is gone. - The weapon's rest placement (its proxy) is not in the clips. So the right wrist is anchored at `trigger_axis` and the left wrist placed relative to it. The fit test is that the left wrist lands on the barrel: ahead of the trigger, short of the muzzle, within 0.15 m of the bore. - `load_clip` falls back to the committed evidence when the Animations/ tree isn't present. **Result on the real clips and manifests:** | Weapon | Pose | Span | Left wrist forward | To bore | |---|---|---|---|---| | A88 | EF88_Vg | 0.242 | 0.219 m | 0.103 | | A88G | AUG_GL | 0.293 | 0.277 m | 0.095 | | A4 | ar15_8in | 0.311 | 0.293 m | 0.105 | | A416 | hk416 | 0.338 | 0.319 m | 0.113 | | A25 | ar15_10in | 0.349 | 0.332 m | 0.106 | | A89 | Minimi | 0.301 | 0.293 m | 0.070 | The left wrist sits 6–9 cm to the left of the bore, which is where a left wrist sits beside a handguard. ### FILES CHANGED `Tools/Common/adfrc_grip.py`, `Tools/Common/test_adfrc_grip.py`, `Docs/WEAPONS_ANIMATION_PLAN.md`, `Docs/PROJECT_AUDIT.md` (R-32 note), `Docs/CHANGELOG.md`. ### TESTING - `python Tools/Common/test_adfrc_grip.py`: **exit 0, 24/24 PASS.** - Synthetic: rest joints recovered exactly; the weapon composes under Spine1; the decoder's hierarchy gives a different answer; the rest-to-MLOD rotation is proper (det +1) with forward down the barrel on three axis layouts; a left wrist behind the trigger is refused. - Real data: rest joints reproduced from the committed clips (rms 0.8 mm and 0.25 mm), mirror symmetry, **all six weapons fit** on their committed `trigger_mlod`/`muzzle_mlod`, the 8-inch span is shorter than the 10-inch, and every clip is in the manifest. - `python Tools/Common/adfrc_grip.py --solve-rest`: lefthand [0.5872, 0.0743, 0.0665] rms 0.00081 m, righthand [−0.5858, 0.0744, 0.0659] rms 0.00025 m. - **NOT RUN (the producer's machine):** `build_adfrc_weapons.py` with the new module (sockets written into the FBX), `setup_weapons.py`, the build, and left-hand IK in game. The sockets' absolute position in the hand (the trigger anchor) is unverified; their spacing is what the pose measures. ### ASSETS None. ### RISKS - **R-64 (open, high):** the decoded animation tree (`Animations/Rig/*`) and anything built from it inherit two errors: the quaternion handedness, and the claim that BMTR transforms are parent-relative bone offsets. This includes the `Animations_UE` FBX clips and any retarget (W5 `GestureReloadAUG`, R-32). Fix it in `rtm_rigs.py` / `anim_to_fbx.py` before any Arma clip is retargeted, and check it with the same fixed-point test. - **R-32 (update):** the Arma rest joint *positions* are recoverable from the clips themselves (fixed points, under 1 mm), so `SkeletonPivots.p3d` may not be needed for positions. Rest *orientations* are still unmeasured. Left OPEN. - **R-62 (resolved by design):** W2 no longer assumes the model's origin is on the `weapon` bone. It anchors the right wrist at `trigger_axis` instead. ### DEFECTS FOUND - The quaternion handedness in the decoded clips, and the parent-relative interpretation in `rtm_rigs.py`. Found by the fixed-point consistency test on the committed evidence. - My Session 055 fix (the Spine1 re-parenting) was right about the parent but insufficient: it still read the transforms as bone offsets with the wrong handedness. Found by the producer's run (0 of 7 again) and the evidence it committed. ### NEXT ACTION **On the producer's machine:** 1. Pull. 2. Run `python Tools/build_adfrc_weapons.py` and check each `manifest.json → grip.fit`. Expect `true` for A88, A88G, A4, A416, A25 and A89. 3. Run `Tools/Unreal/setup_weapons.py`, confirm `SOCKET_LeftHandGrip`/`SOCKET_RightHandGrip` exist on each `SM_*` mesh, then build and test.### ADDENDUM — grip fit confirmed 6 of 6 on real data; A25 semi-auto was working all along; ADR-038 corrects ADR-037 **Grip fit, real data: 6 of 6 `fit: true`.** A88 0.2418 · A88G 0.2929 · A4 0.3113 · A416 0.3381 · A25 0.3486 · A89 0.3013 m hand span — matching the published table. `Tools/Common/test_adfrc_grip.py`: **0 failures** (24 checks). `setup_weapons.py`: **ok true**, 36 steps, 0 errors, each weapon mesh carries its left- and right-hand grip points. **`A25:semi_auto` was reported failing but was already correct.** The step said "0 set(s) now fire ..._Pistol_C". Reading the granted abilities directly: the A25's equipment definition grants `GA_Weapon_Fire_Pistol_C` through its own `AbilitySet_SS_A25_Semi` — the semi-auto fire ability **is** in place. The step could never report success twice: it only acts on sets that still contain `GA_Weapon_Fire_Rifle_Auto`, so after the first successful swap there is nothing left to replace, `swapped` stays 0, and `ok = swapped > 0` is false on every subsequent run. **Fixed** to accept the end state ("already correct") as success, and the report now distinguishes swapped-this-run from already-correct. `setup_weapons.py` is **ok: true, 36 steps, 0 errors** after the fix. **Build and tests on `41348693`:** `Result: Succeeded`. Automation: **57 found, 57 Success, 0 Fail, exit 0, no crash.** **ADR-038 — I was wrong about leader pose, and the producer was right.** ADR-037 recorded the Quantum modules as "not leader-pose compatible" and the pivot as blocked. `SetLeaderPoseComponent` matches bones **by name**; child bones with no counterpart in the leader are simply not driven and hold their reference pose. My probe asked "does the child have bones the leader lacks" (192 of 351) and treated the answer as "unusable", which is not the test. Leader pose would drive the **159 shared bones and leave 192 static**, and the mesh renders fine. The genuine cost is the one identified independently: straight fingers on a rifle. The `IKRetargeter` route also needs no export. **Neither route has been run.** **The three new candidate bodies, measured in Blender (`Build/probe_candidate_bodies.json`):** all three are in the Fab library cache as **source files** (FBX/GLB), which is the one thing better than Quantum. None is better or easier than what we have: | | Rig | Bones | Verts | Slots | Textures | Manny-compatible | |---|---|---|---|---|---|---| | Free Pack – Male Base Mesh (FBX) | **none** | 0 | 4,395 | 1 | **none** | n/a | | FSB Operator (GLB) | Mixamo | 52 | 108,616 | 41 | yes | **0/52** | | SWAT Operator (GLB) | Mixamo | 51 | 57,761 | 34 | yes | **0/51** | The Male Base Mesh is the **worst** of the three, not the easiest: no armature and no textures, so a rig would have to be built from nothing. FSB is the heaviest. SWAT is the best of the three and still not good enough — 34 slots, no modularity, the same straight-finger problem, and its materials reference **KSVR**, a real Russian camouflage brand, which is an accuracy problem for the main player model as well as a trademark one under R-57. ### NEXT ACTION (Session 057 addendum) **Prototype the Quantum modules on the leader (fast, straight fingers) or on an IKRetargeter (slower, articulated hands), in camo, beside the current soldier** — that is the one comparison that settles ADR-036, and the producer's own recommendation is not to switch on paper. The new packs are assessed and none of them displaces Quantum. --- ## Session 058 — 2026-09-28 — W2 Finished In Code: The Left Hand Goes To The Grip Socket, First And Third Person ### COMPLETED - **`USSHandIKMeshComponent`** (`Plugins/SouthernSpearLyraBridge`, the one module allowed to touch Lyra): a skeletal mesh component that puts its left hand on the held weapon's `SOCKET_LeftHandGrip` (Session 057). - **Where it runs.** It overrides `FinalizeBoneTransform`: after each animation evaluation, before the pose is published, it runs a two-bone IK on the left arm in component space. Everything that follows the mesh by leader pose reads that published pose, including the visible 3 ACR / MAF soldier parts and the physics bodies. **No Lyra asset changes and no Animation Blueprint is needed**, which matters because the repo has no headless route to author one. - **Where the target comes from.** Each tick, before the evaluation, it finds the attached static mesh that carries the grip socket. It stores that socket's transform relative to the socket or bone the weapon hangs from; both come from the same (last) frame, so the offset is exact. After evaluation it rebuilds the target from the new pose of that bone, so there is no frame lag. - **The solve (`FSSHandIK::Apply`, pure, written here rather than calling the engine's).** The elbow bends in the plane of the animated elbow. There is no stretching, and the hand stops at full reach. The hand keeps its animated rotation, and twist bones and fingers are carried with their bones. The effector blends by alpha. - **When it switches off.** It fades out at 8 per second: - while a montage named Reload, Equip, Holster, Draw or Inspect plays; - while Lyra's `DisableLHandIK` curve is up; - while ragdolled; - when the target is beyond 1.05× the arm's length; - when the weapon has no grip socket (the A9 pistol). `ss.HandIK 0` turns it off for side-by-side comparison. - **Bones** are found from candidate names, so one class serves Manny (`upperarm_l`, `lowerarm_l`, `hand_l`) and the Fab first-person arms (`LeftArm`, `LeftForeArm`, `LeftHand`). It logs the chain it found, or warns once and stays off. - **Wired in two places:** - `ASSCharacter` swaps its body mesh class, as it already does for movement: `SetDefaultSubobjectClass(ACharacter::MeshComponentName)`. - `USSFirstPersonSubsystem` creates the first-person arms as the same class. `Play()` suppresses the IK while a Draw, Holster or Reload clip plays, since those clips move the left hand themselves. The subsystem's one-off grip measurement uses the left hand only in a log line, so it is unaffected. - A Python port of the solver was run on the test's own scenarios before handing over. The hand landed on the target with error 0.0, bone lengths held to 1e-15, the finger and twist offsets were unchanged, full reach was 57.462 cm against a 57.463 cm limit, and alpha 0.5 landed half-way. ### FILES CHANGED `Plugins/SouthernSpearLyraBridge/Source/SouthernSpearLyraBridge/Public/SSHandIKMeshComponent.h` (new), `Private/SSHandIKMeshComponent.cpp` (new), `Private/Tests/SSHandIKTests.cpp` (new), `Private/SSCharacter.cpp`, `Private/SSFirstPersonSubsystem.cpp`; `CLAUDE.md` (test list); `Docs/WEAPONS_ANIMATION_PLAN.md`; `Docs/CHANGELOG.md`. ### TESTING - `python Tools/validate_architecture.py`: **exit 0**, PASS. - New automation test `SouthernSpear.Bridge.HandIK.Solve`. It checks: - the hand lands on a reachable target, and the shoulder stays fixed; - both lengths are kept, and the hand keeps its rotation; - the finger and twist bone are carried, and bones outside the arm are untouched; - the elbow stays on the upper arm's axis and the wrist on the forearm's axis; - an out-of-reach target stops at full reach, and alpha 0.5 lands half-way; - alpha 0, a broken chain and mismatched parents are all refused and leave the pose unchanged. - Python port of `FSSHandIK::Apply` on those scenarios: all as expected (above). - **NOT RUN (the producer's machine):** the build (C++ written here, uncompiled). Also not run: the automation suite (expect 58), and the in-game look in both views, including `LogSSHandIK` naming the chain for `CharacterMesh0` and `SS_FirstPersonArms`. ### ASSETS None. ### RISKS - **R-65 (open, medium):** the IK relies on `FinalizeBoneTransform` being called with the new pose still in the editable buffer. That is how `USkeletalMeshComponent` publishes a pose, but it is unverified here. If the hand lags a frame or doesn't move, that is the place to look. The alternative is a post-process Animation Blueprint on a copy of the mannequin mesh. - **R-66 (open, low):** the right wrist is anchored at `trigger_axis` (Session 057), so how deep the weapon sits in the palm is an estimate. If the left hand looks consistently a few cm off on every weapon, adjust the anchor once in `adfrc_grip.py`. The spacing between the hands is measured, not estimated. ### DEFECTS FOUND None new. ### NEXT ACTION **On the producer's machine:** build and run the tests (expect 58/58). Then, in a match, compare `ss.HandIK 1` against `ss.HandIK 0` with the A88 in both views (`-SSShotAt` captures). Also check the left hand leaves the grip for a reload and returns after it. --- ## Session 059 — 2026-09-28 — W3 Shell Ejection, And The Hand IK Was Looking For The Wrong Socket Name ### COMPLETED - **Found and fixed a defect in Session 058's hand IK before it cost a build.** Unreal's FBX importer strips the `SOCKET_` prefix from Blender empties: `SOCKET_Muzzle` arrives as `Muzzle`, which is the name every existing lookup uses. The hand IK was looking for `SOCKET_LeftHandGrip`, which no imported mesh has, so the left hand would never have moved. `USSHandIKMeshComponent::GripSockets` now tries `LeftHandGrip`, then `SOCKET_LeftHandGrip`. `setup_weapons.py` now reports, per weapon, which of `Muzzle`, `LeftHandGrip`, `RightHandGrip`, `Eject` and `EjectEnd` exist, so the next run proves the names. - **W3 shell ejection, first half:** - **Sockets.** `Tools/Blender/adfrc_weapon.py` writes `SOCKET_Eject` at the ejection port (`nabojnicestart`) and `SOCKET_EjectEnd` where the case is thrown (`nabojniceend`). Every committed weapon manifest lists both memory points. Two sockets, not one rotated socket, so the throw direction survives the FBX axis conversion. It is read at runtime as the direction between them. The manifest gets an `eject` entry, or the reason there is none. - **`USSShellEjectSubsystem`** (Lyra bridge, clients only; `ss.Casings 0` turns it off). On each rifle or pistol fire cue, `ASSCharacter` asks it to throw one case. The pistol cue also covers the semi-auto A25, which uses the pistol's fire ability. - The case leaves the `Eject` socket towards `EjectEnd` at 2.5–3.8 m/s, with some lift, a little rearward, and the shooter's own velocity. - The local player's case comes from the first-person weapon (the only-owner-see view model); everyone else's from the third-person one. - It flies ballistically with light drag and spin. A line trace per moving case, per frame, bounces it off the world at 35% restitution and 35% friction. It settles on its side and lies there for 30 s or until its slot in the 48-case pool is reused. - A case still in flight after 6 s (off the map) is dropped. The step is capped at 1/30 s so a long frame can't tunnel a case through the floor. - Presentation only (ADR-004): no collision, no gameplay effect. - **The case** is the engine cylinder in a brass tint (`BasicShapeMaterial` with its `Color` parameter), sized per calibre from the mesh name: 5.56×45 by default, 7.62×51 for the A25/A417, 9×19 for the A9, and 7.62×39 for the MAF weapons. No new asset. - A weapon without the sockets throws nothing; there is no guessed port. ### FILES CHANGED `Tools/Blender/adfrc_weapon.py`, `Tools/Unreal/setup_weapons.py`; `Plugins/SouthernSpearLyraBridge/Source/SouthernSpearLyraBridge/Public/SSShellEjectSubsystem.h` (new), `Private/SSShellEjectSubsystem.cpp` (new), `Private/Tests/SSCasingTests.cpp` (new), `Private/SSCharacter.cpp`, `Public/SSHandIKMeshComponent.h`, `Private/SSHandIKMeshComponent.cpp`; `CLAUDE.md`; `Docs/WEAPONS_ANIMATION_PLAN.md`; `Docs/CHANGELOG.md`. ### TESTING - `python Tools/validate_architecture.py`: **exit 0**, PASS. - `python3 -m py_compile Tools/Blender/adfrc_weapon.py Tools/Unreal/setup_weapons.py`: exit 0. - New automation tests: - `SouthernSpear.Bridge.Casings.Motion`: gravity, a floor bounce (105 up, 65 along, spin 12), a slow case resting, and a case thrown from 1.4 m resting within 3 s after at least 2 bounces, 1–4 m out. - `SouthernSpear.Bridge.Casings.Calibre`: mesh name to weapon, and the calibre per weapon. - A Python port of `Step`/`Bounce` on the thrown scenario: rests at 1.17 s after 5 bounces, 2.77 m to the side. - A unity-build name check over the bridge found no clashes for the new file-level names. - **NOT RUN (the producer's machine):** the build (expect 60 tests: 57 + HandIK 1 + Casings 2). Also not run: `build_adfrc_weapons.py` (the eject sockets), `setup_weapons.py` (its `sockets` report), and cases in game in both views. ### ASSETS None (engine cylinder and material). ### RISKS - **R-67 (open, low):** the brass tint assumes `/Engine/BasicShapes/BasicShapeMaterial` exposes a `Color` vector parameter. If it doesn't, cases render the material's default (pale grey). A brass material asset fixes it. ### DEFECTS FOUND - The hand IK looked for `SOCKET_LeftHandGrip` while imported meshes carry `LeftHandGrip`. Found while wiring the eject sockets: every existing lookup of the muzzle socket uses `Muzzle`, which only works if the importer drops the prefix. ### NEXT ACTION **On the producer's machine:** 1. Pull, then run `python Tools/build_adfrc_weapons.py` and `Tools/Unreal/setup_weapons.py`. Check the report's `sockets` table: every rifle should have `LeftHandGrip`, `Eject` and `EjectEnd`. 2. Build and run the tests (expect 60/60). 3. In a match, fire the A88 in both views: check the left hand is on the handguard (`ss.HandIK 0/1`) and cases leave the right side and land. --- ## Session 060 — 2026-09-28 — W3 Muzzle Light: Every Shot Lights Its Surroundings ### COMPLETED - **A muzzle flash already existed.** Lyra's fire cue plays its flash sprite and tracer at its hidden rifle's muzzle, and `ASSCharacter::AlignLyraMuzzle` (an earlier session) moves that muzzle onto our visible barrel, in first person onto the view model's. What a real shot adds, and Lyra's sprite doesn't, is light. - **`USSMuzzleLightSubsystem`** (Lyra bridge, clients only; `ss.MuzzleLight 0` turns it off). On each rifle or pistol fire cue, `ASSCharacter` flashes a warm point light (1.0, 0.62, 0.28) just ahead of the muzzle of the weapon the viewer sees. It lights the shooter's hands and face and the ground and walls near the muzzle, and full-auto fire strobes. - It fades quadratically from its peak to zero over the flash (`FSSMuzzleLightRules::Intensity`). - Peak intensity varies ±20% per shot. - The flash per weapon (`SpecFor`): | Weapons | Peak | Radius | Duration | |---|---|---|---| | 5.56 rifles | 2,500 cd | 7 m | 45 ms | | 7.62×51 (A25/A417) | 3,500 cd | 9 m | 55 ms | | A9 pistol | 1,200 cd | 4.5 m | 35 ms | - Eight pooled lights, no new asset. Shadows are off by default for cost; `ss.MuzzleLightShadows 1` turns them on. - **Shared lookup.** "Which weapon mesh does this viewer see for this shooter" moved from the casing code to `FSSWeaponPresentation::FindViewerWeapon`. The first-person view model is found for the local player, the third-person weapon for everyone else. The imported socket names (`Muzzle`, `Eject`, `EjectEnd`) are tried before the Blender `SOCKET_` names. Casings and the muzzle light both use it. - `Docs/ASSET_REGISTER.md`: E-001 muzzle flash describes what exists now, and new row E-005 is spent cases. ### FILES CHANGED `Plugins/SouthernSpearLyraBridge/Source/SouthernSpearLyraBridge/Public/SSMuzzleLightSubsystem.h` (new), `Private/SSMuzzleLightSubsystem.cpp` (new), `Public/SSWeaponPresentation.h` (new), `Private/SSWeaponPresentation.cpp` (new), `Private/Tests/SSMuzzleLightTests.cpp` (new), `Private/SSShellEjectSubsystem.cpp`, `Public/SSShellEjectSubsystem.h`, `Private/SSCharacter.cpp`; `CLAUDE.md`; `Docs/ASSET_REGISTER.md`; `Docs/WEAPONS_ANIMATION_PLAN.md`; `Docs/CHANGELOG.md`. ### TESTING - `python Tools/validate_architecture.py`: **exit 0**, PASS. - A unity-build name check over the bridge found no clashes for the new file-level names. - New automation test `SouthernSpear.Bridge.MuzzleLight.Rules`: - the light is at its peak at the shot, at a quarter by half time, and dark at and after its duration; - zero duration gives no light, and the fall is monotonic; - a pistol is dimmer, smaller and shorter than a rifle, and 7.62 brighter and wider than 5.56; - every flash is under 60 ms. - **NOT RUN (the producer's machine):** the build. Expect **61 tests**: 57 existing, HandIK 1, Casings 2, MuzzleLight 1. Also not run: the look in game. The brightness is a first estimate; judge it indoors and at dusk. ### ASSETS None. ### RISKS - **R-68 (open, low):** a daylight scene's exposure may make 2,500 cd barely visible outdoors while indoors it reads strongly. That is correct behaviour, but check it looks intended. The values are in `SpecFor`. ### DEFECTS FOUND None new. ### NEXT ACTION **On the producer's machine, one combined run for Sessions 058–060:** 1. Pull, run `python Tools/build_adfrc_weapons.py` and `Tools/Unreal/setup_weapons.py`, and check the report's `sockets` table (every rifle: `LeftHandGrip`, `Eject`, `EjectEnd`). 2. Build and run the tests (expect 61/61). 3. In a match, with the A88 in both views, check: - the left hand is on the handguard (`ss.HandIK 1` against `0`); - cases leave the right side and land; - each shot lights the surroundings (`ss.MuzzleLight 1` against `0`, best indoors). --- ## Session 061 — 2026-09-28 — Website: A Place For In-Engine Captures, a Roadmap Meter and a Press Kit ### COMPLETED - **In-engine screenshot pipeline, ready for the first captures.** Frozen captures go in `Docs/images/screenshots/` with an entry in its `screenshots.json` (file, title, alt, map, captured date, optional session and note; the README there gives the steps). `python Tools/build_site_assets.py --only screenshots` writes AVIF/WebP/JPEG derivatives at 960 and 1920 px to `Site/assets/screenshots/` and the view model `Site/data/screenshots.json`, listing each file at its real width. An entry missing `alt`, `map` or a `YYYY-MM-DD` date, or pointing at a missing file or an LFS pointer, stops the build (exit 1). `build_site_assets.py` gained `--only ` so one step can run without the other sources. - **Media section** now opens with *In-engine captures*, rendered by `site.js` from `data/screenshots.json` with a teal `In-engine` badge, the map, date and session under each, the newest at full width, and a work-in-progress note. With no captures it shows a dashed "none published yet" note. The concept art below keeps its orange labels; the section note now says only `In-engine` images come from the game. - **Lightbox** resolves its triggers when opened rather than once at start-up, so script-rendered captures join the same previous/next sequence. - **Roadmap meter** above the timeline: one segment per phase, coloured by the state the roadmap records, each linking to its phase, with "1 of 7 phases complete · Phase 1 in progress". No percentage (the Development section says none is tracked). - **Development pulse**: sessions recorded, first session and latest session, counted from the changelog. - **FAQ "What PC will I need?"**: says minimum/recommended specs are not measured yet and will be set by testing and listed on Steam; lists only what `Config/DefaultEngine.ini` fixes (DX12 + SM6, Lumen with hardware ray tracing and software fallback, virtual shadow maps, TSR) and the TDD §9.1 target (60 fps at 1080p on an RX 9070 XT), labelled as a target, not a minimum. - **Press kit page** `presskit.html`: fact sheet, one-line/short/long descriptions drawn from existing site copy, the logo, emblem and artwork as downloads (artwork labelled as artwork), contact (Discord, GitHub) and the disclaimers. Linked from every footer, added to `sitemap.xml` and to `publish_site.py`. ### FILES CHANGED `Site/index.html`, `Site/site.js`, `Site/styles.css`, `Site/changelog.html`, `Site/presskit.html` (new), `Site/sitemap.xml`, `Site/README.site.md`, `Site/data/screenshots.json` (new); `Tools/build_site_assets.py`, `Tools/publish_site.py`; `Docs/images/screenshots/README.md` (new), `Docs/images/screenshots/screenshots.json` (new); `Docs/Website/WEBSITE_DESIGN_SYSTEM.md`; `Docs/CHANGELOG.md`. ### TESTING - `python3 Tools/build_site_assets.py --only screenshots` on a scratch copy with two fixture images (one 1672 px, one 2560 px wide): exit 0, 12 derivatives, view model lists the 1672 px file at 1672w. - Same, with an entry missing `alt`, `map` and `captured`: **exit 1**, names the entry and the missing fields. - Headless Chromium 1194 over `Site/` served locally, 1440 px and 390 px, reduced motion: no page or console errors on `index.html`, `changelog.html` and `presskit.html`; no horizontal scroll at either width; captures render and open in the lightbox, arrow keys step into the static gallery; meter reads "1 of 7 phases complete · Phase 1 in progress"; pulse reads 63 / 26 Sep 2026 / 28 Sep 2026; with the empty view model the "none published yet" note stays and `aria-busy` clears. Every screenshot was looked at. - **NOT RUN:** `python Tools/publish_site.py` (not published from this session); the full `build_site_assets.py` (the sources in `Docs/images/` are LFS pointers in this checkout); the `Build/audit/text_audit.js` font-floor check (not present in this checkout). ### ASSETS None. The press kit links existing derivatives only. ### RISKS - **R-69 (open, low):** the press kit's largest logo is the 420 px lockup. Press usually wants a large transparent logo; a 1024 px press derivative from `logo.png` would fix it, and needs the LFS source. ### DEFECTS FOUND - None in existing code. Found in review of the new code: a source narrower than 1920 px was advertised in `srcset` as 1920w. Fixed before commit by recording each file's real width. ### NEXT ACTION **On the producer's machine:** carry out Session 060's NEXT ACTION (the weapon checks, the build, both views in a match), and in that same run capture the first in-engine screenshots with `-SSShotAt` or `HighResShot 1920x1080` (game viewport only, no window frame), then follow `Docs/images/screenshots/README.md` and publish. --- ## Session 062 — 2026-09-29 — The animation decoder is fixed (R-64), and the guards that missed it Four workstreams were requested together and run in one session. Their file sets do not overlap, and their ID ranges were reserved rather than guessed: the decoder takes **R-70–R-74** (ADR-040 stays free — no decoder decision needed one), the two guard checkers take **R-75 onward**. The standing rules stand: one branch per workstream, merged to `main` only when that workstream's own tests pass. This session ran on `main` in a shared checkout; its work is one commit, rebased onto the W3 and site commits before pushing - which is why this entry is 062 and not the 060 it was written as. ### COMPLETED **1. R-64 — the ADFRC animation decoder, fixed and tested (pure Python; no Unreal, no Blender).** `Docs/Sourced/ADFRC/rtm_rigs.py` carried the two errors Session 057 found in the decoded data. Both are fixed, and the fix is the same model `Tools/Common/adfrc_grip.py` already uses: - **The quaternion reading.** `rotation_from_stored(q)` now returns `mat_from_quat((-x, -y, z, w))`, and `world_from_local` uses it. Reading the components as stored leaves every bone's transform without a consistent fixed point (0.007–0.17 m per arm joint on the committed clips). - **The transform model.** A stored transform is a rotation of the bone **about its own rest joint**, relative to its parent, so `p = J - R J` and the bone's joint is `R·J + p`. `solve_rest_joints` recovers `J` as the least-squares fixed point of `(I - R) J = p` over every frame of every clip on the rig; a new `posed_joint` helper states the model directly. `solve3` gained a **relative conditioning guard**, because the residual alone does not catch an under-determined joint: the AUG shoulder solved to −21 m with a 0.16 mm residual, which would have been written into the armature as a bone head. - **The rig JSON** is now `adfrc-rig/2`: `rest_world` is the solved rest joints (falling back to the mean world position only for a bone the clip set never rotates — an identity rotation pins no joint down), with `rest_joints`, `rest_joint_rms_m` and `rest_joints_solved` alongside. - `rtm2json.py`'s `notes` no longer tell a reader the transforms are parent-relative bone offsets. `anim_to_fbx.py` holds no pose maths (it consumes the rig JSON), so it needed only a docstring note that the decoder applies the correction once. **Measured on the committed evidence (`Docs/evidence/w2_grip_clips`).** The eight clips are two rigs: six share a bone list, the two AUG-family clips order theirs differently. Per rig: | Rig | Arm joints solved | Worst arm-joint rms | Wrists mirror | Stored (x,y,z,w) reading | |---|---|---|---|---| | 6 clips (EF88/A4/A416/A25/A89/Minimi) | 9 | **0.062 mm** | 0.3 mm | 0.007–0.159 m | | 2 clips (AUG, AUG_GL) | 8 | **0.956 mm** | — | 0.007–0.166 m | `lefthand`'s stored translation ranges 0.081–0.416 m across the reference rig's poses — a bone offset could not vary; its *joint* is one point to 0.03 mm. The recovered wrists are at x = ±0.586 m, and the decoder and `adfrc_grip.py` agree on them to 1e-6 m. The AUG family is a different skeleton: its left elbow fixed point differs by ~15 mm and its right arm is not mirror-symmetric (R-71). **2. Architecture guard — new rule SS010: Core holds shared types, not content.** `Tools/validate_architecture.py` now scans `Plugins/SouthernSpearCore/Source` and fails any line that names a content path (`/Game/`, `/SSExp_`, `/ShooterCore/`, `/SouthernSpearUI/`) or loads an asset (`LoadObject`, `ConstructorHelpers`, `FSoftObjectPath`, `FSoftClassPath`, `StaticLoadObject`). A module that reaches content must be the module that owns it — the structural version of the Session 059b call that moved a capture harness out of Core. One existing hit is **accepted and reported as a NOTE, not a failure**: `Public/SSFonts.h` loads the UI font faces from `/SouthernSpearUI/Fonts`. It predates the rule, every UI module already depends on Core, and moving it is a separate change — tracked as **R-75**. Every *new* hit is a violation. `Finding` gained an `allowed` flag, `--json` gained `notes` and an `SS010` count, and the human report prints notes after the verdict. `Tools/test_architecture_guard.py` is the CLAUDE.md negative test done properly: a scratch copy of `Tools/` and the SS plugins, a Core asset load injected (exit 1), a Core content path injected (exit 1), the accepted SSFonts.h still a NOTE, and removing the injection back to exit 0. `CLAUDE.md`'s architecture section documents the rule and the exception. **3. A unity-build name clash checker.** `Tools/check_unity_names.py` walks every module under `Plugins/SouthernSpear*`, collects the names defined inside anonymous namespaces and as file-scope `static`s in `.cpp` files, and flags any name defined in **more than one file of the same module** — the thing that is fine when files compile separately and a redefinition when Unreal merges them. It also flags, best-effort, a local variable that shadows a function or member defined in the same file (the Session 048 `Settings`/C4459 case). Exit 1 on a finding; `--json` for the report. It found a real one on its first clean run: `StatCount` is an anonymous-namespace helper in **both** `SSHudStateSubsystem.cpp` and `SSScoreboardSubsystem.cpp` (module `SouthernSpearLyraBridge`). They have different signatures, so a unity build would have accepted them as overloads — but the two names mean different things and the rule is name-based, so they are now `ItemStatCount` and `PlayerStatCount`. The checker is **0 findings across 8 modules** afterwards. `Tools/test_check_unity_names.py` covers the fixture cases (duplicate helper, duplicate file-static, shadow, unique name, exit codes). `CLAUDE.md`'s build-and-test list gains this command and both negative tests. **4. Risk register reconciled.** `Docs/PROJECT_AUDIT.md` now carries **R-50–R-66** from Sessions 048–058: R-50 (Session 048 uncompiled) and R-56 (Session 049 uncompiled) closed by the later builds; R-62 closed by design (Session 057); R-64 closed here; the rest OPEN, each with the changelog's own wording. No status was invented — where the changelog did not state one, the row is OPEN. `CLAUDE.md`'s "Open risks" line was rewritten to match, and no longer lists R-56 as unverified. ### FILES CHANGED - Decoder: `Docs/Sourced/ADFRC/rtm_rigs.py`, `rtm2json.py`, `anim_to_fbx.py`; new `Tools/Common/test_rtm_rigs.py`. - Guards: `Tools/validate_architecture.py` (SS010 + `allowed` findings), new `Tools/test_architecture_guard.py`; new `Tools/check_unity_names.py`, new `Tools/test_check_unity_names.py`; `Plugins/SouthernSpearLyraBridge/Source/.../SSHudStateSubsystem.cpp` and `SSScoreboardSubsystem.cpp` (`StatCount` renamed). - Docs: `CLAUDE.md` (architecture SS010, build/test list, Open risks line), `Docs/PROJECT_AUDIT.md` (R-50–R-66, date), `Docs/CHANGELOG.md`. ### TESTING - `python Tools/Common/test_rtm_rigs.py` — **exit 0, 18/18 PASS**: the stored reading leaves the arm chain without a fixed point; the (−x,−y,z,w) reading resolves it; every solved arm joint is under 1 mm; `R·J + p = J`; the wrists and upper arms mirror; the decoder and `adfrc_grip` agree; and the writer round-trips a rig JSON through `main()` in a temp tree. - `python Tools/Common/test_adfrc_grip.py` — **exit 0, 24/24 PASS** (unchanged, as R-64 required). - `python Tools/validate_architecture.py` — **exit 0**, one SS010 NOTE (SSFonts.h). - `python Tools/test_architecture_guard.py` — **exit 0, 6/6 PASS**. - `python Tools/check_unity_names.py` — **exit 0**, 8 modules, 0 findings (was 1 real clash before the rename). - `python Tools/test_check_unity_names.py` — **exit 0, 9/9 PASS**. - `python -m py_compile` on `rtm_rigs.py`, `rtm2json.py`, `anim_to_fbx.py`, `validate_architecture.py`, `check_unity_names.py` — **exit 0**. - **NOT RUN — needs the producer's machine or Blender:** the full decode over the 165 clips (it needs a clean `Animations/Rig/` output; `main()` was exercised only on a temp tree), the Blender FBX export, the editor build (including the `StatCount` rename), `Automation RunTests SouthernSpear`, and the in-game checks. R-64's code is fixed and tested; **its output artifacts are not regenerated yet**. ### ASSETS None. No asset imported, created or modified. ### RISKS - **R-70 (open, medium):** the rest-joint solve only pins a bone the clip set actually rotates. On the grip clips 25–38 of 67 bones solve; the rest fall back to the mean world position, so a rig re-decoded from a small or static clip set has an approximate skeleton for bones that never move. A full locomotion set is what pins them. - **R-71 (open, medium):** the AUG-family clips (`AUG`, `AUG_GL`) are a **separate rig** — a different bone order, a left elbow fixed point ~15 mm from the other six, and a right arm that is not mirror-symmetric. Consistent with Session 055's "A88G is a different rig" (1.13 m hand span). Treat the AUG handAnim poses as their own skeleton until proven otherwise. - **R-72 (open, low):** the rig JSON schema moved to `adfrc-rig/2` and `rest_world` changed meaning (solved joints, not accumulated translations). The decoded tree and any FBX already built from it must be regenerated; stale `adfrc-rig/1` output should be discarded. - **R-73 (open, low):** the decoder was fixed and unit-tested, but the full 165-clip decode and the Blender FBX export were not run here. The first producer run is the integration test. - **R-75 (open, low):** `Public/SSFonts.h` is a Core header that loads `/SouthernSpearUI/Fonts` assets. Accepted by SS010 and reported as a NOTE; move it to a UI module (which may then depend on it) when convenient. - **R-76 (open, low):** the unity-name checker is a line-based heuristic. It does not see a name that a macro introduces, a generated file, or a name defined only inside a class in the `.cpp`; and it flags same-name overloads that a unity build would in fact accept. It is a tripwire, not a proof. - No risk number was guessed: the decoder used R-70–R-74, the guards R-75 onward. ### DEFECTS FOUND - **The decoder's two errors (R-64), fixed.** Found by Session 057's fixed-point test and carried into code here; the AUG shoulder's −21 m "solution" showed that a residual check alone is not enough. - **A real unity-name clash,** `StatCount` in two files of `SouthernSpearLyraBridge`, found by `check_unity_names.py` on its first clean run. Renamed. - **The first run of the checker had two false-positive classes** (multi-line declarations read as two declarations; anonymous-namespace constants read as locals), found by running it against the real, already-building tree and requiring zero findings. Both are fixed; the fixture test locks them in. ### NEXT ACTION **On the producer's machine:** pull, then run `python Tools/check_unity_names.py`, `python Tools/validate_architecture.py`, `python Tools/test_architecture_guard.py` and `python Tools/Common/test_rtm_rigs.py`, and regenerate the ADFRC decoded tree and FBXs off the fixed `rtm_rigs.py` (`python Docs/Sourced/ADFRC/rtm_rigs.py`, then the Blender export) so R-64's output artifacts match its code. Then build and run `Automation RunTests SouthernSpear` to confirm the `StatCount` rename compiles. --- ## Session 063 — 2026-09-28 — Quantum prototype: the gate is answered, and it is mostly no; and W2 hand IK does not compile-block, but will not run (The Quantum work is ADR-039. The hand IK finding is the important part of this entry.) **W2 hand IK, first real compile: passes.** `3a6364e8` builds clean. Automation is **58 found, 58 Success, 0 Fail**, exit 0, including the new `SouthernSpear.Bridge.HandIK.Solve`. Architecture guard passes. The maths is fine; the wiring is not, and it fails silently. **R-65 is worse than "might lag a frame": the IK will never run.** `USSHandIKMeshComponent` overrides `FinalizeBoneTransform()`. Searching the whole engine, that method is called from `USkeletalMeshComponent::TickAnimation` only in the `TickFunction == nullptr && ShouldBlendPhysicsBones()` branch - a manual/editor refresh - and from editor tools (Sequencer, FBX export), physics, MovieScene and the new Animation Constraints system. `PostAnimEvaluation` does not call it; it calls `DoInstanceFinalizeAnimation` -> `UAnimInstance::FinalizeAnimation`. **Nothing in the normal game animation frame calls it.** The override will not be entered in a match, so the left hand will not move, and the component logs nothing on the way - it reports the bones it found and then does nothing, which is the most expensive failure shape in this project. The fallback the producer named - a small Animation Blueprint on a copy of the mannequin - is the right shape of answer, and the C++ can stay: `UPoseableMeshComponent` is already proven in this session (ADR-039) as the per-bone lever that 5.8 keeps, so an Anim BP that drives the pose, or a component that overrides `RefreshBoneTransforms` and edits the editable component-space buffer before calling Super, both reach the same place. I have not changed the component: it is the other agent's code and the choice between those is theirs. **This is the one thing to fix before anyone looks for the hand in a match.** **Hand IK bone names resolve on every mesh it is wired to** (`Tools/Unreal/check_handik_bones.py`, `Build/handik_bones.json`). Body `SKM_Manny` -> `upperarm_l > lowerarm_l > hand_l` (164 bones, indices 11/12/20); first-person Rifle and Pistol arms -> `LeftArm > LeftForeArm > LeftHand` (55 and 61 bones, indices 26/27/28). So the `LogSSHandIK` line the producer asked me to look for will read correctly for both - it just will not print, because the component never ticks into `ResolveBones` in a match. The ancestry half of the check is a topological-order proxy, not `FSSHandIK::IsAncestor`, and is labelled as such in the report. **Free finding: the Quantum body needs no new arm bone names.** Quantum's `SKM_Jeans` resolves `upperarm_l > lowerarm_l > hand_l` on the component's *existing* body-side list (351 bones, indices 11/12/15), because Quantum uses the UE mannequin names. If the prototype is ever adopted, the third-person hand IK works on it unchanged; only the first-person arms, being a separate mesh, would need their own candidates. **R-66 stands unmeasured.** Nothing here needed the palm-depth constant, because the IK never ran. **Screenshots, again.** The in-match checks the producer asked for - A88 with `ss.HandIK 1` and `0`, both views, one reload - cannot be run: `Tools/run_map_capture.sh` still stalls at 32 log lines before `LoadMap` on every map, with and without `-nullrhi`, with `-RenderOffscreen`, with `-NoLoadingScreen` (ADR-039). That check stays open until both the capture path and R-65 are fixed. **Housekeeping, and a mistake of mine.** My Session 058 / ADR-039 text was lost partway through this session: I wrote a backup to a Git Bash `/tmp` path that Windows Python cannot see, and the `open()` failed, but the `git checkout` that reverted the docs had already run. Recovered by rewriting it - this entry is renumbered to **059** because the other agent's hand IK took 058. The lesson for the next session: do not revert tracked files until the backup has been read back and verified, and on this machine use a project-relative path rather than `/tmp` for anything Windows tooling has to open. **Traps added.** `SK_Mannequin` is the Skeleton and `SKM_Manny` is the mesh, and they share a name, so `LoadObject` on the former returns a `USkeleton` and every cast fails silently. A UFUNCTION with a `bool` return **and** a `FString&` out-param is uncallable from 5.8 Python - the bool becomes `None` and the out-param is dropped - so return the string. And overriding `FinalizeBoneTransform()` on a `USkeletalMeshComponent` looks like the documented post-animation hook and is not, because 5.8 never calls it during a game frame. ## Session 063b — 2026-09-28 — Producer review: three of four points were right, and the fourth exposed a vacuous proof The producer reviewed the Quantum report and raised four corrections. Three were correct and one was based on a stale read. All four are answered here; ADR-039 has been rewritten to match. **1. The reference-pose question is settled, and I was wrong about the order.** Unreal composes child component space as `Local * ParentComponentSpace` - `UPoseableMeshComponent::FillComponentSpaceTransforms` says it in a comment and calls `FTransform::Multiply(Dest, Local, ParentCS)`. I had it the other way round. `GetRefBonePose()` always returned parent-relative transforms; the multiply order was the bug. After the fix, `upperarm_l` to `hand_l` reads **51.87 cm** in both skeletons, which is a correct upperarm-plus-forearm span, against the 144.4 cm it reported before. **The producer's real point was better than the one they made.** A, B and C all compared the retarget against reference poses composed by *the same function* that produced them, so **none of them could detect a wrong composition** - they were self-referential, and "rest drift 0.0001 cm" was an artifact rather than a result. A check that shares a function with what it is checking proves nothing. There is now a guard for the class: the proof composes the reference pose and checks it against absolute human proportions - a forearm 20-35 cm, a head 55-80 cm above the pelvis - and fails the report outright otherwise. It earned its keep immediately, by rejecting my own first band, written on the assumption that `lowerarm_l` sat at the wrist. It does not: it sits at the elbow, so the span is the forearm, 27.25 cm. After the fix, all four modules: sanity passes, A drift 0.0001 cm, B 30.0 of 30.0 degrees, and C **27.2511 -> 27.2511 cm** - the bone length preserved exactly, not within a tolerance. **2. The stall is not the prototype, and not a stray process - proven, not asserted.** The producer was right to challenge "pre-existing". A working commandlet log goes from the ini loads directly to `LogPluginManager: Found matching target receipt`; every `-game` log stops before printing that line. So the `-game` process stalls **before any Southern Spear module loads**, which means my prototype and the hand IK are both incapable of causing it - neither can run before the point it stops at. `tasklist` shows zero Unreal processes. I then tested the game-target hypothesis, and it is *not* the cause: building the target changed nothing about the stall. **But the detour found a real break.** The **game target had never been built on this machine and did not compile**: `SSObjectiveTests.cpp` and `SSSectionAssaultTests.cpp` pass `NAN` and `INFINITY` - compile-time constants - into tests that are about non-finite values, and a game-target build folds them into constant arithmetic (MSVC C4756). Replaced with `std::numeric_limits::quiet_NaN()` and `infinity()`, which is also what those tests mean. `SouthernSpear Win64 Development` now builds and `SouthernSpear.exe` links. The screenshot is still outstanding and I still cannot produce it. **3. Moved out of Core.** `ASSQuantumProtoStage` is now in `SouthernSpearLyraBridge`. The producer's argument is right and mine was thin: Core is the root of the SS dependency graph and holds the shared types every other module uses, and a capture harness naming specific Quantum and mannequin assets does not belong there even though it touches no gameplay state. `validate_architecture.py` passed it in Core because the guard reads module dependencies, not what the code does - which is itself worth recording. **4. The hand IK was compiled, and the finding stands.** The 57 of 57 was from before the pull; after `3a6364e8` the suite is **58 found, 58 Success, 0 Fail**, including `SouthernSpear.Bridge.HandIK.Solve`, re-confirmed after this session's changes. The producer's belief that the IK "hooks the step where the body mesh publishes each frame's pose" is the part that does not hold: `FinalizeBoneTransform()` is called nowhere in the game animation frame - `PostAnimEvaluation` calls `DoInstanceFinalizeAnimation` -> `UAnimInstance::FinalizeAnimation` instead. That finding is unchanged and is the one thing to fix before looking for the hand in a match. **State.** Build succeeds, architecture guard passes, 58/58. Nothing committed. ## Session 063c — 2026-09-29 — The capture works, the comparison is shot, and the critique's five fixes are in **The capture stall was never a stall.** Two compounding misdiagnoses, both mine. First, `run_map_capture.sh` violated two rules of `Docs/PLAYTEST_COMMANDS.md`: it launched `UnrealEditor-Cmd.exe` instead of `UnrealEditor.exe`, and it passed neither `-abslog` nor `-FORCELOGFLUSH`, so a killed run lost its log tail and the default log was being shared with other agents' commandlets. The "stall at 32 lines" was a lost tail plus the wrong log file. Second, the first "successful" capture rendered a black frame with the HUD on it, because the map had been saved while `ASSQuantumProtoStage` still lived in SouthernSpearCore: on load the class failed to resolve (`CreateExport: Failed to load Outer` for Camera, Sun, Fill, Platform), so nothing took the view and nothing lit the scene. Re-running `setup_quantum_proto.py` re-saved the map against `/Script/SouthernSpearLyraBridge.SSQuantumProtoStage` (verified in the umap) and the same recipe then produced a full frame. The script now follows §3 of the doc exactly. **The producer's critique of the first real frame was right on every point**, and the scratch probe (`Build/probe_camo_chain.py` -> `Build/probe_camo_chain.json`) found the cause of the white shirt: the camo master and both instances were created and assigned in commandlet memory but **never saved** - `main()` saved only the four meshes, so the process exit deleted the material out from under the slots. `M_SS_ADFRC_Camo`, `MI_SS_ADFRC_Camo_Shirt` and `MI_SS_ADFRC_Camo_Jeans` now persist in `/SSExp_ObjectiveAssault/Characters/QuantumProto/` and are re-assigned on every run. Also fixed in the stage: the `SKM_Arms` module (the hands - the shirt is rolled-up sleeves ending at the forearm), the ADFRC vest and helmet leader-posed onto the Quantum body (the same-kit test), a matte floor and matte plinth (WorldGridMaterial read as a wet mirror), stronger fill, a full-length camera pull-back, and a per-tick view takeover because Lyra's pawn takes the view back after BeginPlay (the producer's window showed their own first-person glove mid-frame). **The next run crashed on an engine ensure** (`SkinnedMeshSceneProxyDesc.cpp:455`: a leader-pose component whose bone map does not cover the follower's whole ref skeleton) while spawning the soldier into the stage map - the hand-IK component attaching to the camera is now reaching the stage map through the game feature's pawn. That crash is the boundary of this session, not a fix in it. The producer's window of the same build shows both bodies standing on the plinth, and the frame the producer took is saved as `Docs/evidence/qproto/quantum_vs_g3_captured.png`. **Still open, honestly:** the vest/helmet on Quantum, camo-on-body and hands-on-Quantum are all built into the stage but not yet *seen* in a saved capture, because of that ensure. The mic-boom-in-nose defect and the FP/third-person camo mismatch belong to the other agent. The recommendation stands: keep G3 (ADR-036) unless the fixed shot shows Quantum in the same kit clearly beating it. --- ## Session 064 — 2026-09-29 — W5 Reload Tooling, And A Trap Between The Fixed Decoder And The Grip ### COMPLETED - **A trap between R-64's decoder fix and W2's grip, closed before it fired.** The fixed decoder (`rtm_rigs.py`, Session 062) now converts rotations while decoding and writes *standard* quaternions. It still labelled its output `adfrc-anim-local/1`, the tag the old stored-convention files carry. `adfrc_grip.py` reads the stored convention. So the first weapon build after the full 165-clip decode would have flipped every rotation twice and put the hands in the wrong place, silently, since the grip fit reports only whether a pose fits. - The decoder now writes `adfrc-anim-local/2`. - `adfrc_grip.load_clip_frames` converts /2 files back to the stored convention on load. - /1 files and the unlabelled committed evidence are left as they are. - **W5 tooling (the A88 bullpup reload):** - `adfrc_grip.load_clip_frames` returns every frame of a clip, from the decoded tree or from `Docs/evidence/w5_reload_clips`. - `python Tools/Common/adfrc_grip.py --export-frames Docs/evidence/w5_reload_clips` exports the four reload clips (`GestureReloadAUG`, `…Prone`, `MPP_Fast_Reload`, `MPP_Slow_Reload`) in full. - `Tools/Common/adfrc_reload.py [clip] [samples]` turns a reload clip into the left wrist's path relative to the right-hand grip, in the weapon's own axes: forward from the grip to the muzzle, the shooter's right, and up. It writes `Build/reload_path_.json` with the sampled keys, the start and end offsets, the static grip pose's offset (the game uses it to check its right-axis sign), the distance from the path's start to the grip, the largest per-frame step, and the magazine switch phase (0.48 for the EF88 family, from the config registry). - `Tools/Blender/probe_weapon_parts.py ` lists a weapon MLOD's named selections with vertex counts. It answers whether the magazine is a separate part that can come off in the hand. - **What the reload clips hold** (from `ASSET_MANIFEST.json`): - `GestureReloadAUG` and `…Prone` have 165 frames each, in both trees. The Source copies are absolute RTM; the Workshop copies are BMTR. - Their 66-bone rig has `weapon`, both arms and hands, and no magazine bone. Arma moves the magazine through the weapon model (its `magazine` selection is hidden and swapped at `magazineReloadSwitchPhase`). ### FILES CHANGED `Docs/Sourced/ADFRC/rtm_rigs.py` (schema /2), `Tools/Common/adfrc_grip.py`, `Tools/Common/test_adfrc_grip.py`, `Tools/Common/adfrc_reload.py` (new), `Tools/Common/test_adfrc_reload.py` (new), `Tools/Blender/probe_weapon_parts.py` (new), `Docs/CHANGELOG.md`. ### TESTING - `python Tools/Common/test_adfrc_grip.py`: **exit 0, 26/26.** New checks: a /2 copy of a real clip gives the same hands as the stored evidence (1e-9), and /1 and unlabelled frames pass through untouched. - `python Tools/Common/test_rtm_rigs.py`: exit 0. - `python Tools/Common/test_adfrc_reload.py`: **exit 0, 10/10.** - On a synthetic reload the path is recovered exactly: 25 cm forward and 5 cm left at the start; 5 cm forward and 12 cm down at the magazine well half-way; back to the start; no frame step over 3 cm. - The weapon axes are forward down the barrel, z up and the shooter's right. - A one-frame clip is refused. - On the real A88 grip pose, the left hand is 22.2 cm forward, 9.4 cm left and 1.5 cm below the right. - `python3 -m py_compile` on the new scripts: exit 0. - **NOT RUN:** the reload clips themselves (they exist only on the producer's machine), the Blender probe, and anything in Unreal. ### ASSETS None. ### RISKS - **R-77 (open, medium):** the reload path is relative to the weapon, so it can only drive the left hand once the hand IK actually runs. It doesn't in 5.8 (R-65, Session 063). W5 waits on the R-65 fix. ### DEFECTS FOUND - The decoder's /2 output was still labelled /1. Found by reading the fixed decoder's writer before planning W5. Fixed at both ends. ### NEXT ACTION **On the producer's machine:** fix R-65. Confirm in the 5.8 source which post-evaluation hook runs in a game frame (see the prompt in the Session 064 reply), and build the replacement. --- ## Session 065 — 2026-09-29 — R-65 Answered From The 5.8 Source; The Full Decode; Three W5 Probes Four tasks, run in order on `e4fff018`. The R-65 answer is the one that changes a decision, so it comes with the engine quotes and line numbers and **no fix was written** — the producer writes that. ### COMPLETED **1. R-65 (the hand IK never runs) — answered from the UE 5.8.3 source, and the Session 059 diagnosis was incomplete.** **a. A per-component post-process Anim Blueprint override exists.** `USkeletalMeshComponent` (`Objects/Components/SkeletalMeshComponent.h`): ```cpp /** Post-processing AnimBP to use for the given skeletal mesh component, overriding the one set in the skeletal mesh asset. */ UPROPERTY(transient) TSubclassOf OverridePostProcessAnimBP; // line 410 ENGINE_API TSubclassOf GetPostProcessAnimBPClassToBeUsed() const; // line 417 UPROPERTY(transient) TObjectPtr PostProcessAnimInstance; // line 423 UFUNCTION(BlueprintCallable, Category = "Components|SkeletalMesh") ENGINE_API void SetOverridePostProcessAnimBP(TSubclassOf InPostProcessAnimBlueprint, bool ReinitAnimInstances = true); // line 433 ``` Also available: `ToggleDisablePostProcessBlueprint()` (437) and `GetDisablePostProcessBlueprint()` (441). **b. `PostAnimEvaluation` is not virtual, and — this is the correction — `FinalizeBoneTransform` *is* reached in a normal game frame.** Declaration (SkeletalMeshComponent.h:2271, public): ```cpp ENGINE_API void PostAnimEvaluation(FAnimationEvaluationContext& EvaluationContext); // NOT virtual ``` `USkeletalMeshComponent::PostAnimEvaluation` is `SkeletalMeshComponent.cpp:3181-3381`. After the evaluation itself, in order (`if (bDoEvaluation || bDoInterpolation)` opens at 3276; the whole block closes at 3363): | line | call | |---|---| | 3193 | `EvaluationContext.AnimInstance->PostUpdateAnimation()` | | 3198 | `PostProcessAnimInstance->PostUpdateAnimation()` (when `ShouldPostUpdatePostProcessInstance()`) | | 3211 / 3217 / 3222-3227 | cache copies: `CachedCurve`, `CachedAttributes`, `CachedComponentSpaceTransforms`, `CachedBoneSpaceTransforms` (per `bDuplicateToCache*`) | | 3240 / 3245 / 3250 | `OnUROPreInterpolation()` on the main, linked and post-process instances | | 3262 | `FAnimationRuntime::LerpBoneTransforms(...)` | | 3263 | `GetSkeletalMeshAsset()->FillComponentSpaceTransforms(...)` | | 3266 / 3269 | `AnimCurves.LerpTo(...)`, `UE::Anim::Attributes::InterpolateAttributes(...)` | | 3279 | `ResetMorphTargetCurves()` | | 3291 | `AnimScriptInstance->UpdateCurvesPostEvaluation()` | | 3297 | `LinkedInstance->CopyCurveValues(*AnimScriptInstance)` | | 3302 | `UpdateMorphTargetOverrideCurves()` | | 3310 / 3315 | post-process curve copy / `UpdateCurvesPostEvaluation()` | | 3322 | `DoInstancePostEvaluation()` (when `bDoEvaluation`) | | 3325 | `DoInstanceFinalizeAnimation(EvaluationContext.bDoEvaluation)` | | 3327 | `bNeedToFlipSpaceBaseBuffers = true` | | 3337-3338 / 3343-3344 | `UpdateKinematicBonesToAnim(...)`, `UpdateRBJointMotors()` (bodies or per-poly collision) | | **3354 / 3360** | **`FinalizeAnimationUpdate()`** — editor branch / `!ShouldBlendPhysicsBones()` | | 3366 | `DoInstanceFinalizeAnimation(false)` (the else branch) | | 3377 | `ConditionallyDispatchQueuedAnimEvents()` (the else branch) | | 3380 | `AnimEvaluationContext.Clear()` | `USkeletalMeshComponent::FinalizeAnimationUpdate()` is defined **in `PhysicsEngine/PhysAnim.cpp:468`**, not in `SkeletalMeshComponent.cpp`, and its first act is `FinalizeBoneTransform();` at **PhysAnim.cpp:473**. `USkeletalMeshComponent::FinalizeBoneTransform()` (`SkeletalMeshComponent.cpp:5167`, virtual, overriding `USkinnedMeshComponent::FinalizeBoneTransform` declared at `SkinnedMeshComponent.h:1712`) is where `ConditionallyDispatchQueuedAnimEvents()` runs (5186) and where **`OnBoneTransformsFinalizedMC.Broadcast()` is (5188)** — the only broadcast site in the engine. Every normal-frame route to it, in full: - `PostAnimEvaluation` → `FinalizeAnimationUpdate()` when `!ShouldBlendPhysicsBones()` (3354/3360); - with physics blending, the tick calls `BlendInPhysicsInternal` (`SkeletalMeshComponentPhysics.cpp:3456`, under `if (ShouldBlendPhysicsBones())`) → `FinalizeAnimationUpdate()` at `PhysAnim.cpp:459` (serial), or `FParallelBlendPhysicsCompletionTask::DoTask` (`PhysAnim.cpp:97`) → `CompleteParallelBlendPhysics()` (`PhysAnim.cpp:526`) → `FinalizeAnimationUpdate()` at 530 — the default path, `a.ParallelBlendPhysics 1`; - `RefreshBoneTransforms` (2862) calls `FinalizeBoneTransform()` directly at 3039, but only when `TickFunction == nullptr && ShouldBlendPhysicsBones()` (3036). `ShouldBlendPhysicsBones()` itself is `PhysAnim.cpp:400`: `Bodies.Num() > 0 && CollisionEnabledHasPhysics(...) && (bBlendPhysics || DoAnyPhysicsBodiesHaveWeight())`. The two routes are complementary — when it is false PostAnimEvaluation finalises, when it is true the physics blend does — so the broadcast does happen in a normal game frame, on any frame that evaluated or interpolated. The remaining callers are editor-only: Sequencer, the FBX exporters, the physics-asset editor and `UPoseableMeshComponent`. **Historical Session 065 conclusion (revisited in Session 066): R-65's premise needed further observation before a fix was written.** Session 059 searched `SkeletalMeshComponent.cpp` for callers of `FinalizeBoneTransform` and found only the `TickFunction == nullptr` branch; `FinalizeAnimationUpdate` lives in `PhysAnim.cpp` and was missed. If `USSHandIKMeshComponent::FinalizeBoneTransform()` does run, the failure is elsewhere — the ordering (a hook there runs *before* `UpdateChildTransforms`, `UpdateBounds` and `MarkRenderDynamicDataDirty`, PhysAnim.cpp:483-523), the post-process route, or the component simply not being the one that renders. **c. `AnimGraphService` can build part of this post-process graph, not all of it.** From `Build/vibeue_python_api.json` and `Plugins/VibeUE/Source/VibeUE/Public/PythonAPI/UAnimGraphService.h`: - Creating the Anim Blueprint is not a VibeUE call — use the engine API the `animation-blueprint` skill documents: `unreal.AnimBlueprintFactory` with `target_skeleton` and `parent_class` (any `UAnimInstance` subclass, so a C++ parent works) + `unreal.AssetToolsHelpers.get_asset_tools().create_asset(name, path, unreal.AnimBlueprint, factory)`, then `unreal.EditorAssetLibrary.save_asset(path)`. `SkeletonService.set_post_process_anim_blueprint` and `SkeletonService.save_asset` exist for the skeleton side. - Two Bone IK: **`add_two_bone_ik_node(anim_blueprint_path, graph_name, pos_x, pos_y)`** — that is the whole signature; it creates a `UAnimGraphNode_TwoBoneIK` and sets only its graph position. It cannot set the IK bone, the effector location or its space, the joint target, or alpha. - **There is no Input Pose node anywhere in the plugin** (no `add_input_pose`, and no generic node-add in `AnimGraphService`; `BlueprintService.create_node_by_key` is for ordinary Blueprints). - Pins are connected, not bound: `connect_anim_nodes(abp, graph, source_node_id, source_pin_name="Pose", target_node_id="", target_pin_name="Result")`, `connect_to_output_pose(abp, graph, node_id, "Pose")`, `get_output_pose_node_id(abp, graph)`, `disconnect_anim_node(...)`. For a value, `BlueprintService.set_node_pin_value(blueprint, graph, node_id, pin_name, value)` writes a literal default (`Schema->TrySetDefaultValue`) and `configure_node(blueprint, graph, node_id, property_name, value)` sets a property on the node — neither creates a variable binding, and no method in either service does. - Variables and compile: `BlueprintService.add_member_variable(path, name, type, default, is_array, container_type, instance_editable)` and `BlueprintService.compile_blueprint(path)` still exist in C++, but the skill's guidance is that the engine toolset took them over — `editor_toolset.toolsets.blueprint.BlueprintTools` (`add_variable`, `compile_blueprint`) via `call_tool`. Setting the node's IK parameters and binding its pins to the parent's variables is the part that has to be done in the editor (or by a new plugin method), not by a service call. **2. The full decode, and the grip still fits.** `python Docs/Sourced/ADFRC/rtm_rigs.py` → **13 rigs, 165 clips**, written as `adfrc-rig/2` and `adfrc-anim-local/2` (+ `adfrc-rig-index/1`), exit 0. Then `python Tools/build_adfrc_weapons.py`: 7/7 OK, and every rifle's `grip.fit` is still `true` with a span **identical to the pre-pull run** — A88 0.2418, A88G 0.2929, A4 0.3113, A416 0.3381, A25 0.3486, A89 0.3013 m (all inside the 0.24-0.35 m band). The /2 conversion is therefore correct on the real tree, not just in the unit test; the stop-and-report condition was not met. **3. The reload clip path.** `python Tools/Common/adfrc_grip.py --export-frames Docs/evidence/w5_reload_clips` → 4 clips (GestureReloadAUG 165 frames/66 bones, GestureReloadAUGProne 165/66, MPP_Fast_Reload 54/67, MPP_Slow_Reload 91/67). `python Tools/Common/adfrc_reload.py A88`: ```json {"clip": "GestureReloadAUG", "frames": 165, "max_step_cm": 88.14, "start_cm": [98.89, 50.83, -159.73], "end_cm": [98.9, 50.79, -159.73], "grip_cm": [22.23, -9.4, -1.51], "start_to_grip_cm": 185.84, "switch_phase": 0.48} ``` **4. The A88 magazine probe.** `Tools/Blender/probe_weapon_parts.py` on `ADFRC_EF88_MLOD.blend`: ``` [probe parts] Art/ADFRC_BLEND/adfrc_ef88/ADFRC_EF88_MLOD.blend magazine groups: [{"object": "Memory", "collections": ["point_cloud"], "group": "magazine_axis", "vertices": 2}, {"object": "Memory", "collections": ["point_cloud"], "group": "mag_latch_axis", "vertices": 3}] ``` **5. Muzzle-flash candidates.** A commandlet pass over the asset registry (`-ExecutePythonScript`, the convention `setup_weapons.py` uses): `/Game/Realistic_Starter_VFX_Pack_Vol2` holds 188 assets — 56 `ParticleSystem`, 0 `NiagaraSystem` — and **not one** has Muzzle, Flash or Shot in its name; its particles are `P_Asphalt`, `P_Blood_Splat_Cone`, `P_Destruction_*`, `P_Explosion_*` and similar. The same keyword pass over all of `/Game` (14,558 assets) found 4, recorded in `Docs/evidence/vfx_muzzle_candidates.json`: | class | path | |---|---| | NiagaraSystem | `/Game/Effects/Particles/Weapons/NS_WeaponFire_MuzzleFlash_Rifle` | | NiagaraSystem | `/Game/Effects/Particles/Weapons/NS_WeaponFire_Tracer_Shotgun` | | ParticleSystem | `/Game/AK-47/FX/MuzzleFlash/P_AssaultRifle_MuzzleFlash` | | ParticleSystem | `/Game/Downloaded/VaultCache/Untitled7d3b12f5addbV1/data/Content/AK-47/FX/MuzzleFlash/P_AssaultRifle_MuzzleFlash` | ### FILES CHANGED - `Docs/evidence/w5_reload_clips/` — new: `GestureReloadAUG.json`, `GestureReloadAUGProne.json`, `MPP_Fast_Reload.json`, `MPP_Slow_Reload.json`, `parts_A88.json`. - `Docs/evidence/vfx_muzzle_candidates.json` — new (with the pack inventory and the /Game wide scan). - `Art/Weapons/*/ADFRC/` — 7 `SM_*.fbx` (LFS) + 7 `manifest.json` touched by the rebuild. - `Docs/CHANGELOG.md`. - Not committed, by instruction: `Build/probe_muzzle_vfx.py` (the throwaway probe that produced the VFX list) and the stale `/1` decode tree, moved to `Build/ADFRC_Rig_stale_v1/`. ### TESTING - `python Docs/Sourced/ADFRC/rtm_rigs.py` — exit 0; 13 rigs / 165 clips, schemas `adfrc-rig/2`, `adfrc-anim-local/2`, `adfrc-rig-index/1`. - `python Tools/build_adfrc_weapons.py` — exit 0, 7/7 OK; all six grip spans identical to the previous run and all `fit: true`; A9 unchanged (no handAnim clip, so no grip). - `python Tools/Common/adfrc_grip.py --export-frames ...` — exit 0, "4 clip(s) written". - `python Tools/Common/adfrc_reload.py A88` — exit 0, JSON above. - Blender probe — exit 0, magazine groups line above. - Editor commandlet probe — exit 0, `[VFX]` line written; the report is the JSON. ### ASSETS None imported, created or modified by hand. The 7 weapon FBX were re-exported by `Tools/build_adfrc_weapons.py` and re-committed. ### RISKS No new risk numbers were taken: this session did not reserve a range, and the ID rule says not to guess the next free one. The open items this session raises, for numbering with the others: the R-65 mechanism (below) is now unproven rather than confirmed; the A88 magazine has no detachable named selection (below); the weapon FBX export is not byte-reproducible (below); and the decoder cannot be re-run over its own output (below). ### DEFECTS FOUND - **R-65's mechanism does not hold as written.** `FinalizeBoneTransform` is reached in a normal frame via `FinalizeAnimationUpdate` (`PhysAnim.cpp:468`, called from 3354/3360, 459 and 530). Any fix should start from why the override's work is not visible in the rendered pose, not from "it never runs". - **`rtm_rigs.py` cannot be run twice.** Its input glob skips a file whose *immediate* parent directory is `Rig` (`if os.path.basename(os.path.dirname(f)) != "Rig"`), but it writes its own output to `Rig//.json`, where the parent is ``. The second run therefore reads its own output as input and dies with `KeyError: 'bones'` (165 files) — this is how this session started. The stale `/1` tree was moved aside to `Build/ADFRC_Rig_stale_v1/` and the tree was decoded fresh. One line fixes it: exclude anything under `Rig/`, not just files directly in it. - **The FBX export is not reproducible.** Each `SM_*.fbx` is the same size and differs from the committed one only in the `CreationTimeStamp` bytes at offset ~292 of the header (`cmp` on the A88: first difference at byte 292, inside Year/Month/Day/Hour/Minute/Second). Every rebuild therefore mints 7 new LFS objects (~23 MB) that carry no mesh change, and the manifests show as modified purely through CRLF. - **The A88's magazine is not a detachable part.** The probe reports 67 mesh objects in the MLOD blend and **none** of them carries a vertex group: the named selections Arma uses did not survive the conversion, and the only magazine references left are two memory points on the `Memory` point cloud (`magazine_axis`, 2 vertices; `mag_latch_axis`, 3). Arma's reload works by hiding the `magazine` named selection, so as it stands there is no geometry to hide or swap — the magazine is welded into the gun mesh. That is a W5 finding, not a tooling one. - **`GestureReloadAUG` starts 1.86 m from the grip, with an 0.88 m single-frame step.** The clip is a full-body gesture whose first frame is not a weapon-ready pose; worth knowing before the path is turned into an animation. ### NEXT ACTION Write the R-65 fix from the annotated call chain above, starting with why `USSHandIKMeshComponent::FinalizeBoneTransform` (or the post-process route) does not change the rendered pose — then fix the `rtm_rigs.py` exclusion so the decoder can be re-run, and stop regenerating the FBX until the export timestamp is pinned. --- ## Session 066 — 2026-09-28 — review of Session 065: R-65 reopened as unobserved, the gesture clips don't pose, decoder re-runs ### COMPLETED - **R-65 is not "the hook never runs".** Session 065 traced `FinalizeAnimationUpdate` (PhysAnim.cpp:468) calling `FinalizeBoneTransform` on every evaluated or interpolated game frame. `USSHandIKMeshComponent` edits the editable buffer before `Super` flips it, so the edit should be published. No one has looked at a rendered frame with `ss.HandIK 1` against `ss.HandIK 0`; R-65 is now "unobserved", and that A/B screenshot is the next action. - **`GestureReloadAUG`/`…Prone` do not pose under the grip model**: the wrists come out 1.5-3.2 m apart. The start offset and 0.88 m step in Session 065 were this, not a gesture start pose. `adfrc_reload.py` now refuses any frame with wrists over 0.9 m apart (new test). `MPP_Slow_Reload` passes (steps ≤ 13 cm) but is not the AUG gesture. W5 plan: the left-hand path is authored from the weapon's own points (grip → `magazine_axis` → pouch → back), not decoded. - **The decoder can be re-run.** `rtm_rigs.py` skips everything under `Rig/`, not just files directly in it. ### FILES CHANGED `Docs/Sourced/ADFRC/rtm_rigs.py`, `Tools/Common/adfrc_reload.py`, `Tools/Common/test_adfrc_reload.py`, `Docs/CHANGELOG.md`. ### TESTING - `python Tools/Common/test_adfrc_reload.py` → exit 0, 11/11 checks passed. - `python Tools/Common/test_rtm_rigs.py` → exit 0, 0 failures. - `python Tools/Common/adfrc_reload.py A88` → exits with an error: GestureReloadAUG frame 0 has the wrists 1.95 m apart. - NOT RUN: running the decoder twice on the real tree (the raw pack is on the producer's machine only), the build, the automation tests, the in-game check. ### ASSETS None. ### RISKS - R-65 restated: the hand IK hook runs (Session 065) but its effect on the rendered pose has never been observed. - R-78: the ADFRC gesture clips decode to impossible poses, so the reload can't come from Arma's animation. ### DEFECTS FOUND - The gesture clips give impossible poses (found by measuring the wrist-to-wrist distance per frame). - The decoder crashed when re-run (found by the Session 065 agent; fixed here). ### NEXT ACTION A rendered A/B check of the hand IK: `-game` A88 runs with `-SSShotAt` and `-SSExec=ss.HandIK 0` versus `ss.HandIK 1`, and pixel-diff the left hand. --- ## Session 067 — 2026-09-29 — R-65 observed: the hand pose does change; and how a single pair nearly read as a lie **Later status note (2026-10-01):** this session's controlled A/B is the historical rendered evidence for hook execution/pose change; Session 070 added the measured A88 grip-target distance. Current R-65 is the broader fit/acceptance gap, not that the hook never executes. See the current risk row in `PROJECT_AUDIT.md`. The three handover tasks. The R-65 A/B needed more runs than it was asked for: one cross-run pair is not evidence, because a `-SSShotAt` capture has no deterministic viewpoint. ### COMPLETED **1. R-65 is now observed — `ss.HandIK 0` and `ss.HandIK 1` do render a different left-hand pose.** Four runs on `/Game/Maps/L_DryRiver_01`, A88 Rifleman kit, `-game -windowed -ResX=1600 -ResY=900 -nosplash -nosound -FORCELOGFLUSH -SSNoClassSelect -SSShotAt=20 -SSExecAt=12 "-SSExec=ss.HandIK 0|1"`. The flag reaches the game once the pawn exists, and the echo confirms the value: ``` [2026.09.28-22.23.48:402][954]LogSSObjectives: SSExec at 12.0 s: ss.HandIK 0 [2026.09.28-22.23.48:412][954]ss.HandIK = "0" [2026.09.28-22.23.56:403][764]LogSSObjectives: Requested viewport screenshot at 20.0 s. [2026.09.28-22.26.29:423][145]LogSSObjectives: SSExec at 12.0 s: ss.HandIK 1 [2026.09.28-22.26.29:428][145]ss.HandIK = "1" [2026.09.28-22.26.37:425][ 99]LogSSObjectives: Requested viewport screenshot at 20.0 s. ``` The components the fallback question asked for are already named by the binding log, so `ShowDebug ANIMATION` was not needed. Body = `CharacterMesh0` on `SKM_Manny_Invis` (hidden in first person); the first-person arms = `SS_FirstPersonArms`, re-pointed per weapon: ``` LogSSHandIK: CharacterMesh0 (SKM_Manny_Invis): left-hand IK on upperarm_l > lowerarm_l > hand_l. LogSSHandIK: SS_FirstPersonArms (SK_FP_Arms_Rifle): left-hand IK on LeftArm > LeftForeArm > LeftHand. LogSSHandIK: SS_FirstPersonArms (SK_FP_Arms_Pistol): left-hand IK on LeftArm > LeftForeArm > LeftHand. ``` All four captured runs held the `SM_A88` rifle view model at the shot, so the weapon is not a confound (`LogSSFirstPerson: View model shows SM_A88` is the last view-model line before `Requested viewport screenshot` in every one of them; the `SM_A9` pistol line comes after). | run | `ss.HandIK` | image | log | |---|---|---|---| | off1 | 0 | `Docs/evidence/handik_ab/off.png` | `Saved/Logs/SS_handik_off.log` | | off2 | 0 | `Docs/evidence/handik_ab/off2.png` | `Saved/Logs/SS_handik_off2.log` | | on2 | 1 | `Docs/evidence/handik_ab/on.png` | `Saved/Logs/SS_handik_on2.log` | | on3 | 1 | `Docs/evidence/handik_ab/on3.png` | `Saved/Logs/SS_handik_on3.log` | Mean |RGB delta| / coarse-structure correlation (40x22 luma), so the two same-condition pairs are the noise floor and any real effect has to beat them in *every* cross pair: ``` off1 off2 on2 on3 off1 -- 31.52/0.81 33.22/0.73 31.43/0.78 off2 31.52/0.81 -- 32.32/0.76 30.16/0.79 on2 33.22/0.73 32.32/0.76 -- 26.48/0.82 on3 31.43/0.78 30.16/0.79 26.48/0.82 -- ``` The whole-frame numbers overlap, so they settle nothing on their own. The localised statistic does: per tile of a 16x9 grid, flag the tiles where the **minimum** over the four cross-condition pairs exceeds the **maximum** over the two within-condition pairs by more than 2x. 7 of 144 tiles qualify, and they are all in one place — **x 31-56%, y 66-100% of the frame**, ratios 2.7x to 7.1x. Every other tile is at the noise floor (ratio <= 1.2). That cluster is the view-model region, bottom centre, which is where the gripping left hand sits. Pixel counts over the arms/weapon crop (x 448-960, y 558-900, 512x342 px), percent of pixels moving >30: within-condition **13.3%** (off1/off2) and **6.6%** (on2/on3); cross-condition **42.6%, 46.8%, 43.1%, 42.5%**. All four cross pairs agree; neither within pair does. **What this does not show:** that the hand *grips the foregrip*. Pixels show the view-model pose changed when the IK is on, in the hand's region; they do not name the bone or prove the target. An `-SSAnimDebug` run or a left-hand transform log is what would name it. The R-65 engine question (which post-process AnimBP route the edit needs) is untouched — the producer writes that fix. **2. The decoder re-runs (this was the asked-for pair, and the earlier attempt did not actually run).** `python Docs/Sourced/ADFRC/rtm_rigs.py` twice back to back, on a tree that already held `Rig/` output: run 1 exit 0 in 15 s, run 2 exit 0 in 16 s, zero `KeyError`/`Traceback` in either log (grep counts 0 and 0), both ending `-> E:\SouthernSpear\Content\Sourced\ADF_Extracted\Animations\Rig`. The Session 066 fix (skip everything under `Rig/`) holds for a real re-run. ``` python Tools/Common/test_adfrc_reload.py -> exit 0, 0 failure(s) python Tools/Common/test_rtm_rigs.py -> exit 0, 0 failure(s) python Tools/Common/test_adfrc_grip.py -> exit 0, 0 failure(s) ``` **3. No rebuilt FBX committed.** A decoder rebuild only moves `CreationTimeStamp` in the FBX header, so the regenerated meshes are byte-different and content-identical; committing them would be pure churn (and 7 new LFS objects, ~21 MB, per rebuild). The exporter bug is left unfixed on the producer's instruction. **4. Committed** `Docs/evidence/handik_ab/` and this entry only, with `git commit --only -- `. ### FILES CHANGED - Created: `Docs/evidence/handik_ab/{off.png, off2.png, on.png, on3.png, on_attempt1_hitched.png, diff_map.png, hand_region_crop.png, handik_ab.json}`. - Modified: `Docs/CHANGELOG.md`. - Scratch, git-ignored, not committed: `Build/make_handik_evidence.py`, `Build/probe_handik_ab.py`, `Build/probe_handik_ab2.py`, `Build/handik_null/`. ### TESTING - Decoder run twice in a row: exit 0 / exit 0, no `KeyError` (see above). - `test_adfrc_reload.py`, `test_rtm_rigs.py`, `test_adfrc_grip.py`: 0 failures each. - Four `-game` captures (2 per condition) plus one control pair; all five reached `up for play` and wrote the shot at 20.0 s (exit 124 = the expected timeout kill after the shot). - `git status` after the decoder double-run: no tracked file under `Content/Sourced/` modified. - NOT RUN: the UE build and the Automation suites — no C++ or asset changed this session. ### ASSETS None committed. The evidence PNGs are captures, not content. Rebuilt FBX deliberately not committed. ### RISKS - **R-65 moves from "unobserved" to observed.** The hook fires *and* the rendered pose changes. What is still open is whether the change is the right one (hand on the foregrip) — and the engine-side question in Session 065 is unchanged. - **R-79 (new, this session): a `-SSShotAt` capture has no deterministic viewpoint, so a single cross-run pair cannot support an A/B.** The first ON run was requested at 22:26:37.425, the same second Niagara compiled `NS_WeaponFire_MuzzleFlash_Rifle`/`ShellEject`/`Tracer`; that frame correlates 0.35 with the others, against 0.81 for a same-condition control. Read naively it says "enabling the hand IK changes the whole scene". Kept as `on_attempt1_hitched.png`. Only the two-per-condition design above made the answer safe. (The risk register in `PROJECT_AUDIT.md` still lists only up to R-66, as with R-70-R-78.) ### DEFECTS FOUND - The capture pipeline, not the hand IK, was the thing most likely to produce a false answer — found by running a control pair instead of trusting the first one. - The world half of the frame is unusable at pixel level for A/B: 82.4% of its pixels move >30 between two *same-condition* runs (grass/foliage/TAA), against 77.0% across conditions. Region statistics only; eyeballing a pair proves nothing here. - The decoder's re-run crash is confirmed gone on the real tree, not just in the unit fixtures. ### NEXT ACTION Pin the capture viewpoint — a capture flag that places the pawn and holds pitch/yaw for the shot — because this session spent two extra runs discovering that every future A/B on this project is only as trustworthy as that viewpoint. --- ## Session 068 — 2026-09-29 — The training environment that arrived on its own: MOUT, noted before it is used A producer note, not a build session. A Fab environment pack was downloaded and installed into `Content/` during the Ravenshoe work, nobody wrote it down, and it is the only new environment map in the project. This session finds it, measures what is actually on disk, and registers it as a **candidate** — deliberately without opening it in the editor, so nothing here is a claim about how it looks. ### COMPLETED **The download is `Content/MOUT_Civilian/` — a MOUT ("Military Operation Urban Training") urban kit.** Staged at `Content/Downloaded/VaultCache/ModularM6dfea54fd98cV5/`, written 09:13–09:16 today, installed to `Content/MOUT_Civilian/`. **2.1 GB, 507 files** (504 `.uasset`, 3 `.umap`). Three maps, of which the only environment map is `Demo/FirstPersonBP/Maps/FirstPersonExampleMap` (8.5 MB + 14 MB `_BuiltData`); the other two are `LVL_AssetShowcase` and `LVL_Blueprints`. Contents: three modular mesh sets (Building, Church, Awning), thirteen prop sets (bus stop, bollards, park bench, playground, fire hydrant, clothes line, fencing, flag, fountain, police sign, post box, trash cans, electricity pole), nine building Blueprints, **five interactable door Blueprints**, 219 textures, and a first-person demo character and weapon. **Two findings that were not obvious from the filenames.** 1. **The kit was authored in UE 4.26, not 5.x.** `++UE4+Release-4.26` sits in the demo map's header. Everything in it upconverts on load, and none of it has been opened in 5.8. Raised as **R-67**. 2. **The pack wrote no `metadata` sidecar anywhere** — `find -iname metadata` over all 2.1 GB returns nothing — so the seller and `isAiForbidden` are **unverified**, not known-clear. Recorded per the L-0016c standing action. Raised as **R-68**. **It does not fill `M-006` "Training range" as that entry is written.** The register and VS-17 both say *range*, which is open ground with long lanes, and Saltbush already covers that (built "to test engagement ranges past 200 m"). This is a close-quarters **village**: doorways, rooms, awning-to-ground transitions, lookalike buildings. Scored against GDD §5 that makes it a strong fit for **Induction** (module 1) and **Field Skills** (module 4, where explicit identification training wants exactly this kind of geometry) and a weak fit for **Leadership** (module 5). Marksman lanes still have to come from elsewhere. Whether `M-006` is redefined or this becomes a second training map beside it is left open — that is a producer decision, and `MAPS_TRAININGRANGE.md` §6 lists it as the first of five. **Four register documents touched, in the same change as the install**, per the L-0016b standing rule that a pack's row is added *with* the import rather than after: - `MAPS_TRAININGRANGE.md` (new) — what arrived, what is in it, the two candidate uses with a per-module fit table, provenance, what is not decided, the two risks, and the commands to re-verify every number - `ASSET_REGISTER.md` §4.9j (new) — the MOUT kit plus the four other listings downloaded today - `LICENCE_REGISTER.md` L-0016 table — the same five listings - `PROJECT_AUDIT.md` §7 — **R-67**, **R-68** **Two corrections to existing rows, both found by re-reading them against the disk.** §4.9i and the L-0016 row for *Old Abandoned Rusty Cars* both say "downloaded, not imported". It was installed to `Content/RustyCarsFree/` today (69 MB, and it ships its own `Overview/AssetsOverview.umap`). Both rows now say installed and still used by no map — the single Renault wreck on Ravenshoe remains the project's car. ### FILES CHANGED Created: `Docs/MAPS_TRAININGRANGE.md` Modified: `Docs/ASSET_REGISTER.md` (§4.9j new, `M-006` row, header date), `Docs/LICENCE_REGISTER.md` (L-0016 table, rusty-cars row, header date), `Docs/PROJECT_AUDIT.md` (§7 R-67/R-68, risks note), `Docs/CHANGELOG.md` ### TESTING | Check | Command | Result | |---|---|---| | Pack identity and version | `strings -n 4 .../FirstPersonExampleMap.umap \| head` | **PASS** — `++UE4+Release-4.26`; the demo map's engine version read from the file, not inferred from the folder name | | Size and file counts | `du -sh`, `find … \| wc -l`, extension histogram | **PASS** — 2.1 GB, 507 files, 504 `.uasset` / 3 `.umap` | | Map inventory | `find Content/MOUT_Civilian -iname "*.umap" -printf "%s %p\n"` | **PASS** — 3 maps, sizes recorded | | Seller / AI flag | `find …/ModularM6dfea54fd98cV5 -iname metadata` | **PASS** — empty, sidecar absent; recorded as unverified. (The sibling `FabLibrary/listings_v1.db` was queried too and holds only the five *FabLibrary* listings, not this pack) | | Untracked, as intended | `git status --porcelain` | **PASS** — `Content/MOUT_Civilian/` untracked, raw pack stays git-ignored (ADR-021) | | Editor opened on any of the three maps | — | **NOT RUN** — deliberately. Nothing is claimed about how the kit looks or performs | | Navmesh, lighting, objectives, ADR-016 look check | — | **NOT RUN** — the kit has no layout and no name, so there is nothing to check it against yet | | `validate_architecture.py`, editor build, automation suite | — | **NOT RUN** — no source or content change. Documentation only | ### ASSETS - No asset imported, modified, moved or deleted. The pack was already in `Content/` before this session and is recorded, not touched. - Five listings registered that arrived today: the MOUT kit (installed, used by nothing), plus *Mega Moduler Apartment Building*, *Modular 3D hospital environment*, *American Road with Parking Lot* and *IFAK* (cache only). Two carry `isAiForbidden: true`; neither is installed in a map. - No new `M-` id opened, and no `L_*` map asset created. The kit is registered as a candidate against `M-006`. ### RISKS - **R-67 (new):** UE 4.26 kit, never opened in 5.8, 2.1 GB, 219 legacy textures, unknown LOD/nav/texel cost. - **R-68 (new):** seller and `isAiForbidden` unverified for the whole pack (no `metadata` sidecar). - Neither blocks reading the kit; neither should hold up the decision about whether to design on it. No map depends on it, so an upconversion failure costs an idea and nothing else. ### DEFECTS FOUND - **A 2.1 GB environment kit sat in `Content/` with no row in any register** — the exact gap L-0016b was opened to close, recurring two days later on a different pack. It survived because nothing referenced the pack, so no failed build and no audit ever looked at it. The standing rule already covers this (row added in the same change as the import); what is missing is a check that catches an *unreferenced* installed pack, which is the failure mode L-0016b also could not catch. - **Two register rows were stale**: "downloaded, not imported" for the rusty-cars pack, which was installed today. Corrected in both documents. ### NEXT ACTION Open `MOUT_Civilian/Demo/FirstPersonBP/Maps/FirstPersonExampleMap` once, headless, and answer the three questions R-67 exists to ask — does the 4.26 content survive upconversion, what does the geometry measure, and what does it cost — so the producer can decide between "training range" and "urban training facility" on evidence instead of on a folder name. ## Session 069 — 2026-09-29 — Ravenshoe gets the Session 041 world treatment; and the nav-bake doctrine, corrected by experiment The producer's ask: bring Ravenshoe up to how Dry River and Red Gum were actually made. This session ported the passes the map never inherited, then — chasing the frozen bots — tore apart the project's nav-bake doctrine and rebuilt it from measurements. The ground, the horizon and the audit are fixed. The bots are still frozen, and the session ends with the one action that fixes them, now known to be achievable. ### COMPLETED - **The Dry River Session 041 world treatment, ported as `Tools/Unreal/expand_ravenshoe.py`** (idempotent, report `Build/expand_ravenshoe_report.json`): outer skirt `SS_MAP_Ravenshoe_Skirt` (new `Tools/Blender/ravenshoe_skirt.py` + `Tools/Common/ravenshoe_world.py`; 55,862 verts, z −25.9..+137.8 m, **0.0 m edge-height error** against the terrain's 2 m grid — seamless by construction); the pack's `MI_Ground_Dirt_01` on terrain and skirt (the terrain override had silently fallen back to the flat `MI_SS_Raven_Road` — the white ground); four blocking volumes at terrain + 30 m; `SM_Horizon_01` ring at ~2.6 km plus a VolumetricCloud (the black horizon band); nav bounds sized to the play space with read-back correction (X ±130 m, Y ±180 m, Z −25..+40 m — asserts coverage of every deployment, both ramp feet, bed and crest). - **The RecastNavMesh tile pool: 1024 → 4096**, persisted on the actor. The play space needs ~1,285 tiles at TileSizeUU 1000. The "2448" figure in the Session 045 fix history was the *required tile count of the old oversized volume*, never the pool — the real pool was the 1024 default, which is what capped the one interactive bake the map ever got. - **`build_ravenshoe_nav.py` repaired:** it was re-imposing 115/165/45 bounds on every build run — silently reverting the §0 bounds fix; it now asserts bounds coverage and pool instead; single BUILDPATHS (a second call rebuilds on live tiles); save via `LevelEditorSubsystem.save_current_level` like `build_dryriver_nav.py` (`EditorLoadingAndSavingUtils.save_map` serialises no fresh tiles here). - **Audit repaired, 35/35:** the stale "two deployments placed" check counted the 16 Lyra starts as a failure — a regression the starts fix introduced and nobody re-ran the audit on. - **The nav forensics** (logs `Saved/Logs/SS_probe_rav*.log`, `SS_probe_dr*.log`): a fresh cube gains no nav poly after ten spaced BUILDPATHS on **either** map — headless generation consumes no new geometry on this machine, so Dry River's "headless bake works" rode its persisted interactive tiles. The editor with a real RHI builds (1.2 s passes) but consumes almost none of this map's geometry; `bForceRebuildOnLoad=True` set and persisted, does not fire in game worlds. The pipeline's ini flag (`bWaitForAsyncLoadingBeforeBuildingNavigationAutomatically=False`) makes the automatic load-time build run **early, on partial geometry** — part of the fault, not the fix. - **`PLAYTEST_COMMANDS.md` compliance restored** after the producer's correction: own `-abslog` on every engine invocation (the shared `SouthernSpear.log` had another agent's crash in it and produced a false conclusion), scratch probes moved out of `Tools/`, screenshot re-taken properly next round. - **Both new Fab packs git-ignored** (`MOUT_Civilian`, `RustyCarsFree`) per ADR-021, complementing the registration work in Session 068. ### FILES CHANGED Created: `Tools/Common/ravenshoe_world.py`, `Tools/Blender/ravenshoe_skirt.py`, `Tools/Unreal/expand_ravenshoe.py`, `Content/Art/Blockout/SS_MAP_Ravenshoe_Skirt.fbx` + `.uasset`, `Docs/evidence/ravenshoe/SSShot_ground_skirt_0854.png` Modified: `Tools/Unreal/build_ravenshoe_nav.py`, `Tools/Unreal/audit_ravenshoe.py`, `Content/Maps/L_Ravenshoe_01.umap` (pool + flag + persisted fixes), `.gitignore`, `Docs/HANDOVER_RAVENSHOE.md` (§0b), `Docs/CHANGELOG.md`. Scratch (untracked, `Build/`): `probe_raven_*.py`, `probe_dr_*.py`, `fix_raven_rebuildonload.py` and their reports. ### TESTING | Check | Command | Result | |---|---|---| | Skirt geometry | `blender -b --factory-startup -P Tools/Blender/ravenshoe_skirt.py` | **PASS** — `SS_SKIRT` 55,862 verts / 55,080 faces, `edge_height_error_m: 0.0` | | Spec verifier | `python Tools/Blender/verify_ravenshoe.py` | **PASS** — spec-only, 2/2 | | Expand pass | expand_ravenshoe.py | **PASS** — all steps ok, nav bounds ~1,285 tiles, pool 1024→4096 | | Map audit | audit_ravenshoe.py | **PASS 35/35** (was 32/33) | | Nav build (build mode) | build_ravenshoe_nav.py + env + ini flag | **RUN, exit 0, no crash**; report honest: 0/32 routes, ~12 persisted tiles | | Cube test, both maps | Build/probe_dr_cube.py, Build/probe_raven_cube.py | **FINDING** — 10 spaced builds, zero polys on a fresh cube | | Live game, bots | `-game` 6 bots, 210 s, own log | **RUN** — 0 spawn fails, rounds cycle, `Steered 0 idle bot(s)`, captures 0 | | Interactive editor bake (Build ▸ Build Paths, attended) | — | **NOT RUN** — needs a human at the editor; this is the one remaining action | | Look check in a real window | `UnrealEditor.exe ... -SSShotAt=40` (§3), `Docs/evidence/ravenshoe/SSShot_lookcheck_1011.png` | **PASS (measured)** — sky quarter RGB (135,151,163) vs ground quarter (178,128,91): the ground renders as red dirt, not white; 2163 distinct colours in frame. First capture attempt used `-Cmd -windowed` (§3 violation), re-taken with the full editor | | MOUT / RustyCars look checks (ADR-016) | — | **NOT RUN** — Session 068 scope | ### ASSETS - `SS_MAP_Ravenshoe_Skirt` — class F, original, generated from `ravenshoe_world.height()` (which is `ravenshoe_spec.ground_z()` inside the play area and authored hills/rim outside it). - Referenced in place, never modified: `MI_Ground_Dirt_01` + `SM_Horizon_01` (`RuralAustralia`, L-0016), engine VolumetricCloud. New packs on disk registered in Session 068, not used by any map yet. ### RISKS - **R-82 (new; first written as "R-69", renumbered by Session 070's finding):** no scripted process on this machine — commandlet, `-RenderOffscreen`, or the interactive editor unattended — consumes geometry in the nav generator. Every map bake requires the attended editor. CI can build and audit maps but can never verify nav; the Dry River gate's nav evidence is the exception that proves this rule. - **R-83 (new; first written as "R-70", renumbered likewise):** `PLAYTEST_COMMANDS.md` §5 and CLAUDE.md present the async-loading ini flag as part of the working nav recipe. Measured this session: it makes the automatic load-time build run early, on partial geometry. The docs were corrected on 2026-09-29 in the same change; the flag's remaining role (inert in headless CI, harmful as a documented recipe) is stated in `MAPS_DRYRIVER.md` §11.6. ### DEFECTS FOUND - The audit had silently regressed to 32/33: the starts fix changed actor counts and the audit was never re-run — the project's own rule ("re-run the verifier after every change") broken by the fix that the verifier was checking. - `build_ravenshoe_nav.py` re-imposed stale bounds on every run, reverting the §0 fix. One owner per piece of state, or two passes fight. - §0 #1's "navigation can be baked headless" was a misattribution: the underlying bake was Dry River's persisted tiles; Ravenshoe's "success" was the early-build flag answering queries from a partial in-memory mesh. The verify script's huge query extents (up to 50 m) made partial nav read as success. - Process defects, mine: read the shared `SouthernSpear.log` without `-abslog` (another agent's crash nearly became my conclusion); used `UnrealEditor-Cmd -windowed` for a capture (§3 says full editor); put probes in `Tools/`. All corrected after the producer pointed at `PLAYTEST_COMMANDS.md`. ### NEXT ACTION **In the attended editor:** open `/Game/Maps/L_Ravenshoe_01`, run **Build ▸ Build Paths** (the pool is 4096 and the bounds now cover the play space, so yesterday's cap and yesterday's bounds are both gone), save, then confirm `Build/ravenshoe_nav_report.json` reads 32/32 routes and a live `-game` run logs `Steered N idle bot(s)` with N > 0. ## Session 070 — 2026-09-29 — The IK puts the hand on the weapon; the hand then turns to the grip R-65 asked whether the hand IK does what it claims. Measured first, as asked, and the measurement is clean: on the A88 the first-person left hand lands **exactly** on `LeftHandGrip`, with five centimetres of reach to spare. The same runs show why that is not yet "holding the weapon": the solve moves the *wrist*, the hand keeps whatever rotation the clip gave it, and the Fab arms' clip leaves the palm open. So the second half of this session adds a hand rotation behind the position solve, with its correction solved from the third-person body's own Lyra grip and kept in config. ### COMPLETED **1. Where the IK actually puts the hand** (probe `-SSHandIKProbe`, one sample every 0.5 s for 5 s; arms = `SS_FirstPersonArms` / `SK_FP_Arms_Rifle`, weapon = `SS_ViewModel` / `SM_A88`; positions and distances in cm, component space; `gate` = arm length x `MaxReachFactor`): | t (s) | IK alpha | hand -> grip | shoulder -> grip | arm | gate | reach-limited | |---|---|---|---|---|---|---| | 0.40 | 0.00 | **8.37** | 47.98 | 51.80 | 54.39 | no | | 1.01 | 1.00 | **0.00** | 48.66 | 51.80 | 54.39 | no | | 1.52, 2.01, 2.50, 3.00, 3.51, 4.01, 4.51, 5.01 | 1.00 | **0.00** every sample | 48.66 | 51.80 | 54.39 | no | The wrist-only solve moves the hand **8.37 cm** onto the socket and holds it there exactly; the reach gate is never in play (48.66 wanted against 54.39, i.e. 89 % of a 51.80 cm arm at `MaxReachFactor` 1.05). The body's own `CharacterMesh0` (`SKM_Manny_Invis`) reaches the same 0.00 cm and is the mesh that *does* hit the gate: at alpha 0 it stands 42.39-42.94 cm off and logs `reach_limited=1 straight=1` for the first two samples. So the answer to the earlier "stop and report the weapon-frame offset" branch is **no**: the distance is under 3 cm, the position is right, and nothing in `Tools/Blender/adfrc_weapon.py`'s socket axes needs touching. **2. Every mesh component on the pawn in first person, with its mesh and material 0** (same probe, one-shot walk over 3 actors and 2 hand-IK components): | component | class | mesh | material 0 | visible | owner-no-see | only-owner-see | |---|---|---|---|---|---|---| | `CharacterMesh0` | `SSHandIKMeshComponent` | `SKM_Manny_Invis` | **none** | yes | yes | no | | `SS_FirstPersonArms` | `SSHandIKMeshComponent` | `SK_FP_Arms_Rifle` | `MID_MI_FP_Arms_Rifle` | yes | no | yes | | `SS_ViewModel` | `StaticMeshComponent` | `SM_A88` | `MID_MI_A88_adfrc_ef88_co` | yes | no | yes | | `B_SS_A88_Weapon_C_0.SSVisual` | `StaticMeshComponent` | `SM_A88` | `MID_MI_A88_adfrc_ef88_co` | yes | yes | no | | `B_SS_A88_Weapon_C_0.SkeletalMesh` | `SkeletalMeshComponent` | `SK_Rifle` | `MID_MI_Weapon_Rifle` | **no** | yes | no | | `B_SS_Soldier_C_0` x8 | `SkeletalMeshComponent` | 8 ADF/insurgent part meshes | `M_Eye_Source`, `MI_ADF_*`, `MI_MAF_*` | **no** | yes | no | (`CameraProxyMeshComponent_0` is the editor's `MatineeCam_SM`, visible=no.) **The walking tan block is not in this list** - every component that reaches the owner's camera has a real material, and the only untextured one (`CharacterMesh0`, no material 0 at all) is owner-no-see and hidden in first person. The two components that render to the owner are the arms and the view model, so the block is either part of `SM_A88`'s own texture/material state or a component the walk does not reach. The probe carries an ablation switch (`ss.Probe.Hide`) for exactly this and it was **not taken**, so the block stays **unidentified**. **3. The A89's reserve ammo: 200, and it comes from the loadout table.** `Config/DefaultGame.ini`, `[/Script/SouthernSpearLyraBridge.SSLoadoutSettings]` (line 308): `+Weapons=(Weapon=A89,RoundsPerMinute=750,MagazineSize=200,SpareMagazines=1,SpreadScale=1.74,bFullAuto=True)` - a 200-round magazine plus one spare magazine, so 200 in reserve and 400 on the pawn. The producer's A89 shot reads 193/200. The row is applied through `SSWeaponStatsSubsystem` from `SSWeaponStats.h`; nothing else in config sets an A89 magazine count. **4. The hand rotation step (`FSSHandIK::RotateChain`).** After `Apply` has put the wrist on the socket, the hand bone's component-space rotation is set to **socket rotation x `HandRotationOffset`**, slerped by `Alpha`, and every descendant of the hand is rebuilt from its own local transform so fingers and wrist-twist bones follow. The offset is per skeleton and lives in config, not code: `[/Script/SouthernSpearLyraBridge.SSHandIKMeshComponent] HandRotationOffset=(Pitch=25.810,Yaw=146.002,Roll=-40.132)` (quat `[0.298391, 0.258492, 0.897973, 0.194394]`, 157.6 deg). It is not picked by eye: the body holds the same `SM_A88` with Lyra's own animation, so its hand's orientation in the weapon's frame is a known-good grip, and the offset that hands the arms that same grip is `offset = socketrot_arms^-1 * (weapon_in_cs_arms * hand_in_weapon_body)`, with the body's `hand_in_weapon = (-0.8980,-0.1944,0.2984,0.2585)`, the arms' `socketrot = (-0.7071,0.7071,0,0)` and `weapon_in_cs = (0,0,-0.7071,-0.7071)`. Predicted result 0.00 deg from the body's, and the live run after the config write reads the arms' `hand_in_weapon = (-0.8980,-0.1944,0.2984,0.2585)` with `hand_vs_socket = 157.6 deg` (it was 0.0 deg, i.e. the hand was simply left in the socket's orientation) while the body is untouched. `bRotateHandToGrip` defaults off, so this only ever applies to the first-person arms. **5. The solver is committed, not thrown away:** `Tools/Unreal/grip_solve.py` re-derives the offset from a grip log, and it validates its own quaternion-to-rotator conversion against the engine's printed pairs before it answers (0.00000 deg round-trip on all 44 logged pairs; the closed-form inverse was wrong and is gone). This change also carries the two documentation sessions that were sitting uncommitted on disk - **068** (the MOUT kit registered) and **069** (the Ravenshoe world treatment and nav doctrine) - in their own commit. ### FILES CHANGED Created: `Docs/evidence/handik_pos/` (the five raw probe runs, committed as `.txt` because `.gitignore` ignores `*.log`), `Docs/evidence/handik_rot/` (`a88_off000.png`, `a88_fixed.png`), `Tools/Unreal/grip_solve.py` Modified: `Plugins/SouthernSpearLyraBridge/Source/SouthernSpearLyraBridge/Public/SSHandIKMeshComponent.h` (`bRotateHandToGrip`, `HandRotationOffset`, `RotateChain`), `.../Private/SSHandIKMeshComponent.cpp` (the pure `RotateChain`, the `GripCS` rotation and the call in `FinalizeBoneTransform`), `.../Private/SSFirstPersonSubsystem.cpp` (the arms component turns it on), `.../Private/Tests/SSHandIKTests.cpp` (`SouthernSpear.Bridge.HandIK.RotateHand`), `Config/DefaultGame.ini`, `Docs/CHANGELOG.md` Scratch (untracked, `Build/`): `handik_rot/{a88_off000,a88_fixed}.png`, `handik_pos` probes, `make_handik_evidence.py`. Also untracked and deliberately so: `SSHandIKProbeSubsystem.h/.cpp` (`-SSHandIKProbe`) - the throwaway measurement and ablation probe, self-labelled "not for commit", with the scratch probes rather than in the module. ### TESTING | Check | Command | Result | |---|---|---| | Hand position, A88 | `-SSHandIKProbe` game run, `Saved/Logs/SS_handik_pos_a88*.log` | **PASS** - 8.37 -> 0.00 cm, 11/11 samples, `reach_limited=0` | | Reach gate honoured | same logs, body `CharacterMesh0` | **PASS (negative)** - the body trips the gate at alpha 0 (`reach_limited=1`) and the arms never do | | Hand rotation in the live game | `Saved/Logs/SS_grip_a88_fix.log`, 22 GRIP samples | **PASS** - arms' `hand_in_weapon` equals the body's on every steady-state sample; `hand_vs_socket` 0.0 -> 157.6 deg | | Offset solve | `python Tools/Unreal/grip_solve.py Saved/Logs/SS_grip_a88.log` | **PASS** - engine round-trip 0.00000 deg on all 4 pairs, predicted match 0.00 deg | | Rotation unit test | `Automation RunTests SouthernSpear.Bridge.HandIK.RotateHand` | **PASS** - alpha 1 matches the target, position kept, a finger keeps its local transform, alpha 0.5 is the slerp midpoint, zero offset takes the socket, 3 refusals leave the pose untouched | | Position unit test | `Automation RunTests SouthernSpear.Bridge.HandIK.Solve` | **PASS** - unchanged | | Full suite | `Automation RunTests SouthernSpear` | 61 pass, 1 fail - `Network.Gameplay.TwoPlayerAuthoritySmoke` (`ViewportOverlayWidget.IsValid()`). Called environmental here; **corrected in Session 071** - the run that produced it left out `-NoLoadingScreen`, which is the flag the test needs in a headless world. Not an environmental fault. | | Editor build | `Build.bat SouthernSpearEditor Win64 Development` | **PASS - Succeeded** | | A89 position probe | - | **NOT RUN** - only the A88 was sampled | | Tan-block ablation | `ss.Probe.Hide` | **NOT RUN** - component unidentified | | Finger-grip frames in `A_FP_Rifle_*` | - | **NOT RUN** - task stated as report-only | ### ASSETS - No asset imported, modified or moved. The capture pair in `Docs/evidence/handik_rot/` is screenshot evidence, not content. - The arms mesh does carry fingers: `LeftHandThumb1-4`, `LeftHandIndex1-4`, `LeftHandMiddle1-4`, `LeftHandRing1-4`, `LeftHandPinky1-4` (and the right-hand mirror) are in the import sidecar `Build/fp_arms/SK_FP_Arms_Rifle.fbx.json`, from the Fab **M4 - FPS Weapon Animations Pack FREE**, whose six clips were cut from one FBX at the frame ranges that sidecar lists. A closed-hand grip pose is therefore representable on this skeleton; whether any of the six clips contains one was **not measured**. ### RISKS - **R-80 (new):** the arms' finger bones are **not** covered by any live measurement. `RotateChain`'s propagation is proved by the unit test on a synthetic chain, and the grip logs' anatomy columns cannot prove it on `SK_FP_Arms_Rifle` (see the defect below), so "the fingers follow the hand" is unverified in the real mesh until the probe is re-run with a side-correct bone lookup. - **R-81 (new):** the walking tan block is still unidentified, and it is a *rendering* defect that no test can see - it needs the ablation run (or a first-person capture with the view model hidden) to name the component. ### DEFECTS FOUND - **The probe measured the wrong hand.** `ProbeNamedBone` returns the first bone whose name *contains* the stem, and `SK_FP_Arms_Rifle`'s bone order puts the right arm first (`M4_Root, FPS_Camera_j, Spine, RightArm, RightForeArm, RightHand, RightHandThumb1-4, ...` before `LeftArm`), so the arms' `thumb`, `index`, `palmraw` and `mid` columns are **`RightHandThumb1`/`RightHandIndex1`/`RightHandMiddle1` measured against the left hand's position**. The `thumb_vs_barrel=137.6deg` printed for the arms in both grip logs is therefore not the left thumb - it is the right thumb, and it is why the figure did not move when the hand did. The pre-fix finding itself stands, because it was really carried by the hand quaternion (`hand_vs_socket=0.0deg` against the body's 157.6deg, `hand_in_weapon=(0,1,0,0)` against the body's), and the offset solve never used the thumb columns. The body's own `thumb_vs_barrel=12.9deg` is valid: Manny's bone order puts `thumb_01_l` before any right-side thumb. - **The risk register in `PROJECT_AUDIT.md` now lags the changelog by eleven numbers**, and Session 069's entry re-used two live ids: it introduces **R-69 and R-70** while this changelog already carries R-69 (the press kit's 420 px logo, open) and R-70 (the rest-joint solve, open). R-69 and R-70 are therefore each claimed twice. This session's new risks take **R-80/R-81**, the next numbers free in this file. - The position log's A88-only sampling is a gap against the ask, not a report of a failure: the A89 run simply was not taken. ### NEXT ACTION Fix the probe's bone lookup (match the whole bone name, or require the bone to be a descendant of the solved `Hand` index) and re-run `-SSHandIKProbe` on the **A89**, which answers three open things at once: the A89's hand-to-grip distance and reach headroom, the arms' real left-thumb/palm axes after the rotation step, and whether the fingers follow it. Then take the `ss.Probe.Hide` ablation on the same run to name the tan block. ## Session 071 — 2026-09-29 — The hand is turned by its own anatomy, not by the body's hand bone The producer's correction: Session 070 wrote "the transfer is exact", and it was exact only in the bone language. Copying the body's hand-relative-to-weapon quaternion onto the first-person arms assumes both hands point their bone axes the same way, and Manny's `hand_l` and the Fab arms' `LeftHand` do not - which is the ~90 deg roll that stayed in the render. This session measures each hand from its finger bones instead, solves the arms' angle from that directly, and prints the three axis errors at every step. ### COMPLETED **The defect, measured instead of argued.** Each hand is now read off its own finger bones - the distal thumb, index and middle bones *under the solved hand index*, palm normal = thumb x finger, then re-orthogonalised - and both hands' frames are expressed in the weapon's frame, which is the only language two rigs can be compared in. On the A88, before any correction: | hand | thumb vs the socket->muzzle line | frame vs the body's measured frame | |---|---|---| | Manny body (`thumb_03_l`, `index_03_l`) | 11.3 deg | - (the reference) | | FP arms (`LeftHandThumb4`, `LeftHandIndex4`, animated) | **15.0 deg** | palm **82.8 deg**, thumb 11.5, finger **82.1** | So the arms' *animation* already has the thumb down the barrel (15.0 against the body's 11.3). What it is missing is not the thumb at all but **83 deg of roll about it**: the palm normal and the finger direction are both ~82 deg out. A hand-bone-quaternion transfer cannot see that axis, and did not: it reported the arms matching the body while the palm stayed sideways. **The solve, and where it lives.** The arms' measured frame is aligned to the body's measured frame (both taken in the weapon's frame) with `align = twist * swing` - a shortest arc onto the palm normal, then the twist about it that brings the thumb across - and the config offset is `offset = socketrot^-1 * (align * hand)`. Logged by the probe as a ready config line, and written to `[/Script/SouthernSpearLyraBridge.SSHandIKMeshComponent] HandRotationOffset=(Pitch=70.599,Yaw=-79.336,Roll=19.550)` (align magnitude 83.0 deg). In-run residual after the alignment: **0.0 / 0.0 / 0.0 deg**. **Verified live, in the engine's own pose** (full run, rotation step on, the probe reading the published bone transforms): the arms' palm, thumb and finger now sit **0.0-0.2 deg** from the body's measured grip on every steady-state sample, and `thumb_vs_barrel` reads 11.3 deg against the body's 11.3. The body is unchanged to the digit (`hand_in_weapon=(-0.8980,-0.1944,0.2984,0.2585)`, `thumb_vs_barrel=11.3`, `hand_vs_socket=157.6`) because `bRotateHandToGrip` stays off for it. **Composition order mattered, and is now pinned by evidence, not by assumption.** Three angle errors of 6-9 deg remained when the arcs were composed the other way. UE's `A * B` applies **B first**, checked against the engine's own logged triple (`weapon_in_cs * hand_in_weapon` reproduces that mesh's `handrot` exactly), so the probe composes `twist * swing` and `align * hand`; the residual went to 0.0. **Which fore-end the A88 actually has.** `Tools/build_adfrc_weapons.py` takes the A88's grip from the ADFRC pose **`EF88_Vg_static`** (`GRIP_CLIPS`), i.e. the EF88 *with a vertical grip*, and `SOCKET_LeftHandGrip` is that pose's own left wrist (report: `Docs/evidence/w2_grip_clips/EF88_Vg_static.json`). So the hand is placed on a **vertical foregrip**, while both candidate orientations - the body's Lyra hold and the arms' own Fab M4 hold - are *horizontal handguard* holds: in both, the thumb runs within 11-15 deg of the socket->muzzle line and the palm normal is perpendicular to it. Aligning the arms to the body is therefore right in the project's own reference terms and **not yet proven right for this fore-end**: a vertical post wants the palm across the post, not across a handguard. **The smoke-test failure was not environmental.** Asked to check it before calling it environmental, and it does not survive the check. With this change reverted and rebuilt, the single test passes; on the same tree restored, with the change in place, `-NoLoadingScreen` decided it: **without the flag `Result={Fail}` with 4 `ViewportOverlayWidget` ensures, with the flag `Result={Success}` and 0**. The full-suite run that produced "61 pass, 1 fail" omitted the flag, which `CLAUDE.md`'s recipe and commit `ac9b642b` both call for. Session 070's TESTING row is corrected above. ### FILES CHANGED Modified: `Config/DefaultGame.ini` (the offset, 70.599/-79.336/19.550, and what it is solved from), `Docs/CHANGELOG.md`. The measurement instrument is the untracked dev probe `SSHandIKProbeSubsystem.h/.cpp`: this session added the descendant-correct finger lookup, the per-hand frame, the `-SSHandIKProbeNoRotate` measurement pose and the `OFFSET` line that prints the config value. ### TESTING | Check | Command | Result | |---|---|---| | Arms vs body, measured, before | `-SSHandIKProbe -SSHandIKProbeNoRotate` (`Saved/Logs/SS_anat_a88_v3.log`) | **FINDING** - palm 82.8, thumb 11.5, finger 82.1 deg; thumb vs the barrel line 15.0 against the body's 11.3 | | Offset solve | same run, `OFFSET` line | **PASS** - align 83.0 deg, residual 0.0/0.0/0.0, `HandRotationOffset=(Pitch=70.599,Yaw=-79.336,Roll=19.550)` | | Arms vs body, measured, after | probe run with the offset in config (`Saved/Logs/SS_anat_a88_fixed.log`) | **PASS** - 0.0-0.2 deg on all three axes, every steady-state sample | | Body untouched | same log, `CharacterMesh0` | **PASS** - `hand_in_weapon` and `thumb_vs_barrel` identical to Session 070 | | Hand position | same log | **PASS** - held from Session 070: 8.37 cm -> 0.00 cm, gate never fired | | Composition order | engine's own logged triple, `weapon_in_cs * hand_in_weapon == handrot` | **PASS** - `A * B` applies B first; the probe's arcs corrected accordingly | | `TwoPlayerAuthoritySmoke`, change present | `UnrealEditor-Cmd -nullrhi ... -ExecCmds="Automation RunTests SouthernSpear.Network.Gameplay.TwoPlayerAuthoritySmoke"` | **FAIL without `-NoLoadingScreen`** (4 ensures), **PASS with it** - the invocation, not the machine | | `TwoPlayerAuthoritySmoke`, change reverted and rebuilt | same command | **PASS** with the flag - the change is not implicated either way | | A88 grip pose / fore-end | `Tools/build_adfrc_weapons.py` `GRIP_CLIPS`, `Docs/evidence/w2_grip_clips/EF88_Vg_static.json` | **FINDING** - the A88's socket is the `EF88_Vg_static` left wrist: a vertical foregrip | | A89 run | - | **NOT RUN** - the probe's finger lookup is only now correct; the A89 should be re-measured with it | ### ASSETS - No asset imported, modified or moved. `Docs/evidence/w2_grip_clips/EF88_Vg_static.json` predates this session; it is the ADFRC pose the A88's grip socket is fitted from, and it is what names the fore-end. ### RISKS - **R-82 (new):** the correction targets the **body's** grip, and the A88's fore-end is a vertical grip. For this weapon the authoritative reference is the asset's own pose (`EF88_Vg_static`), not Lyra's hold; until that is measured, the arms hold the A88 the way Lyra holds its own rifle, to 0.2 deg. - **R-83 (new):** the probe's `barrel` is the grip socket -> muzzle socket line, and on this asset that line is ~26 deg off the mesh's own +X (`barrel_local=(0.9,0.4,0.3)`), because the socket is the left wrist, not a point on the bore. `thumb_vs_barrel` is therefore a comparison between hands, not a measurement of the thumb against the barrel. ### DEFECTS FOUND - **The probe's perpendicular was degenerate in every earlier run.** `axis_socket_to_barrel` was always `(0,0,0)`, because it was built from the grip socket and the muzzle, and the grip socket lies on that line by definition. The palm was being tested against a zero vector. - **The probe's finger lookup was measuring the wrong hand** (raised in Session 070 as R-80, fixed here at the source): `SK_FP_Arms_Rifle` lists `RightHandThumb1` before `LeftHandThumb1`, so every `thumb`/`index` number the arms printed before this session was the right hand's, relative to the left hand's position. - **Session 070's "the transfer is exact"** was exact only in the bone-frame language; the axis mismatch between two rigs is invisible to a quaternion transfer and was worth 83 deg of roll on this pair. - **The smoke-test failure was misdiagnosed as environmental in Session 070**, and the fix was a missing command-line flag in the run, not a machine fault. ### NEXT ACTION Measure the A88's own pose - the left hand in `EF88_Vg_static` (`Docs/evidence/w2_grip_clips/EF88_Vg_static.json`, via `Tools/Common/adfrc_grip.py`) - in the same three axes, and decide the target on that evidence: if the asset's own hold differs from the body's by the roll a vertical foregrip implies, solve the offset to the asset's pose instead and change one config line. Then re-run `-SSHandIKProbe` on the **A89** (its fingers are now found correctly) together with the `ss.Probe.Hide` ablation for the tan block. --- ## Session 072 — 2026-09-29 — The nav docs stopped teaching a measured fault; and the two packs get a look-check verdict Continuation of Session 069's Ravenshoe work. Two items: close the documentation half of the nav forensics (the docs half of the fault, now that the bake is the one action still owed), and give the producer the desk-level look-check verdict on the two packs Session 068 registered, so the decision about them is grounded before any editor work. ### COMPLETED - **Risk renumbering, correcting Session 069's entry:** its risks were written as "R-69"/"R-70", numbers Session 070 then found already claimed (press-kit logo; rest-joint solve). They take **R-82/R-83**; the changelog entry now says so inline, `PROJECT_AUDIT.md` carries the two rows (R-82 OPEN as a machine/attendant constraint, R-83 CLOSED with the doc correction), and `MAPS_RAVENSHOE.md` §7.1's pointer cites R-82. - **The retired ini flag removed from every working doc and recipe (closes the doc half of R-83):** `PLAYTEST_COMMANDS.md` §5 now says plainly that no headless run bakes nav (R-82) and marks the flag retired; `MAPS_DRYRIVER.md` §11.1 teaches no overrides and gains **§11.6**, the full correction — what the flag actually does (early load-time build on partial geometry; holds the 0x20 lock), why Dry River's headless passes kept "working" (the persisted gate-G1.1 tiles), and the rule now (bake once attended, verify by path query); `CLAUDE.md`'s pipeline block drops the override and points at §11.6; `HANDOVER_RAVENSHOE.md` §0 #1 and the rebuild order are corrected in place; `bake_ravenshoe_nav.py` (superseded, referenced nowhere) gains a do-not-run banner; both `build_dryriver_*.py` headers gain dated corrections; and `.github/workflows/build.yml` stops passing the flag (three sites) — inert in CI, but CI should not teach the fault either. - **Pack look check, desk-level, in `HANDOVER_RAVENSHOE.md` §0b:** MOUT kit **rejected for Ravenshoe** (church/playground/police-signage/European-vernacular per the Singapore Canal precedent; stays a candidate for an urban/CQB map, R-67 measurement still owed); RustyCars **recommended as the deck/bed wreck replacement** — four shells + ivy, seller verified, `isAiForbidden: false`, against the incumbent Renault whose flag is unknown (no metadata) — provenance hardening under ADR-028, UE-native so the M-008i in-place route applies. No content was imported; only the §0b assessment was written. - Session 069's testing table said "look checks NOT RUN — Session 068 scope"; this session records the desk half of that scope. The editor views (MOUT demo map once; RustyCars shells before import) stay owed. ### FILES CHANGED Modified: `Docs/CHANGELOG.md` (this entry + renumbering), `Docs/PROJECT_AUDIT.md` (R-82/R-83 rows + provenance note), `Docs/MAPS_RAVENSHOE.md` (R-82 pointer), `Docs/HANDOVER_RAVENSHOE.md` (§0 corrections, §0b look-check verdicts), `Docs/MAPS_DRYRIVER.md` (§11.1 de-taught, §11.6 added), `Docs/PLAYTEST_COMMANDS.md` (§5 rewritten), `CLAUDE.md` (pipeline block), `.github/workflows/build.yml` (flag removed ×3, comment), `Tools/Unreal/bake_ravenshoe_nav.py` (superseded banner), `Tools/Unreal/build_dryriver_level.py` + `build_dryriver_nav.py` (header corrections). ### TESTING | Check | Command | Result | |---|---|---| | Retired-flag sweep | `grep -rn bWaitForAsyncLoading` across docs, CLAUDE.md, Tools, CI | **PASS** — every remaining mention is inside a retirement/correction context; CI has zero | | Workflow validity | `yaml.safe_load` on build.yml | **PASS** — parses, 22 steps | | Script syntax | `ast.parse` on the three touched Python files | **PASS** | | Line-ending safety | `git diff --numstat` per file | **PASS** — each touched CRLF file shows only its own lines changed; no whole-file churn | | Attended editor bake | — | **NOT RUN** — unchanged from Session 069; still the one action that un-freezes the bots | | MOUT demo map opened (R-67) | — | **NOT RUN** — needs the editor; recorded as owed | | RustyCars shells rendered/inspected | — | **NOT RUN** — needs the editor or a Blender probe; recorded as owed | ### RISKS - No new risks. R-82 stays OPEN (machine/attendant constraint); R-83 is CLOSED by this change. The renumbering itself removes the standing double-claim of R-69/R-70 that Session 070 flagged. ### NEXT ACTION Unchanged and singular: **the attended editor bake** — open `/Game/Maps/L_Ravenshoe_01`, Build ▸ Build Paths, save — then `build_ravenshoe_nav.py` verify mode must read 32/32 and a live `-game` run must log `Steered N idle bot(s)` with N > 0. The RustyCars swap rides the next map-touching session after that. --- ## Session 073 — 2026-09-29 — The hold socket's rotation was inverted by our own transpose, not by the FBX export ### COMPLETED - **Found the cause of the upside-down hold socket.** Session 071's probe read `SOCKET_LeftHandGrip` in game with its thumb axis pointing down, not along the bore, and put that down to the FBX exporter applying its axis change on the wrong side. The fault is in `Tools/Blender/adfrc_weapon.py`. `adfrc_grip.hold_frame_rotation` returns a tuple of rows whose columns are the hand axes (checked: column 0 = the report's palm, [0.1024, 0.5, 0.86] on an A88-like weapon). `mathutils.Matrix()` also takes rows, so it already builds the rotation, and the `.transposed()` added in 520dc043 turned it into the inverse. The transpose is removed. - **The fore-end question is settled.** Session 071 raised it as a risk; that number was renumbered to Ravenshoe in Session 072. The A88 has no vertical foregrip. The `gl*` memory points mark an under-barrel launcher that isn't modelled, and nothing more than 0.9 cm below the bore runs from 9 cm to 42 cm ahead of the trigger. So the hold is the plain handguard; `EF88_Vg_static` supplies position only. ### FILES CHANGED `Tools/Blender/adfrc_weapon.py`, `Docs/CHANGELOG.md`. ### TESTING - `python Tools/Common/test_adfrc_grip.py` → exit 0, 0 failures (no pure code changed). - The layout check that `resolve_hold`'s column 0 equals its reported palm: done in Python, output above. - NOT RUN: the Blender rebuild, the Unreal import, the in-game probe, the automation suite (all on the producer's machine). ### ASSETS `Art/Weapons/A88/ADFRC/SM_A88.fbx` as committed in 520dc043 carries the inverted socket rotation. It must be rebuilt. ### RISKS - R-84: `HandRotationOffset` in `DefaultGame.ini` was solved against the socket's old rotation (before the hold basis). Once the rebuilt socket carries the hold basis, the offset has to be re-solved, or the hand turns by both. ### DEFECTS FOUND - The transpose in `adfrc_weapon.py`, found by reading the matrix layout against `hold_frame_rotation`'s docstring and report. ### NEXT ACTION Rebuild the A88 (`python Tools/build_adfrc_weapons.py`, then `setup_weapons.py`). Confirm with `probe_weapon_socket.py` that the socket's +Z (thumb) lies along the bore. Re-solve `HandRotationOffset` against the new socket, log the three angles (each < 15°), then take a screenshot. --- ## Session 074 — 2026-09-29 — The hold is authored from the weapon, and the weapon says which hand it is The producer's correction, twice over. The Arma-pose derivation is dropped: the finger rest joints of that pose disagree with each other by ~50 deg, so it was never a measurement, only a shape. And a hold carried through the FBX is not a measurement either: the frame authored on `SOCKET_LeftHandGrip` came back in game with the thumb on `(0,0,-1)` while the bore at that socket is `(0.874,0.388,0.291)`. Session 073 found the cause — not the exporter, as I first wrote, but our own `.transposed()` on a matrix that was already the rotation. The fix belongs to whichever route carries a rotation through the export at all. This session does not carry one: socket *positions* survive the round trip exactly, so the hold is built at run time from the sockets, with its per-weapon angle as data in config, and Session 073's R-84 (re-solve the offset against a rebuilt socket) is answered by not using one. ### COMPLETED **Point 1 — does `SM_A88` have a vertical foregrip? No.** Two independent measurements of the shipped asset agree. The `ADFRC_EF88_MLOD.blend` memory points `gl` (+17.3 cm), `gl_axis` (+27.2), `gl_cartridge_axis` (+21.9), `gl_lock_axis` (+12.7) and `muzzle_ugl_pos` (+31.6) all sit 4–5 cm *below* the bore: they are grenade-launcher attachment proxies, and there is no grenade-launcher geometry in LOD0 at all. In `SM_A88.fbx` (78 615 verts), every vertex more than 2 cm below the bore line lies between −35.9 and +9.1 cm — the stock, pistol grip, magazine and trigger guard. From +9.1 cm to the muzzle at +42.4 cm, nothing is more than 0.9 cm below the bore. The fore-end is a plain tube handguard with a top rail, the hand surface runs from about +9 cm to +30 cm ahead of the trigger, and the existing `LeftHandGrip` socket at +21.9 cm sits mid-handguard. `GRIP_CLIPS` still names `EF88_Vg_static`, but “Vg” names the handAnim *clip*, not the gun; it supplies position only, and `build_adfrc_weapons.py` now says so at the line. **Point 2 — the hold is built at run time, and it lives in config.** `FSSHandIK::BuildGripHold` takes the Muzzle and RightHandGrip positions and the weapon component's own up vector and produces the frame: `forward = muzzle − rightGrip`, `up` = component up orthogonalised against it, `right = up × forward` (UE satisfies `F × R = U`, so `U × F = R`), then `thumb = forward`, `palm` = up tilted `PalmTiltDeg` towards `right`, `finger = palm × thumb`. `USSHandIKMeshComponent::UpdateGrip` rebuilds it in the attach frame every time the grip resolves, and a weapon with no muzzle or right-hand socket falls back to the grip socket's own rotation, which is what the code did before. The angle is data, per weapon, in `Config/DefaultGame.ini`: `GripPalmTiltDeg=(A88=30)` and `DefaultPalmTiltDeg=30`, keyed off the held mesh's name so a weapon needs no code. Only the A88 is measured; the rest take the default until someone measures their mesh the same way, because a hold guessed from a weapon's real-world type is exactly the borrowed pose this replaced. The socket-rotation application in `adfrc_weapon.py` is reverted — the FBX is back to position-only and the manifest keeps the resolved frame as documentation. **The sign of `right` is checked against the weapon, not derived.** The ejection port is on the shooter's right and the left hand grip on the shooter's left, so `Eject` must land on `+Right` and `LeftHandGrip` on `−Right`. The probe now prints both every sample: `eject_on_right=1 left_grip_on_left=1 eject_dot_right=+1.625 grip_dot_right=−9.492`. **Handedness, and the bug a 0.0° residual was hiding.** A *left* hand has `palm = thumb × finger`, so `palm × finger = −thumb`: its own (palm, finger, thumb) triad is **left-handed and is not a rotation**. The first version of this work used `finger = thumb × palm` — a *right* hand's relation — and measured the palm as `finger × thumb` to match it. It reported `residual=0.0/0.0/0.0` on all eleven samples, because a mirrored frame still matches a mirrored target, and rendered the back of a closed fist on the handguard. Both sides are corrected. The pre-offset error against the arms' own clip pose fell from 124.8/125.4° to 55.2/54.6°, and the fingers now wrap the tube. `ToHandRotation` is `FRotationMatrix::MakeFromXY(Palm, Finger)` — the engine's own helper, so its row/column convention is the one `FQuat` expects rather than one re-derived — which gives `+X` palm, `+Y` finger and `+Z` the back-of-hand axis. `Tools/Common/adfrc_grip.py` carried the same mirrored triad and is corrected with it; its `plain_handguard` finger axis is `−right` for the same reason. **Point 3 — the offset, solved from `SK_FP_Arms_Rifle`'s own finger bones.** `HandRotationOffset=(Pitch=33.853,Yaw=136.820,Roll=−89.996)`, in `Config/DefaultGame.ini`. Solved on the A88 from its own clip pose with `-SSHandIKProbe -SSHandIKProbeNoRotate`, identical on all eleven samples, and the **three post-offset angles are 0.0 / 0.0 / 0.0** — each far inside the 15° bar. Live afterwards, with the offset in config and the rotation step on, every sample t=0.5…5.0 s reads `palm_err=0.0 thumb_err=0.0 finger_err=0.0`, and the hold's thumb reads `(1.0,0.0,0.1)` in the weapon's frame: on the bore, which is the whole point of rebuilding it at run time. The `thumb_vs_bore=14.7°` figure is not an error — it is how far the arms' own idle clip already holds its thumb off the bore, the anatomical fact Session 071 measured at 11.5°. **The per-weapon data path, and a silent failure worth knowing about.** `GripPalmTiltDeg` is a `TMap`, and its ini form is unforgiving: `+GripPalmTiltDeg=(A88=30.0)` logs `import failed for GripPalmTiltDeg` and leaves the property **empty**, so every weapon silently falls back to `DefaultPalmTiltDeg` with nothing else looking wrong. `(("A88",30.0))` is the form that imports. It was caught by putting a distinctive 45 in the row and reading the realised tilt back — 30.0° with the broken form, 45.0° with the fixed one. Both the ini and the code now say so, because "my per-weapon value is being ignored" is exactly the symptom that sends people looking in the wrong place. **Reading the final live run correctly.** The FP arms settle at `0.0 / 0.0 / 0.0` from t=1.5 s and stay there for the remaining eight samples. The first three read 55.2 / 20.4 / 55.2 on the way there: `Alpha` ramps at `BlendSpeed=8/s`, so the hand is still the clip's own pose at t=0.4 and halfway at t=1.0. That is the blend, not an error, and it is why "every sample reads 0.0" is not a claim this session can make about the fade-in window. **Point 4 — `Build/measure_a88_grip_pose.py` deleted.** It was the cancelled Arma derivation, it was never committed (`Build/` is gitignored), and no committed value depends on it. The socket probe it fed, `Tools/Unreal/probe_weapon_socket.py`, is removed too: the runtime no longer reads a socket rotation, so there is nothing left for it to report. **Point 5 — `TwoPlayerAuthoritySmoke` on main without the change: PASS**, and **PASS with it**. It is not environmental and not this change. The `Fail` in Session 071's evidence is the same test run *without* `-NoLoadingScreen`, which produces 4 `ViewportOverlayWidget` ensures from `GameViewportClient.cpp:3378`; with the flag it passes, and it passed on the tree as it stood before this work began (`SS_smoke_now.log`, exit 0, 0 ensures). The flag, not the code, is the variable. The flag needs recording in `Docs/PLAYTEST_COMMANDS.md` so nobody re-derives it. ### TESTS - `SouthernSpear.Bridge.HandIK.GripHold` (new, pure, no world) — Success. A synthetic bore at +X with up +Z: thumb on the bore, `right = +Y`, `palm = (0,0.5,0.866)` i.e. exactly 30° off up towards right, `palm = thumb × finger`, `palm × finger = −thumb`, `det[palm,finger,−thumb] = +1`, `ToHandRotation` `+Z = −thumb`, 0° and 90° tilts, the frame follows a slanted weapon, and both refusals (coincident sockets, up along the bore) return false with the hold **zeroed** rather than stale — the first cut left `Out.Forward` set on the second refusal, which the test caught. - `SouthernSpear.Bridge.HandIK.RotateHand` — Success. `SouthernSpear.Bridge.HandIK.Solve` — Success. - Full suite: **63 pass, 0 fail** (`Saved/Logs/SS_suite_072.log`, exit 0). - `Tools/Common/test_adfrc_grip.py` 0 failures, `Tools/Common/test_adfrc_reload.py` 0 failures. The grip suite's handedness block was asserting the right hand's relation, so it passed on a mirrored frame; it now asserts the left hand's, and that it is *not* the right hand's. - Editor build clean. ### RISKS - **R-85** — The hold's `PalmTiltDeg` is authored for the A88 only, and only from that weapon's own mesh. Every other weapon takes the 30° default, which is a guess dressed as a default. Open. - **R-86** — The A88's palm still reads as a little high and left of the tube in the capture. The rotation is right (the thumb is on the bore and the fingers wrap it); what is left is the *position* of `LeftHandGrip`, 10.3 cm left of the bore, which comes from the ADFRC handAnim wrist and has never been authored. Open — this is the next thing to look at, and it is a socket-position problem, not a hold problem. - **R-87** — The socket-rotation route was abandoned in favour of building the hold at run time, and `adfrc_weapon.py` no longer writes a rotation to `SOCKET_LeftHandGrip` at all. If a future weapon needs a hold the runtime cannot build from muzzle/right-grip/up alone — a weapon with no muzzle socket, say — it falls back to the grip socket's own rotation, which is arbitrary and, if anyone reintroduces a rotation here, must not be transposed on the way out (Session 073). Open. - **R-88** — `-NoLoadingScreen` is required for `TwoPlayerAuthoritySmoke` and is not yet written down in `Docs/PLAYTEST_COMMANDS.md`, so the next person to run it without the flag will read a `Fail` as a regression. Open. ### FILES `Plugins/SouthernSpearLyraBridge/Source/SouthernSpearLyraBridge/{Public/SSHandIKMeshComponent.h, Private/SSHandIKMeshComponent.cpp, Private/SSHandIKProbeSubsystem.cpp, Private/Tests/SSHandIKTests.cpp}`, `Config/DefaultGame.ini`, `Tools/Common/adfrc_grip.py`, `Tools/Common/test_adfrc_grip.py`, `Tools/Blender/adfrc_weapon.py`, `Tools/build_adfrc_weapons.py`, `Art/Weapons/A88/ADFRC/manifest.json`, `Art/Weapons/A88/ADFRC/SM_A88.fbx` (rebuilt, position-only sockets again), deleted `Tools/Unreal/probe_weapon_socket.py`, evidence `Docs/evidence/handik_hold/SS_hold_frame.txt`. --- ## Session 075 — 2026-09-29 — Casualty care, step 1: the rules, and every asset usable whatever its AI flag ### COMPLETED - **ADR-040 accepted** by the producer, with its build order: rules, then component and kit, then the bridge, then the HUD. - **`isAiForbidden` overruled for every asset (L-0016d, producer).** Recorded in the licence register, CLAUDE.md's content rules and ADR-040. The flag is still recorded at import time; it no longer holds anything back. The Fab IFAK is now `PLANNED` as the medic's kit. - **New module `SouthernSpearCasualty`** (Core only), enabled in the `.uproject`. Step 1 is its rules: `Public/SSCasualtyRules.h`, engine-free like `SSInsigniaRaster.h`. It covers hit zones and bleeding, the downed state and bleed-out, being finished, every treatment row in ADR-040 (who may do it, how long, the outcome), dressings, the medic's kit (charges, lifetime, carry limits) and `Tick`, which never raises health. - **One set of checks, run twice.** `Private/Tests/SSCasualtyRuleChecks.h` runs in the automation suite (`SouthernSpear.Casualty.Rules`) and outside Unreal (`Tools/Casualty/check_casualty_rules.py`, g++ with `-Wall -Wextra -Werror`). ### FILES CHANGED New: `Plugins/SouthernSpearCasualty/` (`.uplugin`, `Build.cs`, `SSCasualtyModule.cpp`, `Public/SSCasualtyRules.h`, `Private/Tests/SSCasualtyRuleChecks.h`, `Private/Tests/SSCasualtyTests.cpp`), `Tools/Casualty/check_casualty_rules.py`. Modified: `SouthernSpear.uproject`, `CLAUDE.md`, `Docs/DECISION_LOG.md` (ADR-040 accepted), `Docs/LICENCE_REGISTER.md` (L-0016d), `Docs/ASSET_REGISTER.md` (IFAK row), `Docs/CHANGELOG.md`. ### TESTING - `python Tools/Casualty/check_casualty_rules.py` → exit 0, "53 checks, 0 failure(s)". - Mutation check (scratch copy, four separate mutations; each caught): head hits never kill → 2 failures; `Tick` regenerates → 3; self-dressing faster than a teammate's → 2; kit treatment costs no charge → 1. - `python Tools/validate_architecture.py` → exit 0 (the new module passes SS001/SS002/SS005). - `python Tools/check_unity_names.py` → exit 1, but the same on `main` without this change: two shadows in `SSHandIKProbeSubsystem.cpp` (`GHaveBodyFrame`, `GBodyFrameInWeapon`) from 520dc043. Not this change's; see DEFECTS. - NOT RUN: the editor build and `SouthernSpear.Casualty.Rules` in the automation suite (producer's machine). ### ASSETS None imported. ### RISKS - R-89: until the bridge wires it (step 3), the rules are inert in play. Lyra's own death still applies. ### DEFECTS FOUND - `check_unity_names.py` fails on `main` in `SSHandIKProbeSubsystem.cpp` (found by running the guard before committing). The parallel session's unpushed 024b9f0d rewrites that file, so it is left to that merge. ### NEXT ACTION Build the editor and run `SouthernSpear.Casualty.Rules` (expect 1 more test, and all to pass with `-NoLoadingScreen`). Then step 2: `USSCasualtySettings` (the `FTuning` numbers in `DefaultGame.ini`), the replicated `USSCasualtyComponent`, and `ASSMedicalKit`. --- ## Session 076 — 2026-09-29 — Loading screens name the operation and its rules; the front end fits at 1080p ### COMPLETED - **Loading screen per operation.** When the destination is an operation, the screen shows: - "LOADING OPERATION" and the map's name, e.g. RED GUM STATION; - its objective count and terrain line, and its description; - the rule set's name, with a short summary written from ADR-018/ADR-031 of what the rules ask of you; - a tip. It shows the map's own art when `T_SS_Load_` exists, else the key art. The front end and unlisted maps get the plain screen, as before. Where the destination comes from: - the front end's request (`USSMenuWidget::PendingMap`, set just before `OpenLevel`); - otherwise the engine's `TravelURL`/`LastURL`, where `Rules=Section` selects the rules. - **One operation list.** `Private/SSOperations.h` holds the maps, art keys, titles, descriptions and meta lines. The front-end cards, their handlers, the loading screen and `setup_ui.py` all read it. - **Drone shots drop in without code.** Place `Docs/images/loadingscreens/.png` (keys listed in its README), then run `setup_ui.py` with `SS_UI_LOADING_ONLY=1`. - **Front end fits at 1080p.** The operation grid is 3 columns (was 2). In the producer's screenshot the RULES row sat under the disclaimer and its buttons were cut off. Cards are 6 px taller so a three-line description no longer touches DEPLOY. - **Scoreboard:** the team total reads "3 KILLS", not the ambiguous "3K". ### FILES CHANGED New: `Plugins/SouthernSpearUI/Source/SouthernSpearUI/Private/SSOperations.h`, `Docs/images/loadingscreens/README.md`. Modified: `SSLoadingScreenWidget.h/.cpp`, `SSMenuWidget.h/.cpp`, `SSScoreboardWidget.cpp`, `SSUIAssets.h`, `Tools/Unreal/setup_ui.py`, `Docs/CHANGELOG.md`. ### TESTING - `python Tools/validate_architecture.py` → exit 0. - `python Tools/check_unity_names.py` → exit 1, the same two pre-existing shadows in `SSHandIKProbeSubsystem.cpp` as in Session 075. Nothing from this change. - `python -m py_compile Tools/Unreal/setup_ui.py` → ok. The art-key regex returns the five keys from `SSOperations.h`. - NOT RUN: the editor build; a front-end deploy to see the new loading screen; a 1080p front-end capture. ### ASSETS None yet. Per-map loading art is waiting on the producer's drone shots. ### RISKS None new. ### DEFECTS FOUND - The front end's RULES row was clipped at 1080p (found from the producer's screenshot). - Scoreboard names show the platform default ("hurleym-…") because no callsign is set, and the only way to set one is the `ss.Callsign` console command. A callsign field on the front end needs a request path through Core, because UI may not depend on Progression (SS001). Not done here. - The class-selection preview holds an M4-pattern rifle while the card says A88. Not investigated. ### NEXT ACTION Build the editor. Deploy to Red Gum from the front end once with each rule set, and screenshot the loading screen (`-SSShotAt` is too late for a load; use the editor's High Resolution Screenshot during the load, or a paused run). --- ## Session 077 — 2026-09-29 — Casualty care, step 2: settings, component and kit; loading art path ### COMPLETED - **Loading art path corrected** to `Docs/images/loadingscreens/` (the producer's folder; `dryriver.png` is already there). `setup_ui.py` now matches file names to art keys **case-insensitively**, so `dryriver.png` is the DryRiver art. - **Casualty step 2 (ADR-040), written blind, not compiled:** - `USSCasualtySettings`: every `FTuning` number, plus the treat range, the kit mesh, `bEveryoneIsMedic`, and the bone-name lists behind `ZoneForBone`. The values are in `Config/DefaultGame.ini`. - `USSCasualtyComponent`: replicated state, bleed, bleed-out, dressings and treatment progress. Server-only entry points for hits, treatments, kit dressings and reset. `OnTransition` (Downed/Died) is for the bridge. Damage cancels treatment both ways, and the patient has to stay in range. - `ASSMedicalKit`: replicated charges and age, self-destroying when spent or old, mesh loaded from settings. - Health only rises inside `FinishTreatment`, through the rules' `Complete`. - **Tests added:** `SouthernSpear.Casualty.Settings` (the ini really imported: empty bone lists mean it failed; zones for Manny's bone names), `.Component` (hit, self-dressing, no regeneration over 60 s, stabilise, finish, cancel), `.Kit` (range, no passive heal, dressing, kit self-treat, charges). ### FILES CHANGED New: `Public/SSCasualtySettings.h`, `SSCasualtyComponent.h`, `SSMedicalKit.h`; `Private/SSCasualtySettings.cpp`, `SSCasualtyComponent.cpp`, `SSMedicalKit.cpp`, `Private/Tests/SSCasualtyRuntimeTests.cpp` (all under `Plugins/SouthernSpearCasualty/Source/SouthernSpearCasualty/`). Modified: `SouthernSpearCasualty.Build.cs` (DeveloperSettings), `Config/DefaultGame.ini`, `Tools/Unreal/setup_ui.py`, `SSOperations.h`/`SSUIAssets.h` comments, `Docs/images/loadingscreens/README.md` (moved), the handover and CLAUDE.md. ### TESTING - `python Tools/Casualty/check_casualty_rules.py` → exit 0, 53 checks, 0 failures (the rules are unchanged). - `python Tools/validate_architecture.py` → exit 0. `python Tools/check_unity_names.py` → the same two pre-existing shadows in `SSHandIKProbeSubsystem.cpp`; nothing from this change. - NOT RUN: the editor build. **None of this C++ has been compiled.** Expect a first-build fix or two (an include, a signature). The three runtime tests have never run. ### ASSETS None. The kit mesh is the engine cube until the Fab IFAK is imported. ### RISKS - R-90: step 2 has not been compiled. The runtime tests build a bare world by hand and tick components manually; if the first run shows `HasAuthority()` false or components not registering, fix the fixture before suspecting the rules. ### DEFECTS FOUND None new. ### NEXT ACTION Build, and run `SouthernSpear.Casualty.*` (expect 4 tests). Then step 3: read Lyra's health and death code (prompt C in `Docs/HANDOVER_CLAUDE_CLOUD.md`) before writing any bridge code. --- ## Session 078 — 2026-09-29 — The grip socket's position gets authored too, and the arm turns out to be the constraint R-86 answered. `GripNudgeCm` joins `GripPalmTiltDeg` as per-weapon data: `(Forward, Right, Up)` in centimetres, in the hold's own axes, applied to the wrist target in `UpdateGrip`. The A88 takes `(Forward=-4.7, Right=5.4, Up=2.1)` and its wrist lands 4.5 cm off the bore, in the 4–5 cm band a hand round a handguard belongs in. The interesting part is not the number, it is that the number **cannot** be chosen by eye, and the first two attempts were wrong in a way that looked like the code was broken. ### COMPLETED **Point 1 — the probe was measuring the wrong point.** It reported `hand_to_grip`, `shoulder_to_grip` and `reach_limited` against the **grip socket**, which is correct only while the nudge is zero. A 6 cm nudge turns the socket into a point the solver is never handed, and the arm's own arithmetic (`reach_limited=0`, `shoulder_to_grip=48.66`) went on describing it. The probe now reads the realised target — `FSSHandIK::ApplyGripNudge(Hold, GripCS, Nudge)` — and redoes the reach figures against that, logging the shoulder, the target and the wrist all in the hold's frame so the numbers can be compared offline. The old line is left in place, still honestly labelled as the socket's. **Point 2 — the arm was at full extension, and that is why the nudge moved the wrist the WRONG way.** `FSSHandIK::Apply` puts the wrist at `A + Dir * clamp(|Target-A|, …, L1+L2-0.001)`. Past `L1+L2` the arm goes straight and the wrist can only ever sit on that sphere. The first nudge put the target 53.7 cm from the shoulder against a 51.8 cm arm: out of reach by 1.9 cm. The `reach_limited` flag read `0` the whole time, because it was reading the socket. Sliding the target around the sphere is not moving it — raising `Right` from 5.5 to 6.0 moved the wrist 0.5 cm *further left*, which looked like a sign error and was not. **Point 3 — reach is why `Forward` is negative.** The first-person shoulder's own bone sits 44.4 cm left of the bore and the whole arm is 51.8 cm, so the socket is already 48.7 cm out: **3.1 cm from straight before any nudge at all.** Every centimetre the wrist moves in towards the tube is a centimetre of arm spent, and it has to be paid back somewhere. Back along the tube is the only somewhere that keeps the hand on the handguard (+9 to +30 cm), so the row pulls the wrist 4.7 cm back as well as 5.4 cm in. Enumerating the whole 4–5 cm ring across the handguard: every point on it needs the entire arm. 1.2 cm of elbow was spent on purpose — 0 cm reads as a pole, not an arm. **Point 4 — solved, then verified.** `Saved/tmp/solve_nudge.py` reproduces the measured run to 0.01 cm (wrist-to-bore 5.54 predicted vs 5.55 logged, `short_by` 1.93 vs 1.92) before it is allowed to predict anything. The live run with the new row: `wrist_to_bore=4.54cm along_bore=17.17cm off_right=-4.09 off_up=-1.96`, `wrist_to_target=0.00cm` (the wrist now lands **on** the target rather than 1.9 cm short of it), `shoulder_to_target=50.60` against `reach=51.80`. Screenshot `Saved/tmp/nudge_a88.png`. **Point 5 — R-88 recorded.** `-NoLoadingScreen` was already written down for the whole-suite command (`PLAYTEST_COMMANDS.md` §7) but not for running a **single** test, which is how the hand-IK work actually runs them. Both forms are now there, with the correct test path (`SouthernSpear.Network.Gameplay.TwoPlayerAuthoritySmoke`, verified against this run's log — not `LyraGame`) and the four `ViewportOverlayWidget` ensures named. **Tests.** 63/63 `Result={Success}`, 0 failures, with the change. `test_adfrc_grip`, `test_adfrc_reload`, `test_rtm_rigs`, `test_architecture_guard`, `test_check_unity_names`: 0 failures each. `validate_architecture.py`: PASS. New pure assertions in `HandIK.GripHold` cover `ApplyGripNudge` (a zero nudge is the identity, the three axes are independent, and a hold is not left stale by one). ### FILES `Plugins/SouthernSpearLyraBridge/Source/SouthernSpearLyraBridge/{Public/SSHandIKMeshComponent.h, Private/SSHandIKMeshComponent.cpp, Private/SSHandIKProbeSubsystem.cpp, Private/Tests/SSHandIKTests.cpp}`, `Config/DefaultGame.ini`, `Docs/PLAYTEST_COMMANDS.md`, evidence `Docs/evidence/handik_hold/SS_hold_frame.txt` (points 9–11). ## Session 079 — 2026-09-29 — The two "shadows" were the checker reading a name's own assignment as a declaration `Tools/check_unity_names.py` reported two shadows in `SSHandIKProbeSubsystem.cpp`: ``` shadow GBodyFrameInWeapon line 692 shadow GHaveBodyFrame line 693 ``` **Both are false positives, and the C++ was never at fault.** Lines 692-693 are assignments, not declarations: ```cpp GBodyFrameInWeapon = InWeapon; // file-scope FProbeHandFrame, declared line 244 GHaveBodyFrame = true; // file-scope bool, declared line 245 ``` A file-scope variable being assigned is not shadowed by anything. The name it is written under is the name it already has, in the same scope, so there is nothing to hide and no C4459 to raise. The build succeeded before this report and succeeded again afterwards with the probe source untouched (`Result: Succeeded`), which is the measurement that settles it. **The earlier cloud claim that C4459 would break the build was wrong.** It read a checker finding as a compiler diagnostic. A real C4459 is a compiler error and would have been in the build output; it was not, and it could not have been, because the two lines are not declarations. Recording this so the next reader does not spend a session renaming correct code. ### DEFECTS FOUND **`declared_name()` treated any identifier before a delimiter as a declaration** (`Tools/check_unity_names.py:162`). It cut the line at the first `[`, `=` or `;` and read whatever identifier sat in front of it as a declared name. Measured directly against the function: | Line | Read as | Correct? | |---|---|---| | `GBodyFrameInWeapon = InWeapon;` | `GBodyFrameInWeapon` | no — an assignment | | `GHaveBodyFrame = true;` | `GHaveBodyFrame` | no — an assignment | | `Alpha += 1;` | `Alpha` | no — a compound assignment | | `Foo[0] = 3;` | `Foo` | no — an element assignment | | `FString Text(GBodyFrameInWeapon);` | `None` | yes — it is a call | A C++ declaration is a **type and a name**: two identifiers before the delimiter. A lone identifier is a use of something declared elsewhere. Found by reading the two flagged lines instead of renaming them. ### COMPLETED **The checker.** One guard after the `if not names: return None` early-out: ```python if len(names) < 2 and delimiter != "(": return None ``` Function definitions (`Type Name(...)`) are kept by the `delimiter != "("` arm, so the anonymous-namespace check the shadow rule relies on is unchanged. Documented in `declared_name()`'s docstring with the three assignment forms it now rejects and the `GBodyFrameInWeapon` case that motivated it. **The regression fixture.** `Tools/test_check_unity_names.py` gains an `ETA` fixture — file-scope `FVector GFrame;` and `bool GHaveFrame = false;` in an anonymous namespace, then a function that only ever assigns them (`=`, `+=`, `[0] =`). The new assertion *"assigning to a file-scope name is not a shadow of it"* fails against the old checker and passes against the new one, and the existing *"only the two real shadows are reported"* assertion still holds at exactly 2, so the guard did not blunt the check it exists to perform. **The probe source is unchanged.** `GBodyFrameInWeapon` / `GHaveBodyFrame` keep their names. The fix belongs in the tool that was wrong, not in the code it misread. ### TESTING | Command | Exit | Result | |---|---|---| | `python Tools/check_unity_names.py` | 0 | `PASS - no name clash across 9 module(s).` | | `python Tools/test_check_unity_names.py` | 0 | 13 PASS, `0 failure(s)`, including the new assignment case | | `python Tools/validate_architecture.py` | 0 | PASS (pre-existing SS010 note for `SSFonts.h`, R-75) | | `Build.bat SouthernSpearEditor Win64 Development -WaitMutex` | 0 | `Result: Succeeded` (no C++ change, so a no-op link) | | `UnrealEditor-Cmd.exe ... -nullrhi -unattended -nosplash -nosound -NoLoadingScreen -ExecCmds="Automation RunTests SouthernSpear;Quit"` | 0 | **67 passed, 0 failed** (63 + 4 casualty) | ### FILES `Tools/check_unity_names.py`, `Tools/test_check_unity_names.py`, `Docs/CHANGELOG.md`. ### ASSETS None. Two loading-screen PNGs (`Docs/images/loadingscreens/dryriver.png`, `redgum.png`) are present in the tree and LFS-matched but still untracked; they are left as found. ### RISKS None new. The real defect here was in the guard, and the lesson generalises: every finding this checker produces is a hypothesis until the flagged line has been read. A finding that asks for a rename in code that already builds should be opened before it is acted on. ### NEXT ACTION Read the flagged line first. The checker's output is a pointer at a place to look, not a verdict on it. ## Session 080 — 2026-09-29 — A callsign field on the Interface tab (cloud; uncompiled) ### COMPLETED - **Callsign setting.** Scoreboards showed the platform name ("hurleym-CB9A…") because the only way to set a callsign was the `ss.Callsign` console command. The Settings screen's INTERFACE tab now has CALLSIGN: a text field, SET, and feedback (the rule in words on entry; "Callsign set to X…" or "Not accepted…" after). It shows only once a service record has loaded. - **Layering respected (SS001).** UI can't depend on Progression, so Core's `USSLocalProfileState` gained `CallsignSetter` and `RequestCallsign()`; `USSPlayerProfileSubsystem` installs the setter in `Initialize` and clears it in the new `Deinitialize` (the lambda captures `this`). The rule itself is still `FSSProgressionRules::IsValidCallsign`; the field calls `SetCallsign`, which saves, publishes and reports the name to the server (`ChangeName`), so an in-match change shows straight away. ### FILES CHANGED `SSLocalProfileState.h` (Core), `SSProgressionSubsystems.h/.cpp`, `SSSettingsWidget.h/.cpp`, `Docs/CHANGELOG.md`, the handover. ### TESTING - `python Tools/validate_architecture.py` → exit 0. `python Tools/check_unity_names.py` → exit 0 (after the Session 079 checker fix). - NOT RUN: the editor build; the field in game. **Uncompiled.** Watch: `UEditableTextBox` styling on the dark panel (foreground set to Sand100; the box's own background is the engine default), and that the six-word INTERFACE page still fits. ### ASSETS None. ### RISKS - R-91: the callsign is local and unauthenticated by design (R-53). Anyone can pick any valid name, including another player's. Fine for the offline-with-bots slice; a real name policy belongs with real accounts. ### DEFECTS FOUND None new. ### NEXT ACTION Casualty step 3 from `Docs/evidence/casualty_lyra_hooks.md` once it is committed (it is still untracked on the producer's machine). --- ## Session 081 — 2026-09-30 — Cloud's callsign field compiles first try; the loading art and the Lyra research are committed Pulled main first thing: `7da97511..6bdfb4c8`, fast-forward, no conflicts. The cloud's Session 080 added the CALLSIGN field (`SSLocalProfileState.h` gains `CallsignSetter`/`RequestCallsign`/`CallsignRule`; Progression wires `SetCallsign` into the setter and clears it in a new `Deinitialize`; `SSSettingsWidget.cpp` builds the Interface-tab block). Its commit message says "uncompiled". **It compiles as committed.** `Build.bat SouthernSpearEditor Win64 Development -WaitMutex`: exit 0, 10 actions (Core, Progression, UI each compile and link), `Result: Succeeded`, 37 s. This session changed nothing in the callsign code — there was nothing to fix. Full suite afterwards with `-NoLoadingScreen`: exit 0, `**** TEST COMPLETE. EXIT CODE: 0 ****`, **passed=67, failed=0** (`Saved/Logs/SS_suite_callsign.log`). ### COMPLETED - Committed last session's untracked work, LFS first: `git lfs push origin main` before the commit and again after it, then `GIT_LFS_SKIP_PUSH=1 git push`. The four binaries — two paintings, two imported textures — are `filter: lfs` per `git check-attr`. - `Content/EuropeanBeech/` (7 GB foliage, not this session's) left untracked and untouched. ### TESTING | Command | Exit | Result | | --- | --- | --- | | `git fetch origin` | 0 | `7da97511..6bdfb4c8 main`, 0 ahead / 1 behind | | `git merge origin/main --no-edit` | 0 | Fast-forward, 7 files +132/−2, no conflicts | | `Build.bat SouthernSpearEditor Win64 Development -Project=... -WaitMutex` | 0 | 10/10 actions, `Result: Succeeded` | | `UnrealEditor-Cmd ... -nullrhi -NoLoadingScreen "Automation RunTests SouthernSpear;Quit"` | 0 | `TEST COMPLETE. EXIT CODE: 0`; 67 `Result={Success}`, 0 `Result={Fail}` | ### FILES - `Docs/evidence/casualty_lyra_hooks.md` — Lyra death-chain and interact research, 301 lines, every file:line verified last session. - `Docs/evidence/ui_session078/` — six 1080p captures (front end, 3+2 card crop, rules-caption crop, three loading screens). - `Docs/images/loadingscreens/dryriver.png`, `redgum.png` — the two paintings (3.6 / 3.8 MB, LFS). - `Plugins/SouthernSpearUI/Content/Textures/T_SS_Load_DryRiver.uasset`, `T_SS_Load_RedGum.uasset` — imported textures (LFS). - `Docs/CHANGELOG.md` — this entry. ### RISKS - The Interface page grew by four widgets (label, 300 px box + SET button row, feedback line). Whether the six-item page still fits 1080p without clipping is not yet measured — that, the live typing behaviour and the scoreboard name are this session's next measurements. ### NEXT ACTION In game, Settings > INTERFACE: capture the CALLSIGN field; type `Dingo 2-1`, press SET, report the feedback text; boot a match and read the scoreboard name; try `x` and require the rule text in red. Evidence to `Docs/evidence/ui_session081/`. ## Session 082 — 2026-09-30 — Wandarra (M-009): the MOUT village built from three packs, nav bake pending ### COMPLETED The producer's 2026-09-29 decision — one big map from the MOUT kit and the other installed-but-unused packs — is now built. `L_Wandarra_01` exists (1,050,985 bytes), laid out by a new single-source spec and placed by a new Dry River-pattern pipeline: - **`Tools/Common/wandarra_spec.py`** — pure-Python layout spec (no unreal): 300 × 300 m site, two crossing streets (7 m + 1.5 m footpaths), 11 building rows, church assembly, 35 furniture rows, 5 fence runs with gate gaps, 10 cars, 6 tree rows, two deployments, three objectives, PROTECTED zones and `run_clears_protected` (corner + Liang–Barsky interior check — the corner-only version first written let a segment pass through a zone interior untouched, caught by unit check before any editor run). - **`Tools/Unreal/build_wandarra_level.py`** — level pass. `LevelEditorSubsystem.new_level` (the proven call), ground slab under `MI_SS_WorldGround_Wandarra` (Ravenshoe gravel on `M_SS_WorldGroundVT`, world-mapped 3 m tile), 11 vendor building Blueprints, 17 ChurchKit parts (nave walls/windows/roof, west tower + cross, east door — the vendor set has no church), 168 fence segments, 35 props, 10 RustyCars wrecks, 38 EuropeanBeech SimpleWind trees (seeded jitter), two tagged deployments (TeamOne depot NW, TeamTwo green SE), three sequential objectives with 900 cm discs, director, NavMeshBoundsVolume (75/75/15 — the ×4 reload factor from R-10 pre-accounted), RecastNavMesh saved into the map, WorldSettings experience set, light_dryriver lighting rig. - **`Tools/Unreal/build_wandarra_nav.py`** — nav pass that verifies the *designed* Depot→A→B→C→Green chain and repairs only what cannot walk it (the spec is the layout authority, not a computed path). BUILDPATHS ×1 per run (a second issue access-violates on this machine). - **`Tools/Unreal/light_wandarra.py`** — idempotent `SS_Light_*` re-light pass, as `light_dryriver.py`. - **Name**: Wandarra — invented locality, "place of the crow", same bird-name convention as Ravenshoe. No real locality is reproduced; `MAPS_TRAININGRANGE.md` §6's three blockers (author, name, layout) are all resolved, so **M-009 opened**. - **Register/audit/licence updates**: `ASSET_REGISTER.md` M-009 row; MOUT kit and RustyCars rows `CANDIDATE`/`NOT_USED` → `IN_USE`; new EuropeanBeech row (§4.9k); M-006 note updated (village committed beside it, range question stays the producer's). `LICENCE_REGISTER.md`: MOUT row in use, RustyCars row in use, new beech row — all three recorded with the L-0016c unverified posture (neither Vault pack wrote a `metadata` sidecar; the beech chunk carries a manifest only). `PROJECT_AUDIT.md`: R-67 first-load half measured, new **R-89** (ADR-016 look check deferred, not passed) and **R-90** (vendor building footprints unmeasured against the 1 m spec grid). - **`.gitignore`**: `Content/EuropeanBeech/` added — the 7 GB pack was installed but unignored (would have been committable; ADR-021/R-14 class defect caught on session-open status). ### FILES CHANGED - `Tools/Common/wandarra_spec.py`, `Tools/Unreal/build_wandarra_level.py`, `Tools/Unreal/build_wandarra_nav.py`, `Tools/Unreal/light_wandarra.py` — new (Class F original). - `Content/Maps/L_Wandarra_01.umap`, `Content/Art/Environment/Fab/MI_SS_WorldGround_Wandarra.uasset` — new build products (LFS). - `Docs/MAPS_WANDARRA.md`, `Docs/evidence/S082_wandarra_level_report.json`, `Docs/evidence/S082_wandarra_nav_report.json` — new. - `Docs/ASSET_REGISTER.md`, `Docs/LICENCE_REGISTER.md`, `Docs/PROJECT_AUDIT.md`, `Docs/CHANGELOG.md`, `.gitignore` — modified. - `Content/MOUT_Civilian/`, `Content/RustyCarsFree/`, `Content/EuropeanBeech/` — raw packs, git-ignored, referenced in place (ADR-021). ### TESTING | # | Command | Exit | Result | |---|---|---|---| | 1 | `python -m py_compile Tools/Common/wandarra_spec.py Tools/Unreal/build_wandarra_*.py Tools/Unreal/light_wandarra.py` | 0 | all four compile | | 2 | spec unit checks (`seg_point_distance`, protected-zone blocking, street clearance) | 0 | 3/3 pass after the Liang–Barsky fix | | 3 | `UnrealEditor-Cmd … -ExecutePythonScript=…build_wandarra_level.py` (`Saved/Logs/SS_wandarra_level.log`) | 0 | report `ok: true`, 12/12 steps, **0 missing assets, 0 Python errors, 0 fatals** (`Docs/evidence/S082_wandarra_level_report.json`) | | 4 | `UnrealEditor-Cmd … -ExecutePythonScript=…build_wandarra_nav.py` | 0 | **expected fail recorded honestly**: 0/3721 grid points on navmesh — headless BUILDPATHS is a no-op under -nullrhi (R-82) (`Docs/evidence/S082_wandarra_nav_report.json`) | | 5 | `layout_spawns.py`, `audit_map_playability.py`, bot match | — | **NOT RUN** — all three need the navmesh bake | ### ASSETS - Created: `L_Wandarra_01` (map), `MI_SS_WorldGround_Wandarra` (material instance). - Referenced in place, unmodified: MOUT kit (4.26 → first 5.8 load **silent**, R-67 evidence), RustyCars `SM_asset_00–04`, EuropeanBeech SimpleWind statics (5.1 native). - Licence rows updated as above; both Vault packs remain seller/AI-flag **unverified** (L-0016c). - Excluded by rule: MOUT `Demo/` character and weapon (ADR-020), Flag skeletal mesh, AwningKit Blueprints, beech PivotPainter/WIG variants, ivy meshes. ### RISKS - **R-89** (new): Wandarra's ADR-016 look check deferred, not passed — nothing has eyeballed the built map. - **R-90** (new): vendor building footprints unmeasured against the spec's 1 m grid; a building may overlap a footpath or fence until bounds are dumped. - R-67 updated: first 5.8 load of the 4.26 kit measured silent; render-cost half still open — and the map now *depends* on the kit, so an upconversion failure is no longer free. - R-82 unchanged: the attended bake is the only way this map gets a navmesh. ### DEFECTS FOUND - The spec's first `run_clears_protected` checked zone corners and endpoints only; a segment could pass through a zone interior without touching a corner. Found by the pre-editor unit check (through-objective-B test returned False-blocked incorrectly); fixed with Liang–Barsky. - `Content/EuropeanBeech/` was installed without a `.gitignore` row — 7 GB of raw pack sat untracked but committable. Found on session-open `git status`; fixed in this change. ### NEXT ACTION Attend the editor: open `L_Wandarra_01`, Build ▸ Build Paths, save (the R-82 attended bake), then re-run `build_wandarra_nav.py` to verify the designed legs and run `layout_spawns.py` with `SS_MAPS=L_Wandarra_01`. Evidence to `Docs/evidence/session082/`. ## Session 083 — 2026-09-30 — The callsign field verified live: Dingo 2-1 accepted, rejected in red, and on the scoreboard The Interface-tab field the cloud added in Session 080 and this shop compiled in Session 081 was exercised in the real front end, end to end, with captures at every step (`Docs/evidence/ui_session083/`, 1920x1080). **What was measured.** SETTINGS opens from the top bar; the INTERFACE tab activates; the CALLSIGN block renders (label, 300 px box with the Sand100 text and engine-default field, SET, rule line). Typing `Dingo 2-1` over a Ctrl+A selection and pressing SET turns the rule line **brass** — measured 855 pixels matching Brass300 `D6C49D` (first sampled pixel exactly `(214,196,157)`) — with the "Callsign set to ..." text. Typing `x` and pressing SET turns it **red** — 758 pixels matching Opfor300 `C98A7C`, zero brass/sage contamination — the "Not accepted." reject. Re-setting the good name returns the brass accept. The game log carries the server-side receipt twice: `LogSSProgression: Service profile: Dingo 2-1 level 2.` (`Saved/Logs/SS_load_s081h.log:2224-2225`). **Scoreboard.** DEPLOY → `Browse: "/Game/Maps/L_RedGum_01?NumBots=8"` → spawned; holding Tab shows the local row highlighted at rank 2, named **Dingo 2-1**, ping 0 ms, between the bots (`06_scoreboard_dingo_2_1.png`). The Session 080 comment "shows on the scoreboard from the next match" is true, and the in-match ChangeName path (`SSServiceRelay::ServerReportProfile`) makes it immediate. **The six-item Interface page fits.** Measured text bands on the 1080p client: FRAME RATE COUNTER y≈359-373, DEVELOPER MESSAGES y≈417-431, CALLSIGN label y≈475-489, box row y≈512-553, rule line y≈570-584; BACK/APPLY at y≈787-843; panel bottom ≈880. Roughly 300 px of spare room below the rule line — no clipping, no scroll. ### DEFECTS FOUND None in the callsign feature. Three environment notes, none actionable here: the engine's AI-toolset Python spams `AttributeError` at boot (engine-side, pre-existing); one D3D12 GPU device loss ended a match mid-session (`D3D12Util.TerminateOnGPUCrash`, crash dir `UECC-Windows-038C...` — the game was closed cleanly afterwards); automation clicks on tab labels only registered a few pixels below centre while every manual click worked, so the driver now finds live hit pixels by hover-scanning (capture-side quirk, recorded in `Saved/tmp`, not a product bug). ### TESTING | Command | Exit | Result | | --- | --- | --- | | `python Saved/tmp/loadshots/csign/boot.py s081h` | 0 | Front end up, client 1920x1080 at (8,31) | | SETTINGS click, INTERFACE tab (hover-verified) | — | Underline moved to x 1082-1176; callsign block present | | Ctrl+A, `Dingo 2-1`, SET | — | Brass accept (855 px `D6C49D`); log line :2224 | | Ctrl+A, `x`, SET | — | Red reject (758 px `C98A7C`, 0 brass) | | Ctrl+A, `Dingo 2-1`, SET (restore) | — | Brass accept; log line :2225 | | DEPLOY → hold TAB in match | — | `Browse: "/Game/Maps/L_RedGum_01?NumBots=8"`; scoreboard shows Dingo 2-1 | ### FILES - `Docs/evidence/ui_session083/01..06` — field, typed states, brass accept, red reject, scoreboard. - `Docs/CHANGELOG.md` — this entry. Capture drivers stay in gitignored `Saved/tmp/loadshots/csign/`. ### NEXT ACTION The callsign feature is verified; the Session 080 handover's "watch" items are all closed. Remaining from the handover: Wandarra's attended nav bake (Session 082), and the casualty-care work still awaits a live test of the Lyra death-chain hooks documented in `Docs/evidence/casualty_lyra_hooks.md`. ## Session 084 — 2026-09-30 — Wandarra dressed in design: awnings and gate doors queued behind a yaw-defect rebuild; vendor doors ruled scenery (ADR-041) ### COMPLETED The producer's "bring the awnings in" direction, executed as design + code + a recorded decision. The interactive editor was open on the project all session (one-writer rule), so **no headless pass ran against the map** — everything below is written, checked, and queued behind one attended sequence: - **`Tools/Common/wandarra_spec.py` extended**: `AWNING_BPS`/`AWNING_ROWS` (6 verandahs from the four `BP_GovernmentAwning_*` types — civic face ×2, west row main-street footpath, east row park side, cross-street store, depot-lane corner) and `DOOR_BPS`/`DOOR_ROWS` (2 scenery doors: depot side gate, green picket gate). Depot gate gaps narrowed to **2.4 m personnel width**; the main gate deliberately keeps an **open** gap, no door — see the defect below. - **`Tools/Unreal/dress_wandarra_awnings.py`** — new pass: idempotent `SS_Dress_*` placement, ground-snapped doors, and the **R-90 measurement** riding along: dumps all 11 vendor building boxes to site coordinates, reports road-corridor overlaps, and pushes each awning out of its building's *measured* box by 40 cm (spec row = intent, footprint = evidence). - **ADR-041 recorded** (`DECISION_LOG.md`): the vendor interactable door Blueprints are **scenery, not gameplay**. Evidence: the shipped BPs carry embedded compile errors (`Could not find a variable named "Left Door Rotation" in 'BP_GreenDoors_Interactable_C'`, read from the compiled asset) and client-local timelines; Lyra's interact input is already assigned to casualty care (ADR-040). Any future gameplay door is a server-owned `USSDoorComponent` on a project-owned actor referencing the vendor *meshes* — the vendor BPs stay unreferenced by gameplay code forever. - **Yaw defect found and fixed (map pending rebuild)**: the site→Unreal rotation shipped as `-yaw`; the correct transform is **`yaw − 90`** (site north = Unreal −Y = Unreal yaw −90; site east = +X = 0; site south = +Y = 90). Every rotated actor in the built `L_Wandarra_01` — building facings, church walls, fence runs, furniture — sits one cardinal direction off, the exact silent-failure class MAPS_DRYRIVER §11.2 documents. Found while deriving the awning push-out vector, not by an editor look. `build_wandarra_level.py` fixed; the on-disk map keeps the defective rotations until the level pass re-runs, and **the nav bake must come after that rebuild**, not before. - **Door-seals-spawn flaw caught pre-placement**: a closed door actor bakes into the navmesh as a blocker, and TeamOne spawns inside the depot compound — a door on the main gate would have made the team's only walkable exit depend on an unwired scenery door. Main gate = open 2.4 m gap; doors only at the side gate and the green gate (no one spawns inside that yard). - Docs: `MAPS_WANDARRA.md` §3.1 (dressings), corrected transform section, R-90 remedy line; ADR-041 in `DECISION_LOG.md`. ### FILES CHANGED - `Tools/Common/wandarra_spec.py` — modified (awning/door tables, gate gaps, main-gate rule). - `Tools/Unreal/build_wandarra_level.py` — modified (yaw transform corrected). - `Tools/Unreal/dress_wandarra_awnings.py` — new (Class F original). - `Docs/DECISION_LOG.md` (ADR-041), `Docs/MAPS_WANDARRA.md`, `Docs/CHANGELOG.md` — modified. ### TESTING | # | Command | Exit | Result | |---|---|---|---| | 1 | `python -m py_compile` (spec, builder, dressing pass) | 0 | compile clean | | 2 | spec table checks (row shapes, gap arithmetic, door/awning coordinates inside fence gaps) | 0 | pass — two coordinate errors caught and fixed pre-commit | | 3 | `dress_wandarra_awnings.py` headless run, level rebuild, R-90 bounds dump | — | **NOT RUN** — interactive editor open (one-writer rule); queued | | 4 | nav bake + verify, `layout_spawns.py`, playability audit | — | **NOT RUN** — still gated on the attended bake, which is now gated on the rebuild | ### ASSETS - Referenced in place, unmodified: `BP_GovernmentAwning_01a/b`, `02a/b`; `BP_WoodenDoor_Interactable` (2 placements). `BP_GlassDoors_Interactable` / `BP_GreenDoors_Interactable` / AwningKit loose meshes: **not used** (glass/green BPs broken; meshes unneeded — ADR-041 consequences). - No new project assets created this session. ### RISKS - **Yaw defect (new, in code fixed / in map pending):** until the rebuild, every rotation in `L_Wandarra_01` is one cardinal off; any measurement taken on the current map is invalid. The queued sequence must be rebuild → dress → bake → verify → spawns. - R-90's dump is now wired into the dressing pass but has produced no numbers yet. - R-89 (ADR-016 look check) unchanged — still no eyes on the map; R-82 unchanged. ### DEFECTS FOUND - **Yaw transform** (above) — found by derivation while coding the awning push-out, not by inspection; the class of failure MAPS_DRYRIVER §11.2 warns about, caught at the second build instead of the first because the spec's own self-checks verify positions, not rotations. - Awning yaw sign error in the first spec draft (a west-row verandah faced away from the street) and three door coordinates that did not sit in their gate gaps — all caught by re-deriving the geometry before commit, none by tooling; the spec should get a rotation-aware self-check. - **Door-seals-spawn** (above) — found by asking "what does the navmesh bake see?" before placing, not after; recorded in ADR-041 so the main-gate rule survives future edits. ### NEXT ACTION Attend the editor once the current session closes it: re-run `build_wandarra_level.py` (applies the yaw fix and the 2.4 m gate gaps), run `dress_wandarra_awnings.py` for the R-90 dump, correct spec rows if the dump demands, then Build ▸ Build Paths, `build_wandarra_nav.py`, `layout_spawns.py` with `SS_MAPS=L_Wandarra_01`. Evidence to `Docs/evidence/wandarra_bake/`. ## Session 085 — 2026-09-30 — The kangaroo easter egg is placed (and one floated in a tree); Dry River gets its overhaul defect list The producer's two new model sets — ghost gum (ENV-003) and kangaroo (ENV-004, both L-0024, provenance initially unverified and recorded as such with R-90a) — arrived in `Content/ghostgum/` and `Content/kangaroo/`. The registers were updated in the same change as the sources landed (`d7c2d93b`). **Later the same day the producer confirmed all models downloaded for or used in the project are free and cleared for game use; R-90a is closed and L-0024 reclassified to Class A under producer risk acceptance.** **The easter egg.** `Tools/Unreal/dress_kangaroo_easteregg.py` imports the kangaroo FBX (legacy FBX path with `combine_meshes` — Interchange split it into 25 pieces first), authors `MI_SS_Kangaroo` as an instance of the project's `M_SS_ScanPBR` master wired to the vendor's own body maps, derives a uniform scale from the mesh's real bounds (63.4 × 24.9 × 53.3 cm source → 2.3668 to stand 1.5 m), and places **two kangaroos per map** on Red Gum Station and Dry River under named dressing trees — terrain-anchored, collision off, labels `SS_EasterEgg_Kangaroo_1/2`, idempotent per label, each map saved. **The defect the producer caught, and what it exposed.** First placement floated two animals in the canopy: the script's downward visibility trace at the trunk hit the tree's own collision top and used *that* z. The producer's screenshot (`Docs/evidence/session084_dryriver/DR_kangaroo_in_canopy.png`) is what found it — the placement script's report said ok. Fix: anchor to the tree's own pivot z (dressing passes spawn trees at terrain height) and re-place all four; re-run clean, 4/4, maps saved. **Process rule recorded in the map doc: no placement pass is done without in-engine captures of the placed actors; script reports are not verification.** **Dry River overhaul.** The producer walked the map and recorded seven defects — front-facing-only assets, untextured assets, the windmill caught in trees, random prop placement, no water in the creek, sparse small foliage versus Red Gum, and a general need for more density. Recorded as D-DR-01..07 with evidence in `Docs/evidence/session084_dryriver/` and standing in §11 of `Docs/MAPS_DRYRIVER.md` as the overhaul's spec. The ghost gum model is imported as part of that overhaul (canopy variety), per the producer's direction. ### TESTING | Command | Exit | Result | | --- | --- | --- | | `UnrealEditor-Cmd -run=pythonscript -Script=.../dress_kangaroo_easteregg.py` | 0 | `ok: true`; textures 9, mesh 1, material 1; Red Gum 2/2, Dry River 2/2; both maps saved | | Re-run after the pivot-z fix (idempotent) | 0 | 0 warnings, 4/4 replaced, both maps saved | | Interchange first import | — | produced 25 piece meshes; purged, re-imported combined (legacy FBX path) | ### FILES - `Tools/Unreal/dress_kangaroo_easteregg.py` — the import + placement pass (idempotent, bounds-derived scale). - `Content/Art/Environment/Kangaroo/` — SM_Kangaroo, MI_SS_Kangaroo, 9 textures. - `Content/Maps/L_RedGum_01.umap`, `Content/Maps/L_DryRiver_01.umap` — 2 easter-egg actors each. - `Docs/ASSET_REGISTER.md` §4.9k, `Docs/LICENCE_REGISTER.md` L-0024 — in the earlier same-session commit. - `Docs/MAPS_DRYRIVER.md` §11 — D-DR-01..07 overhaul spec; `Docs/evidence/session084_dryriver/` — producer evidence. ### RISKS - ~~R-90a~~ CLOSED same day: the producer confirms all downloaded/used models are free for game use (L-0024 reclassified Class A; the missing vendor metadata remains a credits-record note only). - The easter-egg placement is script-verified and pivot-snapped but **not yet re-verified with in-engine captures after the fix** — first item for the overhaul session, along with every D-DR fix. ### NEXT ACTION The Dry River overhaul per §11: ghost gum import, water in the creek, ground-foliage density pass, prop re-composition, textured/backface audit — each fixed with before/after in-engine captures. ## Session 086 — 2026-09-30 — Wandarra rebuilt and dressed; Recast export policy narrowed to walkable geometry ### COMPLETED - Rebuilt `L_Wandarra_01` from the layout spec with the corrected site-to-Unreal yaw (`yaw - 90`) and current gate layout. The level report is `ok: true`: 11 buildings, 17 church parts, 183 fence segments, 35 props, 10 cars and 38 trees; no missing assets or Python errors. - Ran the idempotent dressing/bounds pass. It measured all 11 building bounds, found **0 road-corridor overlaps**, placed **6/6 awnings** (five pushed clear of the measured building boxes, max push 2.7 m), and **2/2 scenery doors**. The report is `ok: true`, with no missing assets or errors. - Tree and car render-mesh components are excluded from Recast geometry export at the placed-actor level (38 trees, 10 cars); world collision remains enabled and no vendor mesh assets are modified. The level builder applies the same policy to newly generated actors. - Removed headless `BUILDPATHS` from `build_wandarra_nav.py`. It now checks saved actor nav policy and saved tile coverage only; this machine's null-RHI bake is a measured no-op (R-82), so a failed/empty report directs the operator to the attended editor bake instead of claiming success or triggering another asynchronous build. - Updated `MAPS_WANDARRA.md` and R-90 in `PROJECT_AUDIT.md` with the measured road clearance, rebuild state, nav-filtering policy and honest outstanding checks. ### FILES CHANGED - `Content/Maps/L_Wandarra_01.umap` — rebuilt and dressed; no Dry River map or assets included. - `Tools/Unreal/build_wandarra_level.py`, `dress_wandarra_awnings.py`, `build_wandarra_nav.py` — actor-level nav exclusions, measured dressing and saved-nav verification. - `Docs/MAPS_WANDARRA.md`, `Docs/PROJECT_AUDIT.md`, `Docs/CHANGELOG.md` — implementation status and risk update. - `Docs/DECISION_LOG.md`, `Docs/PLAYER_MODEL_PLAN.md`, `Docs/HANDOVER_CLAUDE_CLOUD.md` — corrected the G3/Quantum decision framing and documented the live head-assembly mismatch discovered while returning focus to the character model. ### TESTING | Command / evidence | Result | |---|---| | Headless level build report | **PASS** — `Build/wandarra_level.json`: `ok: true`, 12/12 steps, no missing assets/errors | | Headless dressing report | **PASS** — `Build/wandarra_dressing.json`: `ok: true`, 11 bounds, 0 road overlaps, 6 awnings, 2 doors, 38 tree + 10 car nav exclusions | | `python -m py_compile Tools/Common/wandarra_spec.py Tools/Unreal/build_wandarra_level.py Tools/Unreal/dress_wandarra_awnings.py Tools/Unreal/build_wandarra_nav.py Tools/Unreal/layout_spawns.py` | **PASS** (exit 0) | | Saved nav verification | **BLOCKED as expected** — `Build/wandarra_nav.json` reports 0/3721 projected points; attended-editor Build ▸ Build Paths and save still required (R-82) | | `layout_spawns.py`, in-engine visual inspection, playability audit | **NOT RUN** — depend on the attended nav bake and rendered editor validation | ### ASSETS - `L_Wandarra_01.umap` is the only generated project asset changed. Vendor pack assets are referenced in place and unmodified. No Dry River asset is included. ### DEFECTS FOUND - The MOUT building Blueprints measured clear of the defined road corridors (0 overlaps); the headless report does not prove footpath/fence clearance or the visual read. - Vendor tree/car render collision caused high-triangle Recast-export warnings. The fix is per-instance nav exclusion with world collision retained; vendor assets remain unchanged. - The old nav pass attempted `BUILDPATHS` despite measured null-RHI no-op behavior. It is now a saved-nav verifier only. ### RISKS / NEXT ACTION R-82 remains open: perform Build ▸ Build Paths in the editor, save, then rerun the read-only nav verifier and spawn layout. R-89's visual/look check and R-90's footpath/fence clearance are also still open; the measured road-overlap result alone does not close them. Other-agent Dry River changes were intentionally left out. **Next action:** attend the editor for Wandarra's Build Paths bake and save, then verify saved coverage and planned routes. The character-model audit resumed in Session 086 as a documentation correction only: `setup_soldiers.py` still puts the Modern Insurgent 7 head on the ADFRC G3 uniform/gear, and the installed ADFRC source set has no character body/head mesh. ADR-036 no longer overstates the head as a same-pack G3 asset or Quantum as retired; see `PLAYER_MODEL_PLAN.md`. No soldier asset or runtime configuration was changed. The next model work is candidate inventory and a safe same-condition comparison; do not switch skeleton/body before the producer's choice. ## Session 087 — 2026-09-30 — The Quantum body is live for the friendly look (ADR-042): runtime retarget, preview, first-person fixes ### COMPLETED - Producer decision executed: the friendly soldier is now the **Quantum character on its own skeleton** (ADR-042), ending the ADR-036–039 comparison gate. The pawn's gameplay skeleton stays Manny (hit zones, hand IK, sockets untouched); the Quantum modules are retargeted per tick from the pawn mesh's evaluated pose. Opposing MAF look unchanged. The producer's in-game screenshots confirm the Quantum body rendering live with the ADFRC vest and helmet in both the class-select preview and third person. - `ASSCharacterPartActor` rewritten to a minimal retarget path: per-mesh bone maps built lazily with the prototype's rest-pose tolerances (6% height / 30°), pose read from the pawn mesh's component-space transforms, root held at identity, local rotations composed parent-before-child into component space. Rejected the earlier WIP's separate pose-driver component (it could not see the pawn's animation) and its out-of-bounds map indexing. Verified in a headless live run: all 4 pawns spawned `4 retarget + 2 leader + 4 opposing part(s), leader CharacterMesh0, retarget=1`, zero script errors. - `FriendlyLeaderPoseParts` added to the part actor for the Manny-rigged ADFRC vest and helmet layered over the Quantum body; `setup_soldiers.py` now writes the Quantum configuration and `Build/soldiers_setup.json` reports `ok: true` (4 retarget, 2 leader-pose, 4 opposing). - First-person fixes for the skinned-part world (producer screenshots showed the local player's own Quantum head filling the camera in body view, and an arms-view dark mass): body view now hides the head bone on every attached skinned part (the Quantum head is its own component), the arms view-model path hides every attached *skinned* mesh (`USkinnedMeshComponent`, not the skeletal-only cast that missed the poseable Quantum modules), and body view un-hides parts when switching view models. - Class-select preview mirrors the runtime split without a Team dependency (SS001): the Quantum modules self-animate with the pack's own idle, the vest/helmet leader-pose the invisible Manny, and a soldier configured with mannequin parts only still previews the old way. Confirmed live by the producer's screenshot. - ADR-042 written; `PLAYER_MODEL_PLAN.md` §5 rewritten to record the decision (the historical assembly mismatch is configuration history now); this changelog entry. ### FILES CHANGED - `Plugins/SouthernSpearTeam/.../SSCharacterPartActor.{h,cpp}` — retarget implementation, `FriendlyLeaderPoseParts`, cached per-mesh bone maps. `AnimationCore` dependency removed again (the retarget needs only `ReferenceSkeleton.h`, which Engine already exports); Build.cs now byte-identical to HEAD, and the final build after removal is green (15 s, 0 errors). - `Plugins/SouthernSpearLyraBridge/.../SSFirstPersonSubsystem.cpp` — skinned-part-aware head hide, arms-view hiding and body-view restore. - `Plugins/SouthernSpearUI/.../SSClassSelectWidget.cpp` — Quantum preview (self-animated modules, leader-posed kit, legacy fallback). - `Tools/Unreal/setup_soldiers.py`, `Tools/Unreal/setup_character_textures.py` — Quantum friendly configuration; the texture pass no longer authors friendly overrides. - `Docs/DECISION_LOG.md` (ADR-042), `Docs/PLAYER_MODEL_PLAN.md`, `Docs/CHANGELOG.md`. ### TESTING | Check | Command | Result | |---|---|---| | Architecture guard | `python Tools/validate_architecture.py` | **PASS** (exit 0; the accepted SS010 note is pre-existing) | | Unity-name check | `python Tools/check_unity_names.py` | **PASS**, 9 modules, no clash | | Editor build | `Build.bat SouthernSpearEditor Win64 Development` | **PASS ×3** — zero errors, zero warnings (one intermediate failure per wrong engine API, each fixed) | | Automation suite | `UnrealEditor-Cmd … RunTests SouthernSpear` (with `-NoLoadingScreen`) | **67/67 `Result={Success}`, 0 fails** | | Soldier setup | `setup_soldiers.py` commandlet | `Build/soldiers_setup.json` `ok: true`; one prior run failed on the bool's Python name (`retarget_friendly_pose`, not `b_retarget_friendly_pose`) — fixed and re-run | | Live headless smoke | Dry River `?NumBots=4` `-game -nullrhi`, 150 s | `SSCharacterPart … 4 retarget + 2 leader + 4 opposing part(s) … retarget=1` ×4 pawns, 0 `LogScript` errors | | Producer visual | In-game screenshots 2026-09-30 | Quantum body + ADFRC vest/helmet render in class-select preview and third person; first-person defects reproduced by screenshots, C++ fix built **NOT yet captured in game** | | Quantum proof | `prove_quantum_retarget.py` | **NOT RUN** this session (editor lock); the proof targets the prototype stage's copy of the same arithmetic — the runtime path's own evidence above is the spawn log + screenshots | | Camo verification | `probe_quantum_material.py` + asset read-back | **ROOT CAUSE CONFIRMED**: mesh-asset material slots are **read-only from Python in 5.8** — `set_editor_property` on the materials array returns without error and without effect, and `set_material` does not exist on the asset. Session 058's `ok:true` was a silent no-op. Fixed by routing the camo through `FriendlyMaterialOverrides` (the same component-override mechanism as the MAF green uniform, ADR-004): `setup_quantum_proto.py` re-ran `ok: true` with a cleaned graph (`delete_all_material_expressions` after measuring 12 expressions from reruns), `setup_soldiers.py` re-ran `ok: true` writing the overrides, and the saved `B_SS_Soldier.uasset` greps for both `MI_SS_ADFRC_Camo_*` instances and all three Quantum + ADFRC part names. **NOT yet captured in game** — the next play session shows the camo | ### ASSETS - `B_SS_Soldier.uasset` reconfigured (Quantum friendly parts, leader-pose kit, retarget flag). - QuantumProto module meshes and materials unchanged on disk this session (their camo defect is Session 058's, resurfaced; the producer's own untracked Quantum content edits are preserved). ### DEFECTS FOUND - The WIP retarget from the previous session could never have worked: its pose driver was a separately animated skeletal component (reference pose, not the pawn's animation) and its bone map indexing read `FriendlyRetargetComponents[Num]` out of bounds. Found by re-reading the diff before building on it. - First person: attached **poseable** soldier parts escaped both first-person hiding paths (skeletal casts) — the local player wore their own head. Found by the producer's screenshots. - `setup_quantum_proto.py` claimed camo slots it never persisted (Session 058's `ok:true` vs. the blue shirt in game). Found by grepping the saved mesh assets for material names. - UE 5.8 Python booleans drop the `b` prefix (`retarget_friendly_pose`); found by running the commandlet, not by assuming. ### RISKS - **R-91:** CLOSED as a pipeline defect (root cause measured, scripts fixed and re-run, asset read-back verified). The in-game camo confirmation is pending the producer's next capture. - R-58 narrows (visible body no longer Manny-welded; the gameplay skeleton still is). - First-person fix is built but not yet producer-captured; the hide-bone call on poseable meshes is the one untested engine-behaviour claim in this session. ### NEXT ACTION Play a round on the new build: confirm the shirt and jeans render the ADFRC camo, the first-person view is clean (no head, no dark mass), and the vest/helmet and hands sit right in motion. Then set the friendly overrides read-back as a scripted check in a future session. ## Session 088 — 2026-09-30 — Dry River harvest overhaul: D-DR-01..07 fixed from assets already in the project ### COMPLETED - The producer's Dry River overhaul defect list (`MAPS_DRYRIVER.md` §11) is now worked end to end by one idempotent script, `Tools/Unreal/harvest_dryriver.py`. It removes only its own labels (`SS_Overhaul_*`, `SS_Gum_*`) before re-placing, so it never touches hand-placed dressing and can be re-run to iterate. **Report `ok: true`, 0 warnings, 0 errors, 42 actors placed, 100 own actors cleared on the re-run.** - **The harvest came entirely from assets already in this project** — nothing was downloaded. Scene Quarry Slate (L-0005) for the creek surface mesh and bed stones, RuralAustralia (L-0016) for the bush vocabulary, Namaqualand (Session 041) for searsia/rooibos/didelta shrubs, the four-flower pool, dead-quiver driftwood and stones, the ghost gum (ENV-003, L-0024) for the canopy, and one normal map out of the WaterPlane pack. - **D-DR-03 windmill caught in trees.** Measured root cause, not the impression: three `SS_RA_Tree` / `SS_RA_Cover_Tree` with 28–38 m bounds radii overlapped the mill at 24–33 m. They carry no trunk colliders, so deleting them is nav-safe. The mill is now framed by a **ring of 10 imported ghost gums at 10.5 / 13.5 / 16.5 m** — the classic outback windmill-in-gums shot — each with a hidden trunk collider so agents cannot walk through the trunk while the canopy stays collision-free. - **D-DR-05 no water.** Six water segments placed **trace-anchored**: the old placement used the `height()` formula, but the rendered creek terrain is excavated *below* that formula, so all six planes were buried in the ground. Every segment is now `bed_z + 12 cm`, and `water_audit` in the report records the water z, the traced bed z and the segment centre for all six. `M_SS_CreekWater` was authored from scratch (`Tools/Unreal/author_creek_water.py`) because the pack water reads as a glossy orange strip on red dirt. - **D-DR-06 / D-DR-07 density, measured.** The audit previously counted only `StaticMeshActor`s, so the new HISM scatter registered as zero. Fixed to count HISM **instances** through the same subobject path the placement uses: Dry River small-foliage **92 → 592** against Red Gum's 124, of which **500 are instanced pieces** (bed stones 90, bush A 150 / B 90 / C 60, flowers 110) plus 26 driftwood logs. Static-mesh actors 1173 → 1222. Nothing the producer placed by hand was removed. - **D-DR-02 raw grey.** The audit now scans for *visible* actors with empty/default material slots: **0**, with the 10 hidden trunk-collider helpers counted separately so they can never masquerade as a defect. The producer's **grey leafless gum** had the same root cause as the discs — the ghost gum mesh has two slots (`blinn5` trunk, `TH_Gum_Branch_Blinn` **foliage cards**), the branch slot was taking the bark instance, and the tree rendered grey and bare. Slots matching *branch*/*leaf* now take the alpha-masked leaf instance. **The producer confirms the gums are properly leafed.** - **Two placements were retired outright** rather than tuned, because both were wrong as assets: the QuarrySlate `SM_Qua_Sla_Patch_*` round discs (flat discs float on any slope, and their cream albedo fights the red dirt — the producer's "round discs floating" screenshot) and the QuarrySlate "European Spindle" bushes (European broadleaf, near-black; Dry River and Red Gum are both Namaqualand didelta, so they were the wrong continent as well as the wrong colour). Replaced with Namaqualand searsia. Bed stones and driftwood now carry the creek floor. - **Kangaroo, partially fixed.** The animals were sunk into the terrain; they are re-seated on a real trace (+6 cm) and the producer confirms they stand. `MI_SS_Kangaroo`'s texture overrides were found **empty** in the saved asset — the original egg script's parameter writes never persisted — so both maps are now bound by registry lookup and read back through `MaterialEditingLibrary` (`base=znzmoModel-1132355448-0277`, `norm=…-0278`). **The roo still renders clay-grey in game**, so this is not closed; see DEFECTS FOUND. - Nav verified after the new trunk colliders: `build_dryriver_nav.py` `ok: true`. ### FILES CHANGED - `Tools/Unreal/harvest_dryriver.py` *(new)* — the whole overhaul pass; `Tools/Unreal/author_creek_water.py` *(new)* — authors `M_SS_CreekWater`. - `Tools/Unreal/audit_dryriver_overhaul.py` *(new)* — D-DR-01..07 ground-truth audit, extended this session to count HISM instances, to separate hidden collision helpers from visible grey assets, and to restore the Red Gum comparison prefixes it had lost. - `Tools/Unreal/inventory_map_assets.py`, `probe_gum_params.py`, `probe_kanga_water.py`, `probe_mat_api.py`, `probe_sweep.py`, `probe_water_state.py` *(new)* — the diagnostic probes this work needed. Kept: each one documents an engine behaviour that is not obvious. - `Content/Maps/L_DryRiver_01.umap`, `Content/Art/Environment/Kangaroo/MI_SS_Kangaroo.uasset` *(modified)*. - `Content/Art/Environment/DryRiver/M_SS_CreekWater.uasset`, `Content/Art/Environment/Fab/TH_Complete_Full_Ghoast_Gum.uasset`, `…/MI_SS_GhostGum_{Trunk,Branch,Leaf}.uasset`, `Content/Art/Environment/Fab/GhostGum/*` *(new, imported from ENV-003)*, `Content/WaterPlane/Lake/Textures/T_MediumWaves_N.uasset` *(new — the single pack file the water material references; the other 33 files of that 140 MB pack stay local)*. - `Docs/MAPS_DRYRIVER.md` §11 statuses + new §11.1, `Docs/ASSET_REGISTER.md` §4.9h/ENV-003/§4.9l, `Docs/CHANGELOG.md`, `Docs/evidence/session088_dryriver_overhaul/*.json`, `Docs/evidence/ui_session088/*.png`. - **Not touched:** anything belonging to the concurrent character-model/Wandarra thread — the Quantum and ADFRC character assets, `SSCharacterPartActor`, the first-person and class-select C++, the Wandarra map, or `probe_quantum_material.py`. This entry also lands in a working tree that already held that thread's uncommitted Session 087 changelog section; both are additive and neither was rewritten. ### TESTING | Check | Command | Result | |---|---|---| | Overhaul pass | `UnrealEditor-Cmd … -run=pythonscript -Script=Tools/Unreal/harvest_dryriver.py` | **PASS** — `ok: true`, 0 warnings, 0 errors, 42 actors placed, 100 own actors removed on re-run | | Water material | `-Script=Tools/Unreal/author_creek_water.py` | **PASS** — `ok: true`, 1 warning (see DEFECTS FOUND) | | Defect audit | `-Script=Tools/Unreal/audit_dryriver_overhaul.py` | **PASS** — `ok: true`: 0 visible untextured actors (+10 collision-only), 0 windmill tree conflicts inside 12 m, creek bed 65–217 cm below bank, small-foliage DR 592 vs RG 124, 0 paper-thin candidates, 1222 actors | | Water height | report `water_audit` | **PASS** — all 6 segments at exactly `bed_z + 12.0 cm` (e.g. seg 3: water 10.6 / bed −1.4) | | Windmill clearance | report `windmill_clearance` | **PASS** — nearest canopy gap 2.74 m (gum ring; threshold 2.5 m), nearest other scenery 5.99 m (threshold 5.0 m) | | Kangaroo material | report `fix_kangaroos` | **PARTIAL** — MI reads back with both maps bound; the render is still grey. Not closed | | Nav | `-Script=Tools/Unreal/build_dryriver_nav.py` | **PASS** — `ok: true`; two non-gating steps fail on the documented R-10 reload quirk | | Producer visual | screenshots, 2026-09-30 | **MIXED** — gums confirmed leafed, kangaroos confirmed standing, discs and black bushes confirmed gone; **no water seen**, and the kangaroo is still grey | | In-engine captures | six-spot spectator/player capture pass × 6 | **PARTIAL** — 42 frames in `Docs/evidence/ui_session088/`; the final pass matches the current build. The water close-up has not been inspected for D-DR-05 | ### ASSETS - **ENV-003 ghost gum `VENDORED` → `IN_USE`**: mesh + 6 imported maps + 3 `MI_SS_GhostGum_*` instances; 10 placed on Dry River. Register row updated in the same commit. - **ENV-005** `M_SS_CreekWater` and **ENV-006** the gum trunk-collider helpers registered as Class F originals in the new §4.9l. - **WaterPlane pack registered** in §4.9h as a bookkeeping correction: it was installed in `Content/` with no register row and no `metadata` sidecar, so it is recorded as **unverified, never clear** (L-0016c posture), not Class A. One texture file is committed because the material references it. - L-0024 (ENV-003/004 provenance) is unchanged — producer risk acceptance stands. ### DEFECTS FOUND - **The creek water was buried because the placement formula was wrong, not the material.** `height()` is the *design* profile; the rendered terrain is excavated below it by up to 2.2 m. Every water plane placed on the formula was under the ground. Found by tracing the bed instead of trusting the spec function. - **The "windmill caught in trees" report was really crowding at 24–33 m** — no canopy actually overlapped the mill. Re-reading the producer's screenshot changed the fix from "delete trees" to "frame the mill", which is a better result and cheaper. Found by measuring bounds radii against the mill. - **The ghost gum's second material slot is its foliage**, not more bark, despite the name. Mapping bark to it produced the producer's grey leafless gum. Found by reading the mesh's slot names. - **An audit that counts actors silently reports HISM work as zero.** The density pass had landed 500 instances and the audit still said DR 92 vs RG 124. Found by the numbers not moving after a fix that was known to have worked. - **A grey-asset scan that ignores visibility flags reports invisible collision helpers as defects.** 10 hidden cylinders would have kept failing D-DR-02 forever. Found by running the audit after the harvest. - `set_material_instance_texture_parameter_value` returned without error and **wrote nothing** to the kangaroo MI (same class of silent no-op as the Quantum camo in Session 087). The fix is a read-back through a different API, and the read-back is now part of the step. ### RISKS - **New R-92 — the kangaroo grey is unresolved.** The material binding is verified and the height is fixed, so the defect is one of: the kangaroo texture assets, the mesh's UVs, or the parent `M_SS_ScanPBR`'s material-usage flags (a material not flagged for static meshes renders with the engine's flat default material, which is exactly the clay-grey read). **Do not close D-DR-02 on the MI read-back alone.** - **New R-93 — the creek water is unverified in game.** Geometry and material are measured; nobody has seen it rendered. If it is invisible in play, the next suspects are the translucent blend mode at this exposure and the segment width against the excavated bed. - `M_SS_CreekWater`'s panner speed could not be set (`MaterialExpressionPanner.Speed` is protected in 5.8), so the wave animation runs at the node's default rate. Cosmetic; the material still animates. - D-DR-01 is only **partially** evidenced: the audit's `thin_scan` is a bounds proxy, not the producer's look. The assets that showed the defect are gone, which is the strongest single piece of evidence. - D-DR-04's *aesthetic* half (does the composition read well?) is the producer's call and is not closed by a placement rule. ### NEXT ACTION One play session on Dry River with three questions only: **is there water in the creek**, **is the kangaroo still grey**, and **does the gum ring read as a framed windmill** — then fix the kangaroo by dumping the kangaroo texture assets' real state (source size, sRGB, compression) and the mesh's UV channel count before touching the material again. ## Session 089 — 2026-09-30 — The packs git does not carry are now measured, verified and registered ### COMPLETED - Started from the producer's question — *is there a way of opening Unreal without manually importing everything in the content folder?* — and answered it by measurement: **there is no import step.** Every `.uasset`/`.umap` in `Content/` is already imported content, including all 14 vendor packs; the ~4,000 `.fbx`/`.blend`/`.tga` files sitting inside `Content/` are source kept beside their imports and Unreal ignores them. What the question was really hitting is the next finding. - **The repository alone cannot open any map.** Tracked `Content/` is ~350 MB; the tree is ~111 GB, and **14 packs — 25.8 GB, 4,318 files, 685 referenced packages** — are gitignored by ADR-021. Dry River references 207 Namaqualand packages, Ravenshoe 176 Singapore Canal ones, Bluestone and Dry River 102 QuarrySlate ones. This was a known cost of a deliberate rule, not a defect, and it is now a number in a file rather than a fact only the build machine knows. - `Tools/check_asset_references.py` (new, stdlib only, **2.7 s for 9,364 assets / 37,044 references**) reads every committed asset's binary for the package paths it names and sorts each into `ok`, `untracked`, `missing`, `folder`, `artifact` or `external`. This is the check for the silent failure `M_SS_CreekWater` proved: a committed asset naming a file git does not hold renders correctly on the build machine and wrong everywhere else. - `Tools/verify_packs.py` (new) verifies each pack in the manifest is present with the recorded file count and byte size, optionally hashing every file with `--deep` (the check that enforces ADR-021's "never modified" half), and cross-references both registers against live usage. - `Docs/PACK_MANIFEST.md` + `Docs/PACK_MANIFEST.json` (new): the 14 packs with size, file count, referenced package count, licence, register row, restore route and per-pack notes. The JSON is generated; `licence`/`register_row`/`restore`/`note` are hand-maintained and survive regeneration. - **`ASSET_REGISTER.md` §4.9h corrected by the check, not by reading.** World Flags and FP_AKS74U Animation were both marked `NOT_USED` while committed assets referenced them — the MAF weapon meshes take `MI_AKS74U`/`MI_Magazine` from the animation pack, and both flag material instances parent off World Flags. Both now `IN_USE`. A Stone Well row was added for a 1.2 GB pack `L_DryRiver_01.umap` references with **no licence record in either register**; it is written as `UNREGISTERED — PROVENANCE UNKNOWN` so the gate stays red until the listing is identified. - Both checks wired into `.github/workflows/build.yml` as **advisory** (`continue-on-error`), uploading `Build/asset_refs.json` and `Build/pack_verify.json` as build artifacts on every run, with the comment recording exactly what must happen before they can block. ### FILES CHANGED - `Tools/check_asset_references.py`, `Tools/verify_packs.py`, `Tools/asset_reference_baseline.json` (new). - `Docs/PACK_MANIFEST.md`, `Docs/PACK_MANIFEST.json` (new). - `Docs/ASSET_REGISTER.md` §4.9h (two status corrections, one new row, one open-question annotation, and a note that the table is machine-checked), `Docs/CHANGELOG.md`, `.github/workflows/build.yml`. - **Not touched:** any character-model or Wandarra content or C++, and `probe_quantum_material.py`. The manifest *records* that Modern Insurgent 7 and QuantumCharacter are dependencies, which is a note about them, not an edit to them. ### TESTING | Check | Command | Result | |---|---|---| | Reference audit | `python Tools\check_asset_references.py` | **9,364 assets, 37,044 references in 2.7 s** — `ok` 13,495, `untracked` 738, `missing` 101 raw | | Same, after de-noising | with prefix-artifact and folder-reference handling | **`missing` 0 after baselining 31**, `folder` 31, `artifact` 70, `external` 22,679. **No project-authored asset has a dangling reference** | | Pack verification | `python Tools\verify_packs.py` | **12 OK, 2 problems** — Singapore Canal (`REGISTER_MISMATCH`, 11 referencing assets vs a `NOT_USED` row) and Stone Well (`UNREGISTERED`, provenance unknown). 25.8 GB / 4,318 files measured | | Live-read proof | edited a pack row to `NOT_USED` **without** regenerating the manifest, re-ran | `Content/Scene_QuarrySlate` immediately became `REGISTER_MISMATCH` (OK 11, problems 3); register restored. Proves the check reads the registers, not the manifest cache | | CI wiring | `yaml.safe_load` on `build.yml` + both CI commands dry-run with their real arguments | **PASS** — 24 steps, parses; both commands run and write their reports (both exit 1 by design, hence `continue-on-error`) | ### ASSETS - No asset imported, modified, moved or deleted. **Nothing in `Content/` changed this session.** - `Content/WaterPlane/` remains deliberately uncommitted except the one texture `M_SS_CreekWater` references (Session 088); it is not in the manifest because no committed asset depends on the rest. ### DEFECTS FOUND Five, all in the new tooling itself, and all found by disbelieving a result: 1. **The first scan reported 3,931 missing references and was wrong.** The regex demanded a trailing `.AssetName`, but UE stores a HISM or instanced component's mesh as a bare package path with no object name — so it found 16 references in `L_DryRiver_01.umap` where there are hundreds, and would have missed exactly the reference kind that matters. 2. **Asset names containing dots** (`NM_BPSystemEvent.NM_BPSystemEvent`) resolved to themselves and were reported as missing references to themselves. Resolution now tries every dot boundary against what is actually on disk, longest name first. 3. **70 references were truncated-prefix strings** — `/Game/.../SM_Qua_Sla_Rock_S` where the real asset is `SM_Qua_Sla_Rock_S_10`. Those render correctly, so calling them broken buried the real ones. Caught because the "missing" list named Dry River rocks that visibly work in game. 4. **A substring search for `NOT_USED` matched prose.** Section 4.9j's row reads "Row corrected 2026-09-29 to `NOT_USED`, corrected again 2026-09-30 on first map use" on the row that now says `IN_USE`, so the pack it names was reported as unused. The status is now read from the status cell. 5. **The verifier read the register state from the manifest it was checking.** Fixing a register row changed nothing until the manifest was regenerated — the exact staleness the tool exists to catch. The check now recomputes from the registers every run, proven above. ### RISKS - **Historical Session 089 R-94 — `Content/StoneWell` provenance record was absent in that snapshot.** 1.2 GB, referenced by the map, seller/listing details not recorded. That snapshot note did not account for the producer's F2P project-use clearance; StoneWell was registered/producer-cleared in the later 2026-10-01 reconciliation (`ASSET_REGISTER.md` §4.9h, `PACK_MANIFEST.md`). This was not a current release or verifier clearance gate, and the listing identity remains a provenance gap only. The later 2026-10-01 status is documented in the referenced register rows. - **Historical Session 089 R-95 — Singapore Canal's observed references conflicted with an older art-direction note.** Eleven committed assets referenced generic corrugated/wood materials and props, not Asian canal layout or masonry. The later producer F2P direction clears the acquired pack for this project's use (ADR-028/035); `ASSET_REGISTER.md` §4.9h and `LICENCE_REGISTER.md` L-0016 retain the art-direction distinction. This is not a current permission gate. - The reference guard **cannot be made blocking** while ADR-021 holds: 738 untracked references are the expected state, not a fault. Blocking requires restore routes a fresh machine can follow, which is what `PACK_MANIFEST.json`'s `restore` field now records per pack. - The 31 baselined Lyra references are inherited debt. The baseline is a ratchet: it fails on anything new, and deleting a line from it to go green is the failure mode it exists to prevent. ### NEXT ACTION (historical) The original Session 089 next action was a producer ruling on R-95 and listing identification for R-94. The later 2026-10-01 reconciliation supersedes those unresolved-clearance implications; current provenance and project-use status are in `ASSET_REGISTER.md` §4.9h and `LICENCE_REGISTER.md` L-0016. `verify_packs.py` remains a technical dependency check, not an asset-clearance decision. ## Session 090 — 2026-09-30 — Next priorities written down: models, textures, VFX, recoil from real data, maps > **Historical snapshot notice (2026-10-01):** this session records its then-current observations and proposals; several are superseded by the current source/config and inventory review in `NEXT_PRIORITIES.md` §§3–6, `ASSET_REGISTER.md` §4.9m–p, and `MAPS_PLAYABILITY_AUDIT.md`. Specifically, 271 A-series weapon assets are now tracked (so the “0 weapon textures tracked” finding is stale); `Config/DefaultGame.ini` and `USSWeaponStatsSubsystem` apply selected magazine, spare-ammo, spread and RPM values (but do not establish sight zero or recoil tuning, and do not enforce semi-auto); public effective ranges do not establish sight zeros or recoil rankings; R-82 is the attended bake requirement for Wandarra/Ravenshoe, while Red Gum has separate measured playability failures; and the preceding intake's R-94/R-95 clearance concerns were superseded when StoneWell was registered and producer-cleared and Singapore Canal's generic material references were accepted for this F2P project (ADR-028/035; `PACK_MANIFEST.md`, `LICENCE_REGISTER.md`). Keep this entry as a dated record, not current implementation guidance. ### COMPLETED - The producer's playtest verdict — character models look horrible, weapons need better textures, VFX need work, weapons need zeroing and recoil built on real-world data, then the maps — is now a document: `Docs/NEXT_PRIORITIES.md`, integrated into `Docs/HANDOVER_CLAUDE_CLOUD.md` (header pointer, and §3.7 / §2's player-model row marked superseded, since that file is a 2026-09-29 snapshot whose body-selection question ADR-042 has since answered) and into `CLAUDE.md`'s start-of-session list. - **The recoil baseline is measured and it does not exist.** `WID_SS_{A88,A88G,A4,A416,A25,A89}` are copies of Lyra's rifle definitions; a `strings` scan of `WID_SS_A88.uasset` returns no authored recoil, spread, dispersion, damage or range keys, and `B_SS_A88_Weapon` references only Lyra's `/Game/Weapons/B_Weapon`. **All six weapons therefore fire with one generic rifle's recoil, including the LMG.** That is the defect behind the producer's feel, and it is now a written finding rather than an impression. - Real-world figures gathered from primary sources, with derived numbers marked as derived: **EF88** — 5.56, 30-round box, 680–850 rpm, **300 m effective** (Australian Army / Navy). **F89** — 5.56, **100 or 200-round box and belt-capable feed**, 750–1,000 rpm, **400 m point / 600 m area** (ADF Navy F89A1 page; FN MINIMI 5.56 MK3). **HK416** — 5.56, **790 m/s and 1,250 J** per Heckler & Koch's own product page. **M4-pattern** — 5.56, 880–910 m/s, 500 m point. - **The finding that matters is the shape, not the ranking:** the A89 is a different weapon rather than a bigger one (sustained fire without a reload, per-shot recoil comparable to the A88 — the most likely thing to get wrong); the A416 should be the hardest-hitting 5.56 (lowest quoted muzzle energy, piston system, large vertical); A88 and A88G must share ballistics exactly (ADR-004); and effective range is engagement design, not damage falloff. Recorded with an explicit warning that **free recoil energy is not a game recoil value**: the real data fixes the ordering, the ratios and the zero distances; per-shot climb is tuned by playtest. - Weapon texture baseline measured: the A88 carries five ADFRC maps (`adfrc_ef88_co`, `mbus_front_co`, `mbus_rear_co`, `adfrc_spectr_co`, `adfrc_spectr_ca`), but `git ls-files` shows **no weapon texture is in the repository** — they live in the git-ignored Sourced tree — and the pack ships no normal/roughness/AO/metalness (the same finding already recorded for Ravenshoe props, M-008l). The fix is documented as project-owned `M_SS_ScanPBR` instances, which also makes the recoil tuning reproducible. - VFX baseline measured: casing eject and muzzle light are in and tested; **muzzle flash was never placed** (`NS_WeaponFire_MuzzleFlash_Rifle` exists, candidates in `Docs/evidence/vfx_muzzle_candidates.json`); **tracers are still `PLACEHOLDER`** (E-002) and are called out as the notable gap at Dry River's ranges. - Map baseline measured from the audit and Session 088: **R-82 (attended nav bake) is the single blocker for three maps** (Wandarra, Ravenshoe, and Red Gum's remaining pass) and is cheap in effort. ### FILES CHANGED - `Docs/NEXT_PRIORITIES.md` (new), `Docs/HANDOVER_CLAUDE_CLOUD.md`, `CLAUDE.md`, `Docs/CHANGELOG.md`. - No content, no code, no asset touched. **Nothing in the character/animation/VFX/weapon areas was modified** — this session read and measured them only, since the character thread owns that work. ### TESTING | Check | Command | Result | |---|---|---| | Weapon stat baseline | `strings` on `WID_SS_A88.uasset`, `B_SS_A88_Weapon.uasset` | **CONFIRMED** — no authored stat keys; `B_SS_A88_Weapon` → `/Game/Weapons/B_Weapon` only | | Weapon texture commit state | `git ls-files | grep` for weapon textures | **0 weapon textures tracked**; only `T_ADFRC_DPC_camo.uasset` under `Content/Art/Characters/ADF/` | | Texture slots in use | `Build/weapons_setup.json` | A88 `textured_finishes` = 5 ADFRC maps, as listed above | | VFX state | `git ls-files | grep -iE "NS_.*(Flash|Muzzle|Impact)"` | muzzle flash asset present but unplaced; 4 impact systems present; **no tracer** | | Real-world figures | web search + `read_url` on `army.gov.au`, `navy.gov.au`, `heckler-koch.com`, `fnherstal.com` | EF88 / F89 / HK416 figures cited from those pages; M4 figures from standard published data; **derived muzzle energies marked derived** | | Document integration | grep | handover header pointer present; §3.7 superseded; `CLAUDE.md` start-of-session reads `NEXT_PRIORITIES.md`; one accidental duplicated bullet found by grep and removed | ### ASSETS - None added, imported, modified or deleted. ### DEFECTS FOUND - **All six weapons share Lyra's generic rifle recoil**, including the LMG, because the `WID_SS_*` and `B_SS_*` assets are copies of Lyra's with no authored stats. Found by scanning the assets for stat keys, not by playing. The producer's "recoil needs work" is this defect, and it predates this session. - **The handover file's player-model guidance is stale**: it told the next agent that ADR-036 (G3 body) stands and that Quantum is "unapproved". ADR-042 superseded that on 2026-09-30. Left in place with the history, but now marked superseded at both places it appears, so an agent cannot follow it by mistake. - My first edit to the handover duplicated a bullet rather than replacing it (the anchor text also matched the sentence I had inserted above it). Found by grepping the section back after the edit. ### RISKS - **New R-96 — weapon ballistics have no home.** The numbers in `NEXT_PRIORITIES.md` §5.6 need to become a project-owned data asset plus a component override; they must not be edited into Lyra's assets (ADR-004) and must not be hardcoded in C++, or the cloud role cannot check them. Until that exists, the real-world data is documentation only. - **New R-97 — weapon textures are uncommitted.** 0 of them are tracked, so a clean checkout has no weapon finish at all. This is ADR-021 working as intended (vendor content stays out of git) but it means the "weapons look bad" verdict cannot be re-checked on another machine. Interacts with `Docs/PACK_MANIFEST.md`: the Sourced tree is not a pack and is not in the manifest. - `NEXT_PRIORITIES.md` §5.5 is the constraint on R-96: per-shot climb and recovery are **playtest-tuned**, and only the zero distances, feed and rate-of-fire limits come from the real data. ### NEXT ACTION Producer to confirm §8's order — in particular whether weapon textures (Priority 2) really precede VFX (Priority 3) — and to rule on the A89's intended feel, because sustained fire versus per-shot kick is the one recoil decision that is a design choice rather than a derivation from the real weapon. ## Session 091 — 2026-10-01 — Gemini-reported player preview/material changes and character inventory (visual acceptance still open) > **Evidence boundary:** the C++ build, audit run and screenshot below are inherited Gemini-reported results, not independently rerun in this review. They describe the pre-review version of the code. The screenshot path is unavailable in this checkout, and the producer explicitly rejected the resulting class-select appearance. Subsequent working-tree source/material-tool edits are listed separately as unverified corrections; do not treat the reported PASS rows as validation of those edits. ### REPORTED CHANGES - **Suppressed Unreal Editor auto-import popup**: added `[/Script/UnrealEd.EditorLoadingSavingSettings]` with `bMonitorContentDirectories=False`, `bAutoCreateAssets=False`, `bAutoDeleteAssets=False`, `bDetectChangesOnStartup=False`, and `bPromptBeforeAutoImporting=False` to `Config/DefaultEditor.ini` and `Saved/Config/WindowsEditor/EditorPerProjectUserSettings.ini`. Stops the editor prompt to auto-import 60+ GB of loose source files in `Content/Sourced/ADF_Extracted` and `Content/Downloaded/VaultCache`. - **Enabled `UE5AIAssistant`**: verified plugin configuration in `Plugins/UE5AIAssistant` and `SouthernSpear.uproject`, ready to serve HTTP control on `localhost:58080` when `UnrealEditor.exe` is opened. - **Gemini-reported Class Select Preview changes**: - Gemini reported the preview stage had raw mesh components playing `A_MM_Idle` while freezing `StageBody`, and diagnosed floating ADFRC gear plus unoverridden civilian clothing. These symptom/root-cause claims are not independently confirmed from the unavailable screenshot. - Replaced the ad-hoc preview generation with a `B_SS_Soldier` (`ASSCharacterPartActor`) child actor attached to `StageBody`, intended to share runtime retargeting, locality, material overrides and gear pose. - **Gemini-reported skeletal material fix**: - Gemini reported that `M_SS_ADFRC_Camo` lacked skeletal-mesh usage and that this caused a standalone grey checkerboard; the supplied render is unavailable, so this diagnosis is not independently confirmed. - Updated `Tools/Unreal/setup_quantum_proto.py` to set `used_with_skeletal_mesh = True`; Gemini reported recompile and instance saves. Later edits to the setup script/material graph have not been applied in Unreal. - **Reported audit of the active friendly character assembly (inventory only; R-58 remains open and R-59 is measured)**: - Authored and executed `Tools/Unreal/audit_active_character.py` (`Build/active_character_audit.json`). - Gemini reports 107,016 LOD0 vertices across six parts, the Quantum/Manny skeleton split, physics assets on Quantum modules only, and one LOD per part. This is reported inventory, not visual acceptance or performance profiling. - Updated the risk/docs baseline: R-59 records the reported inventory as measured; R-58 remains OPEN for runtime and visual acceptance. - **Gemini-reported in-engine rendering (not independently reviewable here)**: - Session report says `-SSShotAt=8` on `L_DryRiver_01` captured `Saved/Screenshots/WindowsEditor/SSShot.png`, with aligned gear and rendered DPCU camo. - Producer subsequently reported that the last in-game class-select popup did **not** show a good character model. The screenshot is unavailable in this checkout. Treat that direct feedback as a visual rejection; do not close appearance acceptance based on a capture command or mesh audit. - **Source-review corrections after producer feedback (working tree, not built)**: - Code review found locality was applied to the child actor immediately after registration, before its `BeginPlay` built the mesh-component arrays; `ApplyViewerLocality` can then mark itself resolved while showing no parts. Locality is now applied on the widget tick after `HasActorBegunPlay()`. - The old rotation was an intentional three-quarter view (about 55° from the camera-facing +X axis); source alone does not establish that it showed the back. It is now centered toward the camera with a restrained ±25° turn so the face/torso should read more clearly; this composition still needs visual review. - For the 3:4 target, the old 28° horizontal FOV at 330 cm yields about 220 cm vertical coverage. The preview now uses 32° at 360 cm (about 275 cm vertical coverage) for a full-body frame with room for the weapon. This is geometry, not a rendered crop measurement; no fresh runtime capture/build has validated the correction. - Because the runtime character-part components initialize with `OwnerNoSee=true`, the preview explicitly clears that flag on the spawned skinned parts after locality is applied; without it the child actor can still be absent from the capture. - `M_SS_ADFRC_Camo` now samples the generated fabric normal and ORM maps instead of flat roughness/no normal. Texture settings are requested/read errors reported explicitly in the setup script; shader recompilation still requires an Unreal run. - The active-character audit now reports LOD0 triangle totals and marks an incomplete/unmounted Game Feature scan as not OK. ### FILES CHANGED - `Config/DefaultEditor.ini`: added `EditorLoadingSavingSettings` to suppress auto-import prompts. - `Plugins/SouthernSpearUI/Source/SouthernSpearUI/Public/SSClassSelectWidget.h`: tracks `StageSoldier` (pushed); review follow-up adds non-reflected locality state. - `Plugins/SouthernSpearUI/Source/SouthernSpearUI/Private/SSClassSelectWidget.cpp`: pushed child-actor preview; review follow-up defers locality, clears owner-hidden on newly built skinned parts, and adjusts turntable/camera framing. - `Tools/Unreal/setup_quantum_proto.py`: pushed version enabled skeletal-mesh use; review follow-up adds generated fabric normal/ORM inputs and stricter save/error reporting. - `Plugins/GameFeatures/SSExp_ObjectiveAssault/Content/Characters/QuantumProto/M_SS_ADFRC_Camo.uasset`, `MI_SS_ADFRC_Camo_Shirt.uasset`, `MI_SS_ADFRC_Camo_Jeans.uasset`: Gemini reported saving/recompiling them; the pushed binaries predate the review material-graph enhancement and require an Unreal setup run to regenerate. - `Tools/Unreal/audit_active_character.py`: pushed audit script; review follow-up includes triangle totals and fails incomplete measurements. The generated `Build/active_character_audit.json` is not available through the current file reader. - `Docs/PROJECT_AUDIT.md`, `Docs/PLAYER_MODEL_PLAN.md`, `Docs/NEXT_PRIORITIES.md`, `Docs/ASSET_REGISTER.md`, `Docs/DECISION_LOG.md`, and `Docs/CHANGELOG.md`: separate reported inventory/body choice from open appearance acceptance; correct current camo-source and preview-status language. ### TESTING | Check | Command | Result | |---|---|---| | Architecture Guard | `python Tools/validate_architecture.py` | **Gemini-reported PASS** — not rerun in this review; applies to the pre-review tree | | Unity Name Check | `python Tools/check_unity_names.py` | **Gemini-reported PASS** — not rerun in this review; applies to the pre-review tree | | C++ Editor Compilation | `Build.bat SouthernSpearEditor Win64 Development ...` | **Gemini-reported PASS** — not rerun after the current preview changes | | Active Character Audit | `UnrealEditor-Cmd ... audit_active_character.py` | **Gemini-reported result** — 107,016 LOD0 vertices; report is unavailable in this checkout and the current audit script was subsequently edited | | In-engine Rendered Check | `UnrealEditor.exe ... L_DryRiver_01 -game -SSShotAt=8` | **Gemini-reported capture only** — screenshot unavailable here; producer reports the class-select model still looks poor, so appearance is rejected/open | ### RISKS - **R-58 remains open for visual/runtime acceptance.** The reported inventory identifies the intended skeleton boundary, but not successful preview visibility/fit; a later source review found the locality-ordering hazard and a producer-rejected capture. - **R-59 geometry inventory reported** at 107,016 LOD0 vertices; all six parts are LOD0, so cost/LOD work remains open. The updated tool also reports triangles on its next mounted-feature run. ### REVIEW FOLLOW-UP (working tree; not committed or Unreal-verified) - `SSClassSelectWidget.cpp`: locality is deferred until the child actor has completed `BeginPlay`, then the new character's `OwnerNoSee` flags are cleared; the angle is centered toward camera and portrait coverage increased based on FOV/aspect geometry. C++ compiled successfully, but none of this substitutes for judging a rendered capture. - `setup_quantum_proto.py`: builds the camo master from the generated fabric normal and ORM maps as well as the DPC base color. The script now fails material setup if settings/compile/save report failure; generated texture settings and material graph have not been applied or rendered in Unreal. - `audit_active_character.py`: reports triangle counts and marks missing/unmounted parts as a failed audit; it has not been rerun against the mounted Game Feature. - Character-related docs distinguish producer rejection from the body-source choice and historical G3 diagnosis. No claim of visual acceptance remains. ### TESTING (review follow-up) | Check | Result | |---|---| | `git diff --check` | **PASS** for the reviewed working tree at the time of the check; CRLF/LF warnings only | | Python syntax check (`python -m py_compile` on the two edited Unreal tools) | **PASS** | | Project JSON, architecture guard, unity-name check, `git diff --check` | **PASS** — architecture guard notes pre-existing accepted SS010; newline conversion warnings only | | Unreal Editor C++ build (`Build.bat SouthernSpearEditor Win64 Development`) | **PASS** — 20.05 s, including modified class-select source; three pre-existing plugin dependency warnings | | Unreal material setup, mounted-feature audit and fresh in-game screenshot | **NOT RUN** — still required for shader/material data and visible quality review | ### NEXT ACTION Run the updated material setup and mounted-feature audit, then capture the class-select preview in-game and inspect it with the producer. Revise outfit/art direction from the rendered evidence; do not tune or accept the camo from source-only inspection. ## Session 092 — 2026-10-01 — ADFRC friendly assembly in the class-select preview; the shirt sheet's flat panel measured and repainted The producer reviewed the class-select preview twice during this session and reported, in order: no helmet and no webbing with the wrong camouflage; then better camouflage but "a weird gap at the waist"; with the earlier note that the preview is a third-person view of the player and weapon, so it must show the right model, textures and animation. Each change below is a response to a rendered capture, and the captures are quoted rather than described. ### COMPLETED **Friendly assembly changed to the ADFRC gear (`Tools/Unreal/setup_soldiers.py`).** `FriendlyParts` is now the single retargeted Quantum head; `FriendlyLeaderPoseParts` is the ADFRC `SK_ADF_Uniform_G3`, `SK_ADF_Vest_TBAS` and `SK_ADF_Helmet_OpsCore`, so the uniform, webbing and helmet are all fitted to the mannequin the pawn animates. Run: `ok: true`, `1 friendly retarget, 3 friendly leader-pose, 4 opposing`. **Camo reverted to the ADFRC-authored AMCU sheet (`Tools/Unreal/setup_adf_soldier.py`).** The previous edit reassigned the G3's AMC slots to the pack's `Crye_G3_{Shirt,Pants}_DPC_co`. Viewed side by side, AMC is the Multicam-style AMCU skin and DPC is the older Auscam pattern, so that reassignment is reverted: the G3 ships its own AMCU atlas for these UVs. The two DPC textures that edit imported were deleted (no asset referenced them; `grep -rl` over the plugin content). **Game Feature content now loads in a commandlet (`Tools/Unreal/probe_load_routes.py`, measured).** `unreal.load_asset("/SSExp_ObjectiveAssault/...")` returned `None` for every ADF mesh in the probe while `B_SS_Soldier` resolved, and `setup_soldiers.py` hard-fails on a None mesh. Cause: the plugin's content is not in the asset registry until it is scanned. `scan_paths_synchronous(["/SSExp_ObjectiveAssault"], force_rescan=True)` makes all four routes (`load_asset`, `EditorAssetLibrary.load_asset`, registry lookup, `does_asset_exist`) resolve every probed package, including the ones that previously answered None. Added to `setup_soldiers.py` and the probe. **The "weird gap at the waist" was measured, not guessed, and it is a texture.** `Tools/Blender/inspect_uniform_fit.py` (new) on the fitted `SK_ADF_Uniform_G3`: the trunk has faces within 18 cm of the axis in every 2.5 cm slab from 75 cm to 155 cm, so the mesh has no hole; and the torso faces sample UV v 0.02..0.24 of `Crye_G3_Shirt_AMC_co.png`, which `Tools/Common/ss_sheet_probe.py` shows is the sheet's plain khaki under-shirt panel — no camouflage in it at all. The trouser sheet is camouflaged in the same band, so this is the shirt sheet's own layout, not the fit and not the lighting. **`Tools/Textures/patch_adfrc_undershirt.py` (new) repaints that panel in pattern.** It writes `Art/Characters/ADF/T_ADFRC_G3_Shirt_AmcuCamo.png`: every low-variance, non-background 64 px tile is replaced with camouflage mirrored across the sheet from `Crye_G3_Pants_AMC_co.png` (the same pattern at the same scale), so adjacent tiles continue one reflection and the repeat does not read as a grid. `setup_adf_soldier.py` points the G3 shirt slot at it (`T_ADF_G3_Shirt_AmcuCamo`). Two rejected first attempts are recorded in the report: copying the sheet's own most-patterned band gave a near-black belly (brightness now matched to the panel being replaced), and tiling a small window gave a visible 64 px checkerboard with the source's black margins (now a single mirrored source with its black fraction reported as 0.04). **Class-select preview fixes (`SSClassSelectWidget.cpp`).** The rifle idle now loops instead of freezing at 35 % of its length, which is what made the earlier captures read as a mid-animation stumble with the head pitched down and the helmet apparently absent; any weapon component whose asset name contains `Arms` is hidden in this third-person preview (logged, and none was found on `B_SS_A88_Weapon`, which carries `SK_Rifle` hidden and `SM_A88` visible); the stage key/fill/rim lights drop from 9000/7000/8000 to 5200/4200/4600 and the backdrop from (0.16, 0.17, 0.13) to (0.11, 0.12, 0.095), because the camouflage was washing out; the child soldier is added to the capture's show-only list on the tick that applies locality, since it is built after that list is filled. **Retargeted parts are aligned to the leader (`FriendlyRetargetAnchor`, `SSCharacterPartActor`).** The retarget keeps its own skeleton's rest translations, so a module can land at its own bone position rather than where the mannequin's fitted gear is. The whole module is now moved rigidly so `FriendlyRetargetAnchor` (default `head`) sits on the leader's bone of the same name, with the measured gap logged. **Measured today: 0.0 cm on this assembly** — the alignment is a guard for a differently proportioned module, not the fix for the helmet, which was the preview's frozen pose. ### FILES CHANGED - `Plugins/SouthernSpearTeam/Source/SouthernSpearTeam/{Private/SSCharacterPartActor.cpp,Public/SSCharacterPartActor.h}` — part/anchor diagnostics, rigid anchor alignment, locality log trimmed to one line per change. - `Plugins/SouthernSpearUI/Source/SouthernSpearUI/Private/SSClassSelectWidget.cpp` — looping idle, weapon-arm hiding, darker stage, show-only guard, locality log. - `Tools/Unreal/setup_soldiers.py`, `Tools/Unreal/setup_adf_soldier.py` — ADFRC assembly, AMCU revert, patched-sheet wiring, registry scan. - `Tools/Textures/patch_adfrc_undershirt.py`, `Tools/Blender/inspect_uniform_fit.py`, `Tools/Unreal/probe_soldier_parts.py`, `Tools/Unreal/probe_load_routes.py`, `Tools/Common/ss_shot_html.py`, `Tools/Common/ss_shot_grid.py`, `Tools/Common/ss_sheet_probe.py` — new measurement and review tools. - `Art/Characters/ADF/T_ADFRC_G3_Shirt_AmcuCamo.png` — new derivative sheet (untracked). ### TESTING | Check | Result | |---|---| | `UnrealEditor-Cmd -run=pythonscript` × `setup_adf_soldier.py`, `setup_soldiers.py`, probes | **PASS** — `ok: true`, 0 errors; probe reads the CDO back as 1 retarget + 3 leader + 4 opposing with AMCU textures on every gear slot | | Blender `inspect_uniform_fit.py` on `SK_ADF_Uniform_G3.fbx` | **PASS** — 34,061 faces, 4 slots, continuous trunk coverage 0.75–1.55 m, torso UVs on the sheet's plain panel | | `Build.bat SouthernSpearEditor Win64 Development` | **PASS** — succeeded, no errors | | `Tools/run_map_capture.sh /Game/Maps/L_DryRiver_01 20 420` | **PASS** — map loaded, `Requested viewport screenshot at 20.0 s`, image 977,670 bytes; inspected at 2–4× | | `python -m py_compile` on the new/edited tools; `git diff --check` | **PASS** | | Runtime log (`Saved/Logs/SS_capture_20261001_194108.log`) | `SSCharacterPart B_SS_Soldier_C_4: 1 retarget + 3 leader + 4 opposing`, `anchor 'head' rest gap to leader 0.0 cm`, `SSClassSelect: stage locality applied; 8 skinned part(s), 3 in capture list` | ### DEFECTS FOUND - **Game Feature assets resolve as `None` in `-run=pythonscript` until the registry is scanned.** Found by probing four load routes instead of trusting one; it would have failed `setup_soldiers.py` with "required ADFRC friendly mesh/head failed to load" for meshes that are on disk. - **The G3 shirt sheet's torso region is a flat under-shirt panel.** Found by measuring the UVs the torso faces actually sample and reading the sheet at those points — a screenshot could not separate this from a hole, and two renders were spent on that ambiguity. - **My own DPC reassignment was wrong.** Found by rendering the two candidate sheets side by side before committing to either. ### RISKS - **R-92 — The repainted under-shirt panel is a derivative art decision, not a producer one.** `T_ADFRC_G3_Shirt_AmcuCamo.png` fills the shirt sheet's plain panel with the pack's camouflage. It is a texture derivative, so it is reversible by re-running `setup_adf_soldier.py` without the patch; the producer has not yet seen it. - **R-93 — The friendly and opposing heads now come from the same rig family, and the friendly look has no ADFRC head.** The ADFRC source set contains no character head (Session 086), so the friendly head remains the retargeted Quantum module. ### NEXT ACTION Show the producer this capture at full size and settle the remaining art calls on the friendly soldier (head choice, webbing fit, camo finish) before further tuning; the class-select preview path is now measured end to end. ## Session 093 — 2026-10-01 — Both agents' work committed and pushed; the four new installs ignored rather than vendored Producer instruction for this session: **commit all changes to git and GitHub, and update the documents for what the other AI agent did as well as for what I did.** Two agents had uncommitted work in the same tree, so this session is a reconciliation and a commit, not new feature work. Nothing in the committed content changed the runtime state described in Session 092; the character/preview work was already built, captured and measured there. ### COMPLETED **Every change in the tree is now committed and pushed** (see the git testing rows: `git push origin main`, 136 changed paths, 0 files deliberately excluded beyond the ADR-021 pack folders). **The other agent's work is committed, not rewritten.** Its files went into the same commit as mine so the two sets of changes stay reviewable together: `SouthernSpear.uproject` and `.gitignore` (enable and ignore the third-party `UE5AIAssistant` editor plugin), `Config/DefaultGame.ini` (W1 weapon-stats comment rewording only), `Content/Art/Blockout/SS_MAP_DryRiver_01.uasset`, `Tools/Unreal/audit_active_character.py`, `Tools/Unreal/setup_quantum_proto.py`, `Tools/Weapons/adfrc_weapon_data.py`, `Docs/WEAPON_SOURCE_DATA.{md,json}`, `Docs/evidence/asset_inventory_20261001.md`, `Docs/evidence/ui_session088/` (~190 MB of producer-review screenshots and HTML sheets), and the documentation reconciliation across `LICENCE_REGISTER` (ADR-035 F2P clearance reaffirmed), `PACK_MANIFEST`, `MAPS_*`, `SOURCED_ASSET_REVIEW`, `LOCOMOTION_AUDIT`, `WEAPONS_ANIMATION_PLAN`, `HANDOVER_CLAUDE_CLOUD`, `NEXT_PRIORITIES`, `PROJECT_AUDIT` and Session 091 of this file. Session 091's "Evidence boundary" note is intact; no existing ADR was reopened or rewritten by this session. **Four newly imported vendor packs are ignored, not committed (`.gitignore`, producer decision, ADR-021).** `Content/HighPoly_Tree_Model/` (9 files / 37 MiB), `Content/PN_GrassLibrary/` (626 / 1.1 GiB), `Content/Splash/` (2 / 1.1 MiB) and `Content/WaterPlane/` (34 / 139 MiB) are now ignored on the same rule as the existing installed-pack block. This was the one judgement call put to the producer, and the answer was **ignore them per ADR-021**. Verified by measuring rather than assuming: `verify_packs.scan_references()` finds **0 referenced packages** under each of the four tops, so none of them is a map dependency and none belongs in the 14-pack manifest. Had they been committed, they would have added ~1.3 GB to the repository for content nothing committed uses. **One deliberate exception is documented rather than silently kept.** `Content/WaterPlane/Lake/Textures/T_MediumWaves_N.uasset` is already tracked (it feeds the committed material `M_SS_CreekWater`, ASSET_REGISTER ENV-005). A new ignore rule does not untrack a file, and a committed project material legitimately needs its texture, so it stays tracked; the `.gitignore` comment and ASSET_REGISTER §4.9n both say so explicitly so a later reader does not "fix" it as an oversight. **Two real bugs in `Tools/verify_packs.py` were found by running it, and fixed.** (1) It **crashed** with `KeyError: 'bytes'` — rows in the `NOT_IN_MANIFEST` report were built without the `files`/`bytes`/register fields the summary printer reads, so the tool could not report its own findings. (2) It then reported `Plugins/GameFeatures` as an uninstalled pack, a **false positive**: the reference scan walks the whole worktree and counted tracked project content (the Game Feature's ADFRC assets) as if it were an uninstalled vendor pack. `scan_references()` now skips any reference top that git already tracks (`tracked_tops`, compared lower-cased because `tracked_files()` returns a lower-cased set). Both are fixes to the verification tool, so the guard is trustworthy again rather than noisy. **Registers updated for the new installs.** `ASSET_REGISTER.md` §4.9n gains rows for `HighPoly_Tree_Model`, `PN_GrassLibrary` and `Splash` (and the `WaterPlane` row's size corrected to the measured 139 MiB), each marked installed-but-not-referenced-by-committed-assets, plus a paragraph stating the git treatment and the tracked-texture exception. `PACK_MANIFEST.md` §2 gains a short subsection explaining why the four roots are *not* manifest rows, so a reader who sees 14 packs and four more installed folders does not read it as an omission. ### FILES CHANGED - `.gitignore` — the four 2026-10-01 pack roots with the ADR-021 rationale and the tracked-texture exception; `Plugins/UE5AIAssistant/` (other agent). - `Tools/verify_packs.py` — `NOT_IN_MANIFEST` rows carry `files`/`bytes`/register state; `scan_references()` skips tracked reference tops. - `Docs/ASSET_REGISTER.md`, `Docs/PACK_MANIFEST.md`, `Docs/CHANGELOG.md` — this session's records. - Everything else in the commit: the Session 092 source/tool/asset work and the other agent's files listed above. ### TESTING | Check | Command | Result | |---|---|---| | Pack verification | `python Tools/verify_packs.py` | **PASS** — `14 pack(s) ... 685 referenced package(s)`, `OK: 14, problems: 0`, exit 0 (previously crashed, then false-positived) | | New-install reference scan | `verify_packs.scan_references()` filtered to the four tops | **PASS** — 0 hits each; no committed asset references `HighPoly_Tree_Model`, `PN_GrassLibrary`, `Splash` or `WaterPlane` | | Architecture guard | `python Tools/validate_architecture.py` | **PASS** (pre-existing accepted SS010) | | Whitespace / conflict markers | `git diff --check` | **PASS** — clean | | Python syntax | `python -m py_compile` on new and edited tools | **PASS** | | Staging correctness | `git status --porcelain` after explicit `git add` | **PASS** — none of the four pack folders appears in the index | | Push | `git push origin main` | **PASS** — `5450b4d9..9abd9974 main -> main`, `Uploading LFS objects: 100% (79/79), 165 MB`, exit 0; GitHub warned that `Docs/evidence/ui_session088/_sheet.html` (50.05 MB) exceeds its recommended 50 MB maximum. Accepted: the file is committed through LFS and GitHub stores it fine, so the warning is advisory, not a failure | | Remote agreement after push | `git rev-list --left-right --count origin/main...HEAD`; `git ls-remote origin main` | **PASS** — `0 0`, remote `refs/heads/main` = `9abd9974` | | Asset reference audit | `python Tools/check_asset_references.py` | **NOT CLEAN, pre-existing and not from this work** — `MISSING: 31 referenced packages`, all resolved by the ignored packs | | Editor build re-run | `Build.bat SouthernSpearEditor Win64 Development` | **NOT RUN this session** — the C++ sources are unchanged since the clean Session 092 build, so that result still stands and is cited as Session 092's, not re-claimed here | ### ASSETS - Registered as installed-but-unreferenced: `Content/HighPoly_Tree_Model`, `Content/PN_GrassLibrary`, `Content/Splash`, `Content/WaterPlane` (ASSET_REGISTER §4.9n; not `PACK_MANIFEST` rows). None is producer-cleared *for use* — presence is not clearance, exactly as ADR-028/035 is worded; the producer reaffirmed project-use clearance for acquired assets on 2026-10-01 (LICENCE_REGISTER), which covers the acquired set but does not make installed content shipped content. - Committed from this session: `Art/Characters/ADF/T_ADFRC_G3_Shirt_AmcuCamo.png` and its imported `T_ADF_G3_Shirt_AmcuCamo.uasset` (Session 092's repainted under-shirt panel, a texture derivative of the ADFRC sheet; reversible, R-92). ### DEFECTS FOUND - `verify_packs.py` `KeyError: 'bytes'` — found by running the tool while preparing to commit; the reporting path crashed before it could print anything. - `verify_packs.py` false positive on `Plugins/GameFeatures` — found by the same run; the scan treated tracked Game Feature content as an uninstalled vendor pack. - Found, not fixed, and now a known defect: `check_asset_references.py` reports 31 missing packages. These come from the installed-but-ignored packs and predate this session; the tool needs the same "is this top tracked or installable" awareness that `verify_packs.py` now has. ### RISKS - **R-92, R-93 (carried from Session 092, unchanged).** The repainted under-shirt panel still needs producer acceptance, and the friendly soldier's head is still the Quantum module because the ADFRC set has no head. - **R-94 — The four ignored packs are unreproducible from this repository, by design.** They are in the `.gitignore` under ADR-021 and they have no `restore` field in `PACK_MANIFEST.json`, because they are not dependencies of anything committed. A machine that lacks them loses nothing the repository needs, but if any of them is later used in a map, that map has just acquired a new unrecorded pack dependency and the manifest must be regenerated. `Content/Splash/` is the sharpest case: a `Splash.bmp` + `Splash.uasset` pair whose origin and intended use are still not established (§4.9n). ### NEXT ACTION Get producer acceptance on the Session 092 preview capture (R-92, R-93); this session deliberately left the art alone so the commit contains only work that has already been measured. ## Session 095 — 2026-10-03 — First-person arms: gloves read as gloves, and the arms sit lower in the view Producer: first-person models and animations are the priority. Captures of the current build (Dry River, `-SSNoClassSelect -SSShotAt`) showed a fat camouflage forearm across the lower left in the rifle view, bare flat-tan mitts, and pistol arms filling the lower third of the screen. ### COMPLETED - **Arm placement.** `ss.FP.ArmsOffset` default `17 0 -2` -> `13 3 -8`; new `ss.FP.PistolArmsOffset` default `11 0 -8` (the pistol set uses it). The scope view is unchanged (aim correction is computed from the sight position). - **Glove texture.** The glove region of `T_FP_Arms_*_BC` was a flat fill. New `Tools/Blender/bake_fp_arms_ao.py` bakes ambient occlusion from the arms mesh into its own UVs; `make_fp_arms_texture.py` multiplies it into the sleeve and glove, adds a knit weave, and the glove colour changes from coyote (118,98,72), which read as bare skin in sun, to olive-drab (86,82,64). `setup_fp_arms.py` re-imported both sets (`ok: true`). ### FILES CHANGED `Plugins/SouthernSpearLyraBridge/.../SSFirstPersonSubsystem.cpp`; `Tools/Blender/bake_fp_arms_ao.py` (new); `Tools/Textures/make_fp_arms_texture.py` (also carries the other agent's uncommitted `de_plain` change); re-imported `FirstPerson/{Rifle,Pistol}` mesh, material and texture assets. ### TESTING | Check | Result | |---|---| | `Build.bat SouthernSpearEditor Win64 Development` | PASS | | `setup_fp_arms.py` headless | PASS, `ok: true`, no errors | | Captures, rifle and pistol, before/after | `Docs/evidence/s095/fp_{rifle,pistol}_{before,after}.png` — gloves read as olive fabric, rifle forearm no longer crosses the view, pistol hand no longer fills the screen | | Aim-down-sights after the offset change | scope view checked before the glove change only; NOT re-run after it | | Automation suite | NOT RUN | ### ASSETS Derived texture from the Fab M4 FPS pack's `Hand_D.jpg` (ADR-028 cleared) and the ADFRC AMCU sleeve fabric (L-0021). ### RISKS - **R-95** — The hands still have no individual finger or knuckle detail beyond baked occlusion; the pistol palm is a large plain olive area. A purpose-made gloved-hands mesh (the CC BY 4.0 gloves pack needs a credit line) would be the real fix. - Left-hand placement on the rifle handguard is still the accepted A88 solve only (R-85). ### DEFECTS FOUND Found by capture, not by reading code: the arms-offset default put a forearm across the view, and the glove fill read as skin. ### Session 095 addendum — the glove asset on the first-person arms Producer: the hands looked average; use the glove asset on file. `Content/Sourced/Gloves` (Fab "Gloves for fps game", Bobeer, CC BY 4.0; ASCII FBX that Blender refuses) is now read by `Tools/Blender/ascii_fbx_mesh.py`; `Tools/Blender/fp_arms_gloves.py` (run by `fp_arms.py` with `SS_FP_GLOVES=1`) removes the pack's bare hand faces, fits the glove hands onto the pack hand (best of 24 axis rotations, chamfer 0.9 cm, scale 0.86), slides the cuff under the sleeve, and skins the gloves from the pack mesh's own weights (inverse-distance over the 4 nearest vertices; Blender's Data Transfer modifier left every vertex on the forearm bone, so the fingers did not curl in the first capture). `Tools/Textures/make_fp_gloves_texture.py` builds the 2048 glove textures; `setup_fp_arms.py` makes `MI_FP_Gloves` for the new `FP_Gloves` slot. Capture (producer screenshots): the pistol grip is wrapped by olive knuckle-guard gloves; the rifle's left hand is behind the weapon in the hip view. NOT checked: reload, draw and sprint animations (the producer says the animations are still wrong), the aim view, the right-hand grip on the rifle. New risk **R-96**: the glove hands are fitted in the pack's rest pose and follow the pack's finger animation, so grip shape on our weapons is only as good as the pack's M4/G17 grips. Second addendum (producer: "two different animations following changing the weapon / reload, the first one is good, and then it changes to what's currently there"). Cause, from `SSFirstPersonSubsystem.cpp`: draw, holster and reload run with the left-hand IK suppressed (the pack's clip poses the hand), then `UpdateArmsAnimation` switches to the one-frame Idle clip and `Play()` re-enables the IK, so the hand jumps from the clip's pose to the IK solve. Fix: new `ss.FP.HandIK` (default 0) keeps the IK suppressed after every one-shot, so the clip's own hand pose holds. Capture after: left forearm and glove sit on the handguard in the pack's pose (`Docs/evidence/s095/fp_rifle_noik.png`). NOT verified: that this equals the pose the producer called good (the draw clip itself was not captured: a Blender pose dump of the actions returned identical poses for every clip, so that check is invalid and was discarded), reload, sprint. `ss.FP.HandIK 1` restores the old behaviour. The Unreal MCP server was started with `-ExecCmds="ModelContextProtocol.StartServer"` (port 8000, `.mcp.json` generated); it is not auto-start, and this Claude session had to be reconnected to see it. ### NEXT ACTION Producer to review `Docs/evidence/s095/` and the third-person friendly soldier; then fix soldier leg proportions. ## Session 096 — 2026-10-03 — Third-person soldier: legs measured, tunic and sleeves slimmed ### COMPLETED - Measured leg proportions: Manny pelvis 96 cm / head 163 cm, crotch ~46% of height; the G3 uniform crotch matches Manny's. The "short legs" look is the baggy tunic and sleeves, not the legs. - `Tools/Blender/adfrc_gear_rig.py`: new `SS_GEAR_KEEP` env (share of excess kept by the fit cap, default 0.15). Rebuilt `SK_ADF_Uniform_G3.fbx` with `SS_GEAR_CAP=1.5 SS_GEAR_KEEP=0.05` (0.8/0.0 tested and rejected: loses the belt, artefacts). - Re-imported via `setup_adf_soldier.py` and `setup_soldiers.py` headless (both exit 0, `ok: true`). ### FILES CHANGED `Tools/Blender/adfrc_gear_rig.py`, `Art/Characters/ADF/SK_ADF_Uniform_G3.fbx`, re-saved ADF uniform/material assets, `Docs/evidence/s096/`. ### TESTING Blender silhouette comparison (`tunic_caps.png`); class-select capture on Dry River (`class_select_slim_tunic.png`, PASS). Build, automation suite, vest fit, animations: NOT RUN. Not committed. ### RISKS Only the standing silhouette was checked; the vest still sits low; MAF shares the uniform mesh and was not captured. ### NEXT ACTION Producer to review `Docs/evidence/s096/class_select_slim_tunic.png`; then raise/fit the vest and check first-person reload/draw/sprint animations. ### Session 096 addendum — first-person clips reviewed frame by frame - **Tooling:** `ss.FP.DebugClip ` plays one clip on demand, `ss.FP.DebugSprint 1` shows the sprint pose, `-SSShotTimes=a,b,c` writes `SSShot_.png`; `Tools/run_anim_sequence.sh` and `Tools/Common/ss_contact.py` capture a clip and build a contact sheet. - **Defect found by capture:** since Session 095 lowered the idle arms (`13 3 -8`), the rifle reload moved the hands below the screen: eight frames showed a floating, tilted rifle with no hands (`reload_sheet.png`). - **Fix:** new `ss.FP.ClipArmsOffset` (default `17 0 -2`); the arms blend to it while a draw, reload or holster clip plays (`ClipAlpha`) and back to the idle offset afterwards. Result: hands and left forearm in view through the reload and draw (`reload2_sheet.png`, `draw_sheet.png`). - **Checked:** rifle reload, rifle draw, sprint pose (pistol held). **NOT checked:** pistol reload, holster, fire clip, empty reload, aim-down-sights after the glove change. Build PASS (editor closed); automation suite NOT RUN. - **Pistol (G17 arms):** reload shows the support hand bringing the magazine in (`pReload_sheet.png`); idle stays one-hand-forward at the Session 095 offset (raising it fills the bottom of the view with forearms, `pidle_cmp.png`). `ss.FP.DebugClip P:` waits for the pistol arms. Holster clip shows no lowering in the pistol set (not a gameplay-visible path; left). - **MAF/OPFOR third person: NOT reviewed.** `ss.Debug.FollowBot -180/-300` put the camera inside Dry River foliage (`maf_pair.png`: black silhouettes, blown-out leaves), so no usable view of the enemy soldier. Needs a camera that avoids geometry (or a posed preview like the class-select stage). - **MAF third person reviewed:** `ss.Debug.FollowBot` now line-traces and pulls the camera in front of foliage and walls (`SSFirstPersonSubsystem.cpp`). `maf2_pair.png`: the opposing soldier reads as a conventional olive-uniform infantryman with tan webbing and a helmet, firing and running with a believable pose; a second MAF bot is prone in the background. No defect found. - **Friendly vest:** the "sits low" note was my inference from mesh bounds, not a rendered defect; the class-select and running captures show it at chest height. Left as is. - **Empty reload** (`rempty_sheet.png`): hands in view through the charge; **fire clip** captured (`fire_sheet.png`, not inspected closely); **ADS** captured through the scope view (`ads_1.png`, UI off). Sprint remains the procedural tilt. ### Session 096 addendum — Dry River ghost gum foliage - **Defect (producer: "greygum foliage isn't rendering correctly"):** `MI_SS_GhostGum_Leaf` set an OpacityMask texture on `M_SS_ScanPBR`, which is Opaque, one-sided and has no opacity-mask node (probe: `Build/probe_scanpbr.json`), so every leaf card drew as a solid polygon: flat black in shade, blown white in sun (`maf_pair.png`, `gum_after.png`). - **Fix:** `Tools/Unreal/setup_gum_foliage.py` builds `M_SS_Foliage_Masked` (Masked, two-sided, two-sided-foliage shading, BaseColor x `Brightness` 1.5, Normal, OpacityMask on R, `ShadeFill` emissive 0.18) and re-parents the leaf instance. In the editor the gum canopy now shows proper leaf-and-branch silhouettes (verified by viewport capture from several instances). - **Still dark:** broad-leafed Namaqualand shrubs (`SM_searsia_lucida_*`, `MI_Searsia_Lucida_NN`) read near-black in shade; their material is the pack's own (ignored pack, not edited). A brighter project instance would fix it; not done. - **Found, not fixed:** several ghost gum instances are placed with extreme rotations (`SS_Gum_WindmillScreen_03` pitch 71°, `_04` pitch 86°, `_00` roll 180°); intent unverified. - NOT RUN: in-game (`-game`) capture after the final material; automation suite. ### Session 096 addendum — Dry River overhaul after producer review Producer, after my first foliage-only claim: windmill inside a tree and no fan, water tank untextured, no water in the creek, rocks and assets floating off the terrain, few shrubs. All checked in the editor over MCP and fixed with `Tools/Unreal/overhaul_dryriver_s096.py` (idempotent steps, `exec` over MCP) and `Tools/Blender/windmill_fan.py`: - **Creek water:** the old planes were buried ~2.8 m under the bed (the mesh pivot sits 309 cm below its surface) and tilted. 48 flat segments now follow the bed (`creek_water`); `M_SS_CreekWater` rebuilt (translucent, dark teal, scrolling `T_MediumWaves_N`). In-game capture shows water with rocks standing in it (`dr_after_2.png`). - **Windmill:** the ring of `SS_Gum_WindmillScreen_*` gums was moved out to 28 m, three instanced trees within 19 m removed, and a new wheel `SS_Windmill_Fan` (24 blades, rim, spokes, corrugated-iron material) placed on the head (`SS_Farm_Windmill_Fan`). Fan is static (no rotation yet). - **Water tower:** the Fab texture sheet bound to it renders flat tan (sheet is a 4k wood atlas; cause not found). Re-skinned with the project's weathered-timber material; reads as timber, not the Fab tank look. Open. - **Floating props:** 24 static props and 384 instanced rocks/stones/debris sunk 3 cm into the terrain (support-aware trace; rails, wires and tower parts skipped as elevated by design). - **Shrubs:** +3,200 instanced Namaqualand shrubs on flat dry ground outside the creek channel (`SS_S96_Shrubs_*`); the black broad-leaf `searsia` meshes were swapped for the rooibos bush mesh. - **Gums:** the 10 ghost gum assets (also near-black in shade) were replaced by the project's RuralAustralia gum trees (`SS_S96_Tree_*`); `M_SS_Foliage_Masked` remains for any later use. - NOT verified: Dry River playability audit re-run (3 pass / 6 fail), nav rebuild (R-82, attended), kangaroo appearance, performance of +3,200 instances, water shader on the final in-game frame beyond one capture. Not pushed. - **Tank and windmill wheel replaced with Australian designs (producer: Fab tank not good; wheel must be an Australian windmill).** Both are procedural (not from an asset), built by `Tools/Blender/aus_farm_props.py`: `SS_Aus_WaterTank` (round galvanised corrugated tank, domed lid, overflow pipe, on a splayed rusty steel stand with braces and ladder) and `SS_Windmill_Fan` (18 curved, cambered blades on two rings and flat spokes: Southern Cross / Comet multi-blade pattern). Materials: project corrugated iron and rust. Viewport-checked; not yet in an in-game capture. The Fab timber water tower is no longer used on the map. ### Session 096 addendum — the original Fab assets, reviewed on disk (producer: "they all came with textures") Producer was right: my procedural tank and wheel replaced assets that were fine. What the disk review found: - **ROOT CAUSE of grey tank / clay kangaroo / black gum leaves: material compile errors, not missing textures.** `M_SS_ScanPBR` (used by the Fab tower, kangaroo, ghost gum trunk and others) samples Roughness/Metalness/AO as "Masks" and Normal as "Normal", but its default textures were `DefaultTexture` and the instances' maps were imported with default compression, so every instance failed to compile and drew the fallback grey (`MaterialEditingLibrary.recompile_material` listed the errors). Fixed in the master: sane default textures (`/Game/Art/Environment/Fab/Defaults/T_SS_Mask_{White,Black,Rough}`, `DefaultNormal`); tower Normal/Roughness/Metallic re-set to Normal/Masks compression. The same class of error was in `M_SS_Foliage_Masked` (opacity map sampler type). Any other ScanPBR instance in the project may have been grey for the same reason; not audited. - **Water tower:** the pack's own `Water_Tower.fbx` (16 parts, 4k wood atlas) imported as `SM_Fab_WaterTower`, scale 1.3, with its textures: matches the Fab thumbnail (timber barrel, metal bands, timber stand, ladder). The decimated `SS_Raven_water_tower` is not used on Dry River now. My procedural corrugated tank is withdrawn from the map (asset kept: `SS_Aus_WaterTank`). - **Windmill:** the pack's `wind_mill.fbx` has all 13 parts including the wheel (Wings, Wings_Structure, Structure, Hub, Nose_Cone, tail Arrow, Cables, Feet, Ladder, Planks). `Tools/Blender/prep_fab_props.py` dropped the wheel (its discs read as ground planes), so `SS_Raven_windmill` had no fan. New `Tools/Blender/prep_windmill_full.py` keeps every part (7.0 m tower, 2 m wheel) -> `SM_Fab_Windmill` on the map. **The pack on disk contains only the FBX and a thumbnail: no texture files** (the listing promises a 4K PBR set). Slots currently use project rust / timber / corrugated iron; re-download the pack's textures from Fab to finish it. My procedural fan is withdrawn. - **Kangaroo:** 7 body slots each have their own map (0271..0278, plus fur-card and noise maps); the earlier setup put one body map on every slot. Textures were imported as 32x32 placeholders (editor-icon compression): re-imported at true size (667x558 etc.). Per-slot `MI_SS_Kangaroo_*` created and applied. The body still reads pale cream-grey: the source map's mean colour is (224,208,197), so that is the asset, not a fault. - **Ghost gum:** the supplied tree (ENV-003) restored on Dry River in place of the RuralAustralia stand-ins I had swapped in (`SS_S96_Gum_*`, 10, scale 1.3-1.8). With the compile fix its leaves are green and its bark grey. It is a small tree (6.5 m at scale 1). - Still open: Farmstead objective move did not persist (walk parity 58% unchanged); the playability audit still fails 7 of 9 rules (`Build/map_playability_dr_s096.json`: crossing 24 m, sightline 363 m, hard:soft 1.49, cover density 0.16, spawn exposure 25/64, walk parity 18%/58%); 266 cover props were added this session (`SS_S96_Cover_*`). Not pushed. - **Kangaroo (follow-up):** the editor still showed grey because `SM_Kangaroo`'s 11 slots held `WorldGridMaterial` (my earlier headless slot assignment had not stuck). Re-applied in the editor: per-slot `MI_SS_Kangaroo_*` on the mesh and the placed actors; viewport now shows the textured, orange-brown animal. ScanPBR textures with wrong compression (14 Ravenshoe maps) were corrected project-wide. - **Dry River playability audit, re-run** (`Docs/evidence/s096/map_playability_dr_s096.json`, `SS_BUILDPATHS=1`): **5 pass / 4 fail** (was 3/6): open crossing 18 m PASS, cover density 0.97 PASS, spawn exposure 0/64 PASS (was 35/64), starts PASS, close quarters PASS. Still failing: max sightline 355 m (target 220), hard:soft cover 2.12 (target 0.15-0.6; scaled "soft" props read as hard), walk parity Water Point 20%, Farmstead 12% (was 58%; the `SSObjectiveActor` moved to (700,-300) - the same-labelled billboard `Actor` is a separate object and `label()` helpers find it first). ~720 collision cover props (`SS_S96_Cover_*`, simple box collision added to nine pack meshes locally; those pack folders are git-ignored) were placed, which is a lot of props; thin if the map feels cluttered. - **Found, not fixed:** moving `SSObjectiveActor` Water Point from Python did not reproduce reliably between sessions. ## Session 097 — 2026-10-04 — Dry River: homestead, outstations, fallen timber, wildlife, creek water Producer: "give it a full overview, look at the placement of rocks, river beds, farm houses, water tanks etc. ... use [Fab and other maps' assets] and place across the map to give the map more life." **COMPLETED** - Overview (editor bird's-eye and ground captures): vegetation was already dense (about 18 k scatter instances, 300 grass trees, 700+ trees); what the map lacked was a place. The farm was a windmill and a tower in a field of rocks. - **Homestead compound** on the flattest ground north of the creek (best 12 m pads measured at 22-45 cm of relief), `Tools/Unreal/farm_life_s097.py`: RedGum farmhouse, two quarters, shearing shed, stock pens; Ravenshoe barn and old barn (timber and corrugated-iron slot materials as in `farm_dryriver.py`); stone well, hand pump (own MIs), RedGum tank, Fab tractor, two Fab chicken coops, dunny, drums, log piles, picnic table, barbecue, mailbox, five power poles, wreck car, 22-segment yard fence. Every building sits on its lowest ground sample with a stone plinth over the gap; trees, rocks and scatter inside each footprint were removed first. - **Outstations** (`farm_life_s097b.py`): 56 props around the nine existing sheds, lean-tos and tanks, two Fab windmills beside the west tanks, 26 Fab dead trees, 14 Fab fallen trees (collision), 45 Fab branches, four kangaroo mobs (13, existing per-slot MIs, scale 2.37 as the originals), three RuralAustralia kangaroo signs. - **Creek water** `M_SS_CreekWater` edited in place: DepthFade 12 cm on opacity (soft shore, tile edges gone), murkier green body, specular 0.6. Before: pale grey slabs with rectangular seams; after: green pools following the bed. - **Cleanup of Session 096 clutter:** 219 of the 311 randomly scattered crates, sacks and barrels (the ones more than 25 m from any structure) removed; 2 duplicate boulders deleted; 3 floating rocks snapped (one was 4.3 m up). Instance audit of 87 scatter meshes (about 1,000 samples): 4 within 40-59 cm. - Dry River nav is dynamic: path queries to the middle of the farmhouse and barn return partial paths after placement. **FILES CHANGED** `Content/Maps/L_DryRiver_01.umap` (tracked), `Content/Art/Environment/DryRiver/M_SS_CreekWater.uasset` (tracked), `Tools/Unreal/farm_life_s097.py`, `farm_life_s097b.py`, `Docs/ASSET_REGISTER.md` §4.9q, evidence under `Docs/evidence/`. **Untracked, local only:** `Content/Art/Environment/Fab/Props/` (400 MB: dead tree, fallen trees, branch, tractor, coop). Not committed because of size and ADR-021; the committed map therefore references assets a fresh clone does not have. **TESTING** - In-game `-game` captures after shaders settled (60 s): `Docs/evidence/s097_homestead_fp.png`, `s097_creek_fp.png` (an earlier capture at 18 s showed grey foliage and a white rifle: shader compilation, not a defect). Teleport used: `-SSExecAt=60 "-SSExec=EnableCheats|BugItGo x y z pitch yaw roll"`. - `audit_map_playability.py` with `SS_BUILDPATHS=1` (without it: "no walkable ground", expected): `Docs/evidence/map_playability_dr_s097.json`, **5 pass / 4 fail, unchanged**: crossing 16 m PASS, cover density 0.96 PASS, spawn exposure 0/64 PASS, starts PASS, close quarters PASS; max sightline 332.7 m FAIL (was 355), hard:soft 2.38 FAIL (was 2.12), Water Point parity 20% FAIL, Farmstead parity 13% FAIL (was 12%). - NOT RUN: persistence reopen of the saved map beyond the in-game load, PIE, performance (no frame-time measurement), C++ build, automation tests, multiplayer, bot match on the new layout. **ASSETS** ENV-007..011 and the compound, see register §4.9q. Remaining caveats: Fab windmill still has no vendor textures on disk; barn and old barn use pack timber/iron (their FBX shipped without textures); RedGum buildings are low-poly project blockouts with pack materials; tractor glb carries a green ground mat (buried 26 cm); kangaroo mobs are static meshes. **RISKS** R-83: committed Dry River map references untracked Fab props (400 MB). R-84: hard:soft cover and Farmstead/Water Point parity still fail; the compound adds solid mass near the Farmstead and was not re-balanced. **DEFECTS FOUND** Session 096 added about 300 random supply crates/sacks/barrels purely to lift cover density (found by viewing the map: dark boxes in open bush). Duplicate boulders at identical positions and a rock 4.3 m in the air (found by the floating audit). **NEXT ACTION** Walk the compound in PIE with bots (nav, collision, spawn-to-objective routes), then fix Farmstead/Water Point parity and the 332 m sightline by moving or adding mass, not props. ### Session 097 addendum — producer review round, MAF rifle, Red Gum life pass Producer review of the in-game Dry River captures: assets randomly placed, a tree base coming out of the creek bed, rocks sticking out, kangaroos too plentiful, creek water not like UE5 water and no creek bed, textures. Then: "move to other maps that are more complete ... Red Gum ... add quality Fab assets incl. Australian road signs, kangaroos under trees, windmill, tank, farm houses ... then promo shots ... OPFOR/MAF should be using the AKM model, not the AKS-74U". **Review of what I did before this addendum (stock-take)** - Dry River: objectives moved and persisted (Water Point (-1500,-600), Farmstead (3900,-1700)); audit 7/9 (`Docs/evidence/map_playability_dr_s097c.json` is in `Build/`, summary in the previous entry). Creek clutter cut (473 bed stones, 37 driftwood/rocks, jams and signs removed); kangaroos cut from 13 to 4, one pair and three singles, under map-edge trees; five wrecks replaced with the original RustyCarsFree meshes (`SM_asset_03/04`, the Ravenshoe wreck meshes rendered red/black noise); Namaqualand driftwood branch material given the intact branch textures (local, untracked pack); Fab texture compression fixed (`Tools/Unreal/fix_fab_prop_textures_s097.py`). - **NOT fixed:** the Dry River creek is still flat tinted planes, no carved bed, not UE Water. An attempted `WaterBodyRiver` + `WaterZone` did not render and was not saved. This is the producer's open complaint; it needs a carved terrain bed first. - Lost work found the hard way: edits made after an editor restart without a save were gone (kangaroo placement); the map must be saved before the editor is closed or killed. **Red Gum Station (`L_RedGum_01`), `Tools/Unreal/redgum_life_s097.py`, actors `SS_RG97_*` (67)** - Road centreline read from the landscape spline mesh components (30 segments, saved to `Build/redgum_road_segments.json`, local). - Four outstations on flat pads (measured relief): Hendry (10000,16000), Cole (-8000,-13000), Brennan (-20000,14000), Tully (20000,10000): RedGum farmhouse / huts / shearing shed / water tank, Ravenshoe barn and old barn (timber + corrugated iron slot materials), Fab water tower and Fab windmill, Fab tractor, Fab chicken coops, caravan, stone well, dunny, drums, wreck, log piles; trees and foliage instances inside each footprint cleared first. - Homestead (0,0): the procedural RedGum windmill replaced with the Fab windmill; added Fab tower, tractor, coops, old barn, well, dunny, drums, mailbox. - Eight RuralAustralia kangaroo signs placed along the road (facing copied from the existing eight); seven kangaroos under edge trees (two pairs plus singles), kept away from objectives, deployments and the road. - Fab tractor: its glTF carries about 60 green grass-tuft islands (ground patches) in the same mesh; removed in Blender (`Build/fab_fix/SM_Fab_Tractor_noplate.fbx`, local) and reimported over `SM_Fab_Tractor`; the Dry River tractor was raised 36 cm to match. - Captures (in-game, 60 s after load): `Docs/evidence/s097/rg_road2_b.png` (road, signs, outstation), `rg_hendry2_b.png` (Hendry yard, tractor without the green plate), `rg_hendry.png`/`rg_hendry_air.png`. The elevated "air" shots sit inside the canopy: Red Gum is dense forest, so promo shots must be ground level in clearings. **MAF rifle -> Fab AKM (`/Game/AK-47`, "AK-47" pack)** - `Tools/Blender/maf_weapon.py` gained `SS_MAF_SCALE` / `SS_MAF_TRIGGER_FRAC` / `SS_MAF_TRIGGER_Z` for static packs without a trigger bone. The static mesh `SM_AK-47` was exported (`Build/maf_weapons/SM_AK47.fbx`), scaled 0.78 to 0.876 m, origin at the trigger (36% from the butt), muzzle socket at the front; imported over `SM_MAF_R1` and `SM_MAF_S1` (support counterpart is the same mesh, as before) with the pack's `M_Exterior` / `M_Interior`. The AKS-74U source is backed up at `Build/maf_weapons/SM_MAF_R1_aks74u_backup.fbx` (local). - **Implemented but unverified:** grip/hand alignment and look on a MAF soldier in-game were not yet seen (the follow-bot capture was interrupted). `Art/Weapons/MAF/SM_MAF_R1.json` is stale (still names the AKS-74U). **FILES CHANGED (tracked)** `Content/Maps/L_RedGum_01.umap`, `L_DryRiver_01.umap` (already committed), `Content/Art/Environment/DryRiver/Farm/SM_Fab_Windmill` / `SM_Fab_WaterTower` (collision flag only), `Art/Weapons/MAF/SM_MAF_R1.fbx`, `.../Weapons/MAF/SM_MAF_R1.uasset`, `SM_MAF_S1.uasset`, `Tools/Blender/maf_weapon.py`, `Tools/Unreal/redgum_life_s097.py`, `fix_fab_prop_textures_s097.py`, `Docs/ASSET_REGISTER.md`. **Untracked, local only:** `Content/Art/Environment/Fab/Props/`, Namaqualand and RustyCarsFree material edits, `Build/*`. **NOT RUN:** Red Gum audit (`audit_map_playability.py`), nav check of the new buildings, bot match on Red Gum, persistence reopen, performance, C++ build, automation tests. **RISKS** R-85: Red Gum references the untracked Fab props folder, as Dry River does (R-83). R-86: Red Gum is dense canopy, so an aerial promo view is not possible and the 67 added actors were not nav-checked. **NEXT ACTION** Capture a MAF soldier in-game to confirm the AKM mesh/grip, then run the Red Gum playability audit and a bot match before promo shots. ### Session 097 addendum 2 — MAF rifle and camouflage after producer review - Producer: "AKM is only being seen as a grey rectangle. MAF is not MAF camouflage." Causes: the AK-47 pack's FBX carries a `UCX_` collision box that was merged into the display mesh (the grey rectangle), now removed in `maf_weapon.py`; and the MAF uniform materials had been reset to plain green by a later soldier rebuild (`apply_maf_camo.py` must be re-run after `setup_adf_soldier.py`). Producer later saw the AK render correctly in-game. - Producer supplied reference photos of the MAF dress: **cream ground with large rounded blobs of terracotta red, olive, khaki tan and brown**. `Tools/Textures/make_maf_uniform.py` palette and blob scale rewritten to match (the old pattern was dark and muddy). The producer then saw it in-game and reported "legs look structureless": the generator kept only fine detail against a 24 px blur, losing knee-pad shadows and pocket edges. It now keeps the green texture's full tonal structure (luminance / cloth median, exponent 1.25, clamp 0.30 to 1.25). Regenerated and re-applied (`Docs/evidence/s097/maf_pants_new.png`, `maf_pants_compare.png`). - **Implemented but unverified in-game:** the structure fix. Not done: MAF helmet, vest and belt remain flat olive; the reference shows a camo cap and olive tactical vest. The MAF camo is still an original pattern (Class F), not a copy of a real pattern. ### Session 097 addendum 3 — Red Gum audit and promo captures - **Red Gum playability audit** (`SS_BUILDPATHS=1`, `Docs/evidence/map_playability_rg_s097.json`): **3 pass / 7 fail**. Pass: close quarters, starts 8+8, spawn exposure 0/64, Homestead walk parity 4%. Fail: open crossing 30 m (target 20), max sightline 1010 m (the map's own extent), hard:soft cover 3.1, cover density 0.03, North and South paddock parity 12% each. Cover density and the ratio are artefacts: the audit counts only placed static meshes, and Red Gum's cover is foliage instances (trees, rocks, logs). Nav coverage 33% (R-12). Not re-tuned; the new outstations were not nav-checked. - **Promo captures** (in-game, 1920x1080, bot match; kept ones in `Docs/evidence/s097/promo_keep/`): MAF soldier running past a barn (new camo, AK, structure on the trousers visible), MAF squad under trees, 3 ACR soldier at a fallen log, road with kangaroo signs, Hendry farm with the tractor. Rejected: kangaroo pair (camera landed inside a tree), windmill (a tree in front); MAF helmet is still plain green and the AK is small in frame. All were taken with `God` and `ss.Debug.FollowBot`, third-person debug cameras, not a dedicated photo mode. - **NEXT ACTION** A proper photo mode or placed cameras for the hero shots (windmill and tank, kangaroos under a tree, a posed 3 ACR and MAF pair), then the Dry River creek bed. ### Session 097 addendum 4 — soldier legs ("no structure in them") - Producer on the MAF promo shots and in-game: legs weird on both models, helmets not sitting right, "doesn't appear as a person". Diagnosis by Blender front/side overlay of `SK_ADF_Uniform_G3` on the Manny reference (`Docs/evidence/s097/models/legs_front.png`, `legs_side.png`): the Session 096 tunic fit (`SS_GEAR_CAP=1.5 SS_GEAR_KEEP=0.05`) was applied to the whole uniform and shrink-wrapped the **trousers onto the mannequin's legs**, a skin-tight leg with no cargo-pocket or knee-pad volume. In the captures the friendly soldier's helmet and head sit correctly; the MAF helmet is a different asset (PASGT) and looks a little high. - Fix: `adfrc_gear_rig.py` gains `SS_GEAR_LEG_CAP`, `SS_GEAR_LEG_KEEP`, `SS_GEAR_LEG_Z` (default 0.95 m): below the hip the cap is separate. Rebuilt with tunic `1.5 / 0.05` and legs `4.5 cm / 0.6`. Reimported headless (`setup_adf_soldier.py`, `setup_soldiers.py`, `apply_maf_camo.py`: all `ok: true`). The S096 uniform is backed up at `Build/SK_ADF_Uniform_G3_s096_backup.fbx` (local). - Producer's re-check after the change: "seems better. some odd animations and blue botches on MAF uniform." The blue patches are not in the camo textures (0 bluish pixels in `T_SS_MAF_G3_Shirt_co` / `Pants_co`); cause not found. The animation oddities are not diagnosed. Open. - Producer also reported Dry River assets with no back faces (rock ledges seen from behind, floating slabs over the creek). Not yet fixed. ### Session 097 addendum 5 — full Quantum modular body for both sides (ADR-042 re-instated) Producer: "option 1 ... full quantum body for both aussie and MAF, then ADFRC textures, webbing, helmets over the top; alternative helmets / headsets / faces for the different classes too." - C++ (`SSTeam`): `ASSCharacterPartActor` gains `OpposingRetargetParts`, `bRetargetOpposingPose`, `OpposingRetargetMaterialOverrides`; the retarget tick is now `TickRetargetSet`, run for the friendly or the opposing set by viewer locality (the MAF Quantum modules are retargeted from the pawn pose exactly like the friendly ones). Built clean. Renamed `CVarHandIK` to `CVarFPHandIK` in `SSFirstPersonSubsystem.cpp`: it clashed with `SSHandIKMeshComponent.cpp` under an adaptive unity build (`check_unity_names.py` passes). - Assets (`Tools/Unreal/setup_quantum_soldier.py`, `Build/quantum_soldier_setup.json` ok): 3 ACR = Quantum head, rolled-sleeve shirt, trousers and arms in the **real ADFRC AMCU print** (`T_SS_Tile_AMCU`, a seamless patch cut from the G3 shirt sheet by `Tools/Textures/make_camo_tiles.py`), TBAS vest and Ops-Core helmet over it; MAF = the same Quantum modules in the original cream / terracotta / olive / khaki / brown print (`T_SS_Tile_MAF`), PASGT helmet and Peacekeeper vest over it. The G3 uniform is no longer used by either side. - Producer's in-game screenshots of the friendly soldier after the change show a person with a proper head, neck, hands and a fitted helmet: a clear improvement. Visible faults: the left hand is held open (the retarget's finger curl does not read), the AMCU tile has a faint stitch line and wrinkle shading repeating, and the soldier has no boots modelled by Quantum (boots come from the jeans module's feet). - MAF: renders and retargets (verified only on a prone/dead bot in `Docs/evidence/s097/models/q_enemy_1.png`); a standing MAF has not been captured and the PASGT/Peacekeeper fit on the Quantum torso is unmeasured. - **Variants plan (not built):** the pack has one head only, so per-class faces need more heads (Quantum cap `SM_Cap_Bege`, the Modern Insurgent heads, ADFRC boonie/PASGT/OpsCore helmets, headsets are baked into the Ops-Core). Next step is a data table of loadout slots (helmet, headwear, face, vest) selected by class. ### Session 097 addendum 6 — friendly rank markers (first GUI pass item) Producer: a small chevron over each friendly soldier with the rank slide, the service level under it, and the player name only within 15 m or when the crosshair is on them. - New: `USSFriendlyMarkerState` (Core, client-local list of teammates), filled twice a second by `USSScoreboardSubsystem` (bridge); `USSFriendlyMarkerWidget` (SouthernSpearUI) draws a pool of 24 markers from live pawn positions. Same-side players only (viewer-relative, never replicated, ADR-017); bots, which have no service record, get a stable stand-in level 2..60 from their name, flagged `bPlaceholderLevel`, so the markers can be judged in bot matches. - First version flickered (producer: "keeps flickering around"): slots were reassigned by list order, the line-of-sight dimming toggled per frame and the auto-sized slot re-laid out. Fixed: a slot is owned by its soldier, line of sight is sampled 6 times a second and the opacity eased, fixed 180 x 96 slot, 30 Hz position smoothing. Captured: markers hold position across burst frames (`Docs/evidence/s097/gui/marker4_sheet.png`). Then enlarged, put on a dark plate with a brighter chevron and the anchor lowered to the head (`marker5_1.png`). - Seen also in the producer's scoped screenshots: the markers draw over the scope view; not yet suppressed. Junior ranks have no insignia device; the plate shows the number only. - `-SSShotUI` makes `-SSShotTimes` captures include the HUD. - NOT RUN: network/dedicated test, markers on the class-select, performance with 15 teammates, the broader GUI visual pass (HUD plates, menus, scoreboard), which is the next item. ### Session 097 addendum 7 — HUD pass 1, smaller floating rank markers, AMCU crispness, hand grip Producer: base the GUI on the website and on what competing products do well; rank slides smaller, floating, no background. Then: AMCU "not as crisp and clean as the last model"; holding-the-barrel animation looks strange in third person. - **Rank markers:** plate removed, slide 22 px (was 34), level 11 pt, name 12 pt, chevron 10 x 8, shrinks with distance, text shadow only. Captured `Docs/evidence/s097/gui/hud1.png`: legible and unobtrusive on sand. - **HUD pass 1 (`SSPlayerHudWidget`):** health is now a large number with a 20-cell segmented bar on a light plate with a brass edge; ammo shows the weapon name, a 44 pt magazine count, the reserve and a row of magazine pips (one per round, capped at 30, going clay when the magazine is low); a pulsing clay edge wash below 30 % health. Palette unchanged from `Site/styles.css`. NOT done yet: compass, minimap and objective panel restyle, scoreboard, menus, hit markers, kill feed polish, stance and stamina readouts. - **AMCU crispness:** the tile was a 455 px crop upscaled to 1024 (soft) and cross-faded across its whole width (ghosting). Now a native-density 512 px tile with only a 14 % edge feather, tiled 5.0 / 5.5 times (was 2.2 / 2.6): smaller, sharper pattern. In-game look is greener than the old G3 sheet because the tile is cut from the G3 shirt's front panel, which is the greener part of the print; not yet colour-matched. - **Hand grip:** logged the retarget (`SSArmSolve`): the Quantum hand is already at the mannequin hand's position to 0.0 cm, so the problem is hand orientation and fingers, not reach. Added an experimental two-bone arm solve and hand-orientation match behind `ss.Char.ArmSolve` (default 0: it did not read better and I could not verify it); raised the finger curl from 55 to 80 degrees per joint. **The grip is not verified fixed**: the follow-camera captures catch the bots in reload, draw and sprint poses, so no clean standing-hold frame was seen. Needs a deliberate capture of a bot holding the handguard. ### Session 097 addendum 8 — left-hand reach, HUD pass 2 - **Grip (third person).** The mannequin's left wrist is solved onto the weapon's grip socket only when the socket is within `MaxReachFactor` (1.05) of the arm's length; beyond that the solve was skipped and the hand stayed in the animation pose, floating off the handguard (sprint/run frames, `Docs/evidence/s097/models/grip1_sheet.png`). The solve already clamps at full reach, so `MaxReachFactor=1.6` now lets it reach for the grip. Result: the hand lands on the weapon in 3 of 4 frames (`grip2_sheet.png`); the fourth is a reload/draw clip, where the IK is suppressed by design. Not yet checked: crouched and prone holds, other weapons (only the A88 has an authored nudge/tilt row; the AKM, PKM and others use the defaults). - **HUD pass 2** (`Docs/evidence/s097/gui/hud2.png`): rank markers hide while a magnified optic is up; the objective panel no longer repeats "round 1 starts in 7:14" twice before the round; compass strip loses most of its box and gets larger labels, a bolder centre notch and a brass heading; minimap gets a brass frame and is toned down; health, ammo and objective plates are darker so they read over bright sand. - **Not done:** scoreboard, kill feed, hit markers, menus, front end restyle; no in-scope capture of the marker hide (code path only). ### Session 097 addendum 9 — AMCU print, per-class gear, MAF camo cap/vest, SVD and PKM, kill-feed names, boots - **AMCU (`Tools/Textures/make_camo_tiles.py`).** The tile was a feathered crop of the G3 sheet and still carried wrinkles and a stitch line, so each repeat ghosted ("not as crisp and clean as the last model"). Cropping, flattening and mirroring were tried (mirroring reads as ink-blots, `Docs/evidence/s097/models/amcu_tile_new.png` is the kept result). Now: the six colours are measured from a clean patch of the ORIGINAL ADFRC print (k-means) and the shapes rebuilt as layered periodic noise, cut sharply, with a fine weave. In-game it is crisp and clean (`gear1_sheet.png`). It is a rebuilt print in the original's colours, not the original pixels: the sheet cannot tile. - **Per-class headgear and body armour (3 ACR).** New `ISSCosmeticRoleReceiver` (Core), `FSSRoleGear` + `FriendlyRoleGear` map on `ASSCharacterPartActor` (Team), and `USSLoadoutSubsystem::Grant` tells the pawn's part actors the class after granting the kit. Fitted with `adfrc_gear_rig.py`: Rifleman OpsCore + TBAS (PC); Medic TeamWendy + TBAS CFA; Machine gunner Exfil + TBAS MG; Sniper boonie + TBAS Base; Grenadier OpsCore + TBAS CQB. The BRH converted to the NVG mount only and was dropped. Verified in game: log shows each bot's class gear, frames show the different helmets and rigs (`gear1_sheet.png`). Host/standalone only: kit choice is not replicated (R-22). - **MAF.** Helmet cover, belt and vest fabric wear the MAF blob camouflage instead of flat olive (`MI_MAF_CamoGear`, `gear2_maf_sheet.png`); blotches tightened (tiling 1.6/1.8 → 3.4/3.8, gear 3.0); AMCU tiling 5.0/5.5 → 3.0/3.3 once the print read too fine (`amcu_scale3.png`; kill feed confirmed in game: "AKM" and "SVD" for the other side, `boots-170.png`, `boots170.png`). Boots: the jeans mesh is one material with the boots in its top atlas band, so camouflage covered the boots; `apply_quantum_boots.py` builds `M_SS_QuantumTrousers` (boot band V < 0.245 draws the pack's leather texture). Run it after `setup_quantum_soldier.py`. - **MAF SVD and PKM** (producer drop, `Content/SVD`, `Content/PKM`): `maf_weapon.py` -> `Art/Weapons/MAF/SM_MAF_{SVD,PKM}.fbx` (1.225 m / 1.21 m, muzzle +X, side renders `SM_MAF_*_side.png`), `setup_maf_support_weapons.py` imports them (SVD PBR set at 2048; PKM flat metal/wood/brass/ammo-box colours: the pack has no textures). The viewer-relative swap now maps A25 -> SVD, A89 -> PKM, the rest -> AKM. PKM seen in MAF hands in game; **SVD not yet seen in game**. Neither has a LeftHandGrip socket, so the left-hand IK does not run for them (hands follow the animation). - **Kill feed.** For a viewer on the other side the killer's weapon reads AKM / PKM / SVD (A9 pistols unchanged). Display text only. Builds; not yet seen in game. - **Scoreboard** checked in game (`Docs/evidence/s097/gui/scoreboard1.png`): already in the site palette with brass rules, legible, no change needed; the rank column shows "–" for bots (the rank-marker placeholder level is not shown there yet). Kill feed shows "AKM" for MAF killers. - **Not done / next:** menu, front-end and hit-marker restyle; per-class faces; Quantum cap module (textures present in the pack) for a boonie/cap on non-helmet classes; MAF per-class gear; grip sockets for the MAF weapons. ### Session 097 addendum 10 — headwear variants, camera recoil, F1 grenade model - **Headwear.** `FriendlyRoleGear` is now a set of options per class (`FSSRoleGearSet`); each soldier wears one, picked from the part actor's name so it is stable. Five more ADFRC pieces fitted with `adfrc_gear_rig.py` (Airframe, Exfil B, CVC, balaclava; the 135k-vertex OpsCore MT was skipped). Rifleman: OpsCore / Airframe / Exfil B / boonie; Medic: TeamWendy / boonie / Airframe; MG: Exfil / Airframe / CVC; Sniper: boonie / boonie + balaclava / Exfil B; Grenadier: OpsCore / CVC / Exfil. Seen in game (`headwear1.png`). Faces are still the pack's single head. - **Recoil.** New in `Core`: `FSSRecoilRules` (stance scale: aiming 0.7, crouched 0.85, moving 1.2; burst build-up +8 % a shot to 1.4; recovery), per-weapon `RecoilPitchDeg / RecoilYawDeg / RecoilRecoveryDegPerSec` on the weapon rows (`Config/DefaultGame.ini`; A88 0.55/0.22/7, A89 0.50/0.35/5, A417 0.95/0.30/5.5, A25 1.25/0.25/4, A9 1.10/0.35/9). **These are tuning choices, not measured or sourced** (see `Docs/NEXT_PRIORITIES.md` §5). `ASSCharacter::ApplyShotRecoil` (on the local player's fire cue only, bots untouched) adds the kick to the control rotation and 65 % of the climb settles back (`ss.Recoil.Scale`, 0 = off). Lyra's character does not tick, so `ASSCharacter` now enables its tick. Tests: `SouthernSpear.Core.Weapons.Recoil` passes (20 Core tests, 0 failures). In game, `SSRecoilTest 12` (new dev exec) logged a climb of 0.55 → 6.73 deg over 12 shots and a settle to 3.0 deg. **Not tested:** real shots with a real weapon (the exec calls the same function without firing), feel, per-weapon balance, server-side effect of recoil on hit registration. - **Grenade.** Grenade is **G** and melee **V**; Q/E are lean (`setup_tactical_movement.py` moved them off Q), so there is no clash. Lyra's `B_Grenade` is untouched; `USSGrenadeVisualSubsystem` swaps its mesh for the ADFRC F1 grenade (`Art/Weapons/Grenade/SM_F1_Grenade.fbx`, 6 x 6.4 x 10 cm, `setup_grenade.py`, colour and normal maps; `ss.Grenade.VisualScale` 1.3). Verified: `SSThrowGrenade` (new dev exec) activates `GA_Grenade`, the log shows the swap. **Not verified:** the model is hidden in the trail effect in flight (`grenade2.png`, `grenade_crop.png`), so its look is unseen; there is no first-person throw animation (Lyra has none), and the spoon, pin and fuse effect are not modelled. - **Next:** a held-grenade view/animation and a visible grenade for third person; recoil tuning with real firing; per-class faces; MAF per-class gear. ### Session 097 addendum 11 — playtest fixes, baked G3 uniform, flags, crosshair - **Playtest findings (producer):** recoil felt absent (the fire cue did not reach the local player: recoil now follows the magazine count), snipers inaccurate (A25 `SpreadScale` 0.75 -> 0.2), Lyra hit markers (new `USSCrosshairWidget`: own crosshair, white X on a hit, clay X on a kill; Lyra reticles collapsed by class name), Minimi textures (45 weapon materials moved to `M_SS_WeaponPBR` with normal and SMDI maps; the Elcan lens was a white disc and is now dark glass), class preview showed one helmet (preview now applies the class gear), second silhouette in the owner's shadow (mannequin no longer casts). - **Uniform textures ("not the quality you have on file").** The tiled camouflage cannot carry pockets, knee pads, seams or the print's true scale. `Tools/Blender/bake_g3_to_quantum.py` bakes the authentic ADFRC G3 uniform (AMCU) onto the Quantum shirt, trousers and arms UVs (the two rigs share bone positions exactly, so no alignment). First attempt was black: clearing the parent dropped the armature transform; fixed with world-matrix preservation. In game the soldiers now show G3 pockets and knee pads (`bake_ingame2.png`). **Known weaknesses:** the bake is nearest-surface so some areas stream (stretched pattern), sleeves are muted, gaps are dilated rather than re-baked, and the look is flatter than the Arma reference (no separate under-shirt, no wrinkle normals). - **Flags.** `SK_ADF_FlagPatches` (Blender, curved patches on both upper arms, ADFRC flag texture) is in the class gear; not yet seen in game. - **Machine crash, 2026-10-05 ~08:17 (event log review).** A whole-system reset during a capture run (`head2.log` ends 08:17:54; Kernel-Power 41 with bugcheck 0 at the 08:36 reboot, no minidump written) with a burst of GPU watchdog reports (WER LiveKernelEvent 141, 116, 117, 193, 1a8). Same GPU on the same machine (AMD RX 9070 XT, driver 32.0.31044.16) blue-screened three times before with bugcheck 0x116 VIDEO_TDR_FAILURE (2026-07-26, 09-26, 09-28). Not an engine crash: this points at the graphics driver or hardware, with the game's GPU load as a trigger rather than a proven cause. Not yet tried: driver update or rollback, lighter capture settings. The last run's output (head close-ups) is unverified. - **Not fixed:** helmet clipping through the head (the fitted helmets were sized to the mannequin head), A88GL animations, per-class faces. ### Session 097 addendum 12 — baked detail maps, EF88 magazine, shoulder patch - **Uniform detail.** The G3 normal (`_nohq`) and gloss (`_smdi`) sheets are baked onto the Quantum UVs too (`bake_g3_to_quantum.py` `SS_BAKE_CHANNEL`), colour contrast lifted 1.3x and gaps dilated (`Tools/Textures/finish_quantum_bake.py`), and the AMCU materials moved to the cloth master (`M_SS_FabricPBR`). Close-ups (`heads2.png`) show smeared fabric where the bake maps by nearest surface; not solved. - **Shoulder patch.** The G3 sheet has two black diamonds where a flag velcros on, so the shoulder rendered as a dark hole. They are filled with sleeve camouflage (`T_ADFRC_G3_Shirt_AmcuCamo.png`, original kept in `Build/`). The flag mesh was inside the sleeve (radius 5.5 cm against a measured 7-9 cm); rebuilt at 9.2 cm, 8.5 x 5.5 cm, V flipped onto the flag. **Not confirmed in game:** every capture caught the arm in motion. - **EF88 magazine.** The ADFRC EF88 model has an empty magwell. `adfrc_weapon.py` now seats the pack's own 30-round AUG magazine (`SS_MAG_BLEND`, set per weapon in `build_adfrc_weapons.py` `MAGAZINES`) at the `magazine_axis` memory point; A88 and A88G rebuilt and reimported (`SM_A88_side.png`). **Not done:** the reload animation. The magazine is part of the static mesh, so it stays in the rifle during a reload; making it come out and go back needs it as a separate mesh moved by the reload montage. - **A88G (launcher) hold.** Its `LeftHandGrip` socket is 28.4 cm forward, 6.3 cm to the left and 3.8 cm low (A88: 22.2 / 9.4 / 1.5), but only the A88 has an authored palm tilt and nudge in `DefaultGame.ini`, so the A88G hand sits where the A88's would. Not yet adjusted. ## **Addendum 13 - Wandarra (the MOUT town) made playable.** The Fab MOUT pack ships no finished map (its demo is a cube arena, the other two are asset showcases); Wandarra is our own layout from its pieces. Defects found in game and fixed in `build_wandarra_level.py`: ground slab, player starts and nav volume were placed 100x too deep/high (metres vs cm); nav volume covered only 150 m of the 300 m site (bots reported 'start point not on navmesh', now 0); ground used the VT master that rendered solid red (now the Ravenshoe flat terrain instance); fences ran 90 degrees off; roads did not exist (now 11 bitumen/footpath slabs). `wandarra_dynamic_nav.py` sets the navmesh to build at load, so no attended bake is needed. Verified in game: bots path and fight toward objective A, fences correct. NOT DONE: it is still sparse (no trees beyond 38, no Australian trees, wrecks, logs, barricades), roads not yet seen from above, not in the front-end map list. **Addendum 14 - Wandarra dressed (producer: 'no roads, no buildings, fences wrong way, limited vegetation').** Imported the owned Fab Concrete Barrier, Metal Barricade and two Military Trenches sandbag meshes (`import_wandarra_fab_props.py`, `/Game/Art/Environment/Wandarra`; no props were created). `build_wandarra_level.py` now places 30 barricades/sandbag nests, 11 wrecks and drums, 16 logs, Australian gums + grass trees (RuralAustralia pack replaces the European beech) and 28 extra street-frontage buildings (17 bungalows, 11 houses in total). Added dev exec `SSFly x y z` for overview captures (top-down shot confirmed roads, buildings, trees). Wandarra is in the front-end operation list (no loading art yet). Fixed a unity-build clash (`Gap`) in SSCrosshairWidget.cpp. NOT DONE: interiors/doors, tuning of tree density and cover, loading-screen art, attended look review. ## Session 098 — 2026-10-05 — Wandarra on real Australian ground, three-map watchtowers, and the five-day roll-up (Sessions 094-098) Producer: "ground textures / grass isn't rendering ... don't use any Blender assets or things you've created, there should be ample assets in our content library ... a watch tower ... on all maps, but not too many ... classic Aussie things in Wandarra, a kangaroo, a windmill on the corner ... update the changelogs with everything achieved over the last few days to give a big update to the website." **COMPLETED** - **Wandarra ground, roads, grass (map rebuilt from `Tools/Unreal/build_wandarra_level.py`).** The 300 m ground is now the RuralAustralia pack's own world-space grass instance (`MI_Ground_Grass_01`) on a slab, replacing the custom grass material built earlier in the session. Roads use the Fab "American Road with Parking Lot" Asphalt04 texture set on the AutomotiveBridge triplanar master (the previous cracked sheet read as white blotches). Ground cover is about 3,500 grass clumps (PN Grass Library) and 200 shrubs and flowers, now including the Namaqualand shrub and flower meshes. Before/after: `Docs/evidence/s097/wandarra/`. - **Australian landmarks in Wandarra, all from existing library assets:** the Fab windmill with its wheel (south-west lot), the Fab water tower (north-east, by the park), a RuralAustralia kangaroo road sign on the south entry, seven Fab kangaroos around the outskirts, plus the earlier gums, grass trees, fallen logs, wrecks and Fab barricades. - **Watchtowers (new Fab asset, `Tools/Unreal/import_watchtower.py`, `place_watchtowers.py`).** One on Dry River's ridge, one in the Red Gum timber, two on Wandarra (east edge, south-west scrub): few, on purpose. The glb imports as separate pieces sharing one scene offset; every placement spawns them at one transform so they reassemble. - **Capture tooling.** `Tools/quick_shot.sh` opens the game, takes one screenshot, closes it (no more long-running windows); `Tools/capture_wandarra.sh` uses it. - **Roll-up of the last five days (Sessions 094-098) for the website:** - *Soldiers:* full Quantum modular body for 3 ACR and MAF with baked AMCU (colour, normal, gloss) and MAF tiles; per-class gear and per-soldier headwear (airframe, Exfil, CVC, boonie, balaclava); Australian flag patches on both shoulders; gloved first-person arms that sit lower in the view; slimmer tunic and legs. - *Weapons and feel:* camera recoil from per-weapon rows, recoil following the magazine count, tighter sniper spread, left-hand grip reach, finger curl, EF88 magazine seated, SVD and PKM for MAF with kill-feed names, F1 grenade model. - *HUD:* segmented health, ammo pips, compass, minimap frame, objective panel, friendly rank markers, new crosshair and hit markers, class gear in the class preview. - *Maps:* Dry River homestead, outstations, fallen timber, kangaroo mobs and creek water (Session 097); Red Gum outstations and road signs; Wandarra (the MOUT town) playable with roads, fences, doors, barricades, Australian trees, dynamic navmesh and bots pathing, and in the front-end operation list. **FILES CHANGED** `Tools/Unreal/build_wandarra_level.py`, `setup_wandarra_surfaces.py`, `import_watchtower.py`, `place_watchtowers.py`, `wandarra_dynamic_nav.py`, `Tools/Common/wandarra_spec.py`, `Tools/quick_shot.sh`, `Tools/capture_wandarra.sh`, `Content/Maps/L_Wandarra_01.umap`, `L_DryRiver_01.umap`, `L_RedGum_01.umap`, `Content/Art/Environment/Wandarra/Materials`, `Content/Art/Environment/Watchtower`, `Site/index.html`, `Docs/evidence/s097/wandarra/`. Untracked (not committed): the new vendor packs (KiteDemo, Namaqualand, Nanite plants and others), as before. **TESTING** - `UnrealEditor-Cmd ... -ExecutePythonScript=Tools/Unreal/build_wandarra_level.py` exit 0; `Build/wandarra_level.json` all steps ok, no missing assets, grass_clumps 3483, shrubs 203. `setup_wandarra_surfaces.py` ok. `place_watchtowers.py` ok (15 pieces per tower). - In-game captures (`Tools/quick_shot.sh`, 20 s): ground, roads, grass, windmill and watchtowers confirmed in `Docs/evidence/s097/wandarra/{top,street,park,windmill,tower,tower_dryriver,tower_redgum}.png`. - NOT RUN: C++ build and automation suite this session (no C++ edits after the last commit); bots on the rebuilt Wandarra with the new props; Dry River and Red Gum navigation after the tower pieces (tower collision is the pieces' default). **ASSETS** Fab "Old Wooden Watchtower (House 3)" imported to `/Game/Art/Environment/Watchtower`; Fab Asphalt04 textures to `/Game/Art/Environment/Wandarra/Materials/Textures`. Reused: RuralAustralia ground MI, Fab windmill, water tower, kangaroo, Namaqualand foliage. Nothing newly modelled; the Blender barrier script was not used. **RISKS** R-98: Wandarra, Dry River and Red Gum reference untracked vendor packs (as R-83/R-85). R-99: a RuralAustralia Landscape (`MI_Landscape_Example_01`, any layer set, two grid sizes) rendered the engine checker grid in `-game` while the same material renders in the pack's own map; cause not found, so Wandarra uses a ground slab (no runtime grass layers). R-100: the watchtower is reached only by ladder and Lyra has no ladder climbing, so bots and players cannot use the towers yet. **DEFECTS FOUND** Checker-grid ground on the Landscape (producer screenshot). Roads blotchy white (producer screenshot). Capture runs left the game open for minutes (producer: "the player is in the same spot"): fixed by `quick_shot.sh`. **NEXT ACTION** Roadmap item 2: character visual acceptance (uniform detail, helmet clipping, faces, MAF gear), then the weapon material audit. Open Threads | Item | Blocked on | Owner | |---|---|---| | ~~First editor launch of the renamed project~~ | **CLOSED** — G0.10 passed | — | | ~~NavMesh validation for Dry River~~ | **CLOSED** — G1.1 passed, 560 tiles, path verified | — | | **Dedicated server target build (R-09)** | **Producer decision — this engine distribution cannot build Server targets at all. See `PROJECT_AUDIT.md` §6.1 and producer question 5** | **Producer** | | Fab account / engine registration (R-03) | Producer decision | Producer | | Second client machine for 4-client test (R-05) | Producer decision | Producer | | ~~Insignia legal clearance (L-0003)~~ | **CLOSED** — hold lifted by ADR-035 (free-to-play; R-57 accepted) | — | --- ## Conventions for Future Entries One section per work session, newest at the bottom. Always: - **COMPLETED** — exact changes, not intentions - **FILES CHANGED** — created/modified/removed, with the vendoring status of anything untracked - **TESTING** — command, result, or `NOT RUN`. No exceptions - **ASSETS** — created/imported/licence entries/placeholders remaining - **RISKS** — new risks get IDs continuing from R-08; closed risks are struck through, not deleted - **DEFECTS FOUND** — with how they were found, since that is the real signal - **NEXT ACTION** — exactly one, the highest-priority item ## Session 099 — 2026-10-05 — Documentation reconciliation and roadmap review after Session 098 Producer: update the asset register, the backend documents and the changelog, then look at the roadmap. **COMPLETED** - `Docs/ASSET_REGISTER.md` new section 4.9r: Session 098 assets (watchtower ENV-012, Wandarra windmill/water tower/kangaroos ENV-013, ground ENV-014, Fab asphalt ENV-015, Namaqualand dressing ENV-016, UE5 C++ skills TOOL-001) and the new untracked packs not yet integrated (MCO Mocap Basics, MHBOP Lite, Screen Damage Indicator, KiteDemo, STF Landscape Pro, PKM/SVD). - `Docs/PROJECT_AUDIT.md`: risks R-98, R-99, R-100 added with the Session 098 wording; `Docs/MAPS_WANDARRA.md` section 3.2 (the Australian rebuild); `Docs/DEVELOPMENT_ROADMAP.md` current status; `Docs/NEXT_PRIORITIES.md` status block. - Roadmap review (below). No game code, assets or maps were changed this session. **ROADMAP: where to from here** (Phase 1 vertical slice; Phases 2 to 6 not started) 1. *Close the slice's playability gap:* R-100 watchtower access, attended R-82 nav bake for Wandarra and Ravenshoe, Red Gum nav coverage (17%), then a bot match on each of the three playable maps. 2. *The producer's art order:* character visual acceptance (uniform detail, helmet clip, faces, MAF gear), weapon material audit, then VFX (muzzle flash, tracers, impacts). 3. *Weapon feel:* A88G hold, EF88 reload animation, zero and recoil from real data. The new MCO Mocap and MHBOP packs are the likely source for hold, reload and locomotion animation. 4. *HUD:* integrate the Screen Damage Indicator as a C++ presentation widget in `SouthernSpearUI`. 5. *Maps:* use KiteDemo/Namaqualand open-world foliage and decide the Landscape question (R-99, try STF Landscape Pro); Wandarra loading art. 6. *Housekeeping that gates release:* untracked vendor packs (R-98), asset provenance for the new packs, CI and the R-09 server-target limit. **FILES CHANGED** `Docs/ASSET_REGISTER.md`, `Docs/PROJECT_AUDIT.md`, `Docs/MAPS_WANDARRA.md`, `Docs/DEVELOPMENT_ROADMAP.md`, `Docs/NEXT_PRIORITIES.md`, `Docs/CHANGELOG.md`. **TESTING** Documentation only. NOT RUN: any build, test or in-game capture. **ASSETS** Registered only (see 4.9r); no new imports. **RISKS** None new (R-98 to R-100 recorded in the audit). **DEFECTS FOUND** Roadmap status in `DEVELOPMENT_ROADMAP.md` section 13 was last updated 2026-09-26 and still read "Phases 2-6 not started" without the Session 098 state; fixed. **NEXT ACTION** Make the watchtowers usable: add a climb volume (or ramp from an owned asset) to the tower ladders and a bot route, then verify a sniper can hold the east Wandarra tower in a bot match. ### Session 099 addendum 1 - roadmap item 1: watchtowers usable, nav on every map, bots verified (2026-10-05) - **Watchtowers are now climbable (R-100, player side).** `Tools/Unreal/add_tower_access.py` gives each of the four towers a plank ramp (engine cube dressed in the tower's own `Wood_Planks`, 28 degrees, direction chosen by ground traces) with a flat landing deck over the platform floor, and removes collision from the lattice, fascia ring and top rail so the ramp can step onto the floor. Root cause of the earlier "cannot get in": the Interchange import gave every tower piece a solid auto-box as simple collision (a triangle trace passes through it, a pawn does not), so the 186 x 186 x 667 frame was a solid block. The script sets `CTF_USE_COMPLEX_AS_SIMPLE` on all 15 meshes (assets re-saved). The tower's own slabs are open frames/hatches, which is why they were not used for the ramp. - **Verified in game by walking** (new dev command `SSWalk x y z yaw seconds [endPitch]` in `ASSCharacter`, drives the pawn's velocity so it collides and steps like walking): Wandarra east tower climbed from the ground to the deck (feet z 412), crossed and stepped off the far side; Red Gum and Dry River towers likewise (log lines in `Saved/Logs`, evidence `Docs/evidence/s099/ramp_*.png`). Wandarra's second tower got the same script but was not walked separately. - **Runtime navmesh on all five objective maps** (`Tools/Unreal/dynamic_nav_all.py`: `RuntimeGeneration = Dynamic`), so no attended bake is needed (R-82 worked around, not fixed) and new geometry such as the ramps enters the navmesh. - **Bot matches** (`Tools/bot_check.sh `, 8 bots, headless): Wandarra captured objective 0 and moved to 1 (150 s); Dry River moved to objective 1; Ravenshoe bots fought (7 kills, 120 s); Red Gum captured objective 0 after a long walk (300 s, slow, large map). No "not on navmesh" or path-failure lines on any map. NOT verified: that a bot actually climbs a tower ramp. ### Session 099 addendum 2 - roadmap item 2 (characters), capture tooling, first promo set (2026-10-05) - **Shoulder defect fixed.** Producer-visible black patch with white shards on the friendly shoulder: the G3 source sheet has no printed camouflage across the shoulder yokes, so the Quantum bake left a smooth brown smear. `Tools/Textures/fix_quantum_shoulders.py` fills those zones (about 25 % of the shirt sheet) with camouflage cut from the well-baked torso, luminance-matched, feathered; normal and gloss neutral there. The shirt/jeans/arms were also re-baked with a longer ray distance (`SS_BAKE_RAY=0.3`). In game (`Docs/evidence/s099/chars/shoulder_b3.png`) the black patch is gone; the pattern on the shoulder is still stretched and smeary at 1 m, so **texture detail on the uniform is improved, not solved** (needs a proper retexture or a different source shirt). - **Flag patch.** The patch showed the black hook pad, not the flag. V range corrected in `adfrc_flag_patches.py` and the patch enlarged to 8.5 x 5.5 cm at 9.2 cm radius. **Not confirmed**: the shoulder captures show the patch area but the flag is not legible at that size. - **Helmets and heads.** Checked at 20 to 40 cm in game (3 ACR and MAF): no helmet clipping through the head seen; faces read as faces. MAF camo is the pale-cream pattern from last session, left alone. - **Weapon check.** All five classes captured with HUD hidden (`Docs/evidence/s099/weapons/class*.png`): textures present on A88, A88G (grenadier), A89, sniper rifle, with the optic's lens showing sky. The A89's optic lens shows a baked image of a vehicle (reads as an odd sphere). Not changed. - **VFX.** `SSFire` (new dev command) shows the Lyra muzzle flash, smoke and the code-driven muzzle light on the A88 and impact sparks, dust and a visible tracer line on the ground (`Docs/evidence/s099/vfx/`). So flash, tracer and impacts exist and render; scale and tuning are not judged. - **A88G grip.** First-pass `GripNudgeCm` row for the A88G added from the changelog estimate (Forward -8, Right 1.7); the left hand is not visible in the stills so the hold is **unjudged**. - **Capture tooling.** `SSOrbit`, `SSWalk`, `SSFire` dev commands (`ASSCharacter`), `ss.UI.Hide`, and `Tools/promo_shots.sh` (1920x1080 stills: character full/portrait for both sides, four weapon classes, map overview and the three towers, HUD) plus a 18-frame GIF of firing (`Docs/evidence/s099/promo/gameplay_fire.gif`, 960x540, 6 MB; not for the site as is). - **Not done:** weapon texture audit beyond visual, EF88 reload animation, zero/recoil data, damage-indicator integration (the Screen Damage Indicator pack is a demo map plus a blood overlay texture; the HUD already has a hit-direction arrow and low-health wash). ### Session 099 addendum 3 - per-optic scope reticles, MAF uniform bake, promo moved to Red Gum (2026-10-05) - **Scopes (producer: every optic showed the same view).** The ADS scope overlay now shows the optic's own ADFRC reticle: Spectr on the A88/A88G, TA31 ACOG on the A4/A416, C79 ELCAN on the A89, TA648 on the A25 (`Tools/Textures/make_reticles.py` -> `Art/Weapons/Reticles/`, imported to `/SouthernSpearUI/Textures` by `import_reticles.py`; `USSPlayerHudWidget` picks by weapon name, generic stadia only as fallback). Verified in game on A88 and A25 (`Docs/evidence/s099/weapons/scope_class0.png`, `scope_class3.png`). **Not done:** a true 3D scope (a render target of a magnified second view shown on the optic's lens, with the reticle in the lens material). It is feasible with a `USceneCaptureComponent2D` enabled only while aiming a magnified optic, at the cost of a second scene render; recommended as the next step if the overlay is not enough. A4/A416/A89 reticles not yet captured; the A25's scope image is also still a full-screen overlay, not the lens. - **MAF uniform (producer: pockets, knee rolls).** The MAF Quantum shirt and trousers were a flat tiled camouflage. They now use the same bake as the 3 ACR uniform, from the MAF-repainted G3 sheets (`bake_g3_to_quantum.py` with `SS_BAKE_FACTION=MAF`, `finish_quantum_maf.py`, shoulder fill via `fix_quantum_shoulders.py` `SS_FACTION=MAF`): chest and thigh pockets, knee pads and weave show in game (`Docs/evidence/s099/chars/maf_new.png`). The bake still smears at 1 m (same limit as the AMCU one) and has some grey blotches from the shoulder fill. MAF forearms (Arms module) were left as skin. - **SVD (producer: no textures in the promo).** `MI_MAF_SVD` has all four of its maps bound and the UV sampling is consistent, but the texture is dark (wood, brown steel) and in deep shade it reads as plain dark brown. **Not resolved**: no lit close-up of an SVD was obtained (bots mostly carry the AK); needs a dedicated viewer or a kit override, and probably a brighter wood. - **Promo** moved to Red Gum (`Tools/promo_shots.sh`). Red Gum looks far better than Wandarra, but the characters and weapons landed in deep canopy shade; the set needs sunlit clearings picked per shot before it is usable. ### Session 099 addendum 4 - MAF uniform pass 2, promo set published, crash note (2026-10-05) - **MAF uniform.** Re-baked with a shorter ray distance (0.15) and brightened 1.22x (`finish_quantum_maf.py`): the pocket flaps, thigh pockets and knee pads now read in the lit shots; red/olive/cream pattern kept (`Docs/evidence/s099/chars/maf_check1.png`, promo `char_maf_full.png`). Knee **rolls** (raised knee panels as geometry) are not modelled: the Quantum jeans mesh has none; only the printed pads show. Forearms and the vest are unchanged. - **Promo set** (`Tools/promo_shots.sh`, `Docs/evidence/s099/promo/`): Red Gum HUD and tower, Dry River tower, both factions on Wandarra (Red Gum's canopy and ground clutter hid the soldiers, so the character shots are taken on Wandarra), and a 4.7 MB firing GIF on Red Gum. Seven stills published to the site gallery (`Docs/images/screenshots/`, `screenshots.json`). The weapon-class stills were not published (a power pole stands in the frame). - **3D scopes** noted as a later item in `Docs/NEXT_PRIORITIES.md` section 10. - **Crash (producer-reported).** A capture run died in `UnrealEditor_Core` `InheritedContext.cpp:130` (refcount assertion) from a DevHttp thread, i.e. the editor's analytics/HTTP thread at shutdown or startup, not our code; `Saved/Crashes/UECC-...`. `quick_shot.sh` now passes `-unattended` (no crash reporter dialog left open) and kills the editor on completion. A crash-reporter window left open is what looked like a hang. - **Dev tooling added:** `SSGather` (teleport bots to a spot), orbit camera avoids walls and trunks. Not for shipping builds. ### Session 099 addendum 5 - weapon stills, SVD/PKM textures fixed, ten weapons on the site (2026-10-05) - **Root cause of the untextured SVD (producer: "no textures on it").** `M_SS_MafGunPBR`, the master for the MAF SVD and PKM, **failed to compile** in game: the log said `(Node TextureSampleParameter2D) Param2D> Found NULL, requires Texture2D` and "Default Material will be used in game", so both guns rendered the engine's default grey. The maps and parameters were correct; the master had no valid default texture on every parameter. `setup_maf_support_weapons.py` now falls back to a project texture when an engine default does not load, and `SS_REBUILD_MASTER=1` recreates the master. Verified: SVD shows its wood furniture and blued steel, PKM its wood and olive ammunition box (`Docs/evidence/s099/promo/weapon_MAF_SVD.png`, `weapon_MAF_PKM.png`). Lesson recorded: a material that fails to compile does not error in the editor commandlet; grep the game log for `Failed to compile`. - **Weapon stills.** New dev command `SSShowMesh [fov]` floats a weapon mesh above the pawn and frames it with the orbit camera (negative distance = multiples of the bounds radius). `Tools/promo_shots.sh weapons` renders all ten weapons (A88, A88G, A89, A4, A416, A25, A9, MAF SVD, PKM, AK) against the sky; the earlier first-person weapon shots with a power pole in frame are gone. - **Site.** Ten weapon renders added to the in-engine gallery (17 captures total, each captioned and dated; the weapon ones are labelled "Model view", not gameplay). Checked in a local preview before publishing: all 17 images load. - **Not done:** the firing GIF is not on the site (the gallery takes stills only); knee rolls on the MAF trousers; EF88 reload animation; zero and recoil from real data. ### Session 099 addendum 6 - firing clip on the site (2026-10-06) - The Red Gum firing GIF (18 frames, 800x450) is now an animated WebP (1.7 MB, against 4.7 MB as a GIF) and leads the in-engine gallery, labelled "In-engine clip". `Tools/build_site_assets.py` copies a `.webp` manifest entry as is and records `animated` in `Site/data/screenshots.json`; `Site/site.js` plays it in a loop with no resized derivatives. Gallery is 18 items. Checked in a local preview (all 18 images load; the clip is 800x450). The clip is about a second long and silent. ## Session 100 — 2026-10-06 — Bullpup hold and reload, magazines, recoil and zero from real data Producer: "continue with what you suggested. if we can get the GIF on the website, that would be good", then "noticed that the M4, HK416 dont have magazines. We need to work on the animations for holding the EF88, should be different between the EF88 and the GL variant ... once that has been sorted and the reloading improved via animations - then the zero and recoil from real data would be great", and the HUD concept (squad roster bottom left, 8v8, two squads of four). **COMPLETED** - **Firing clip on the site** (Session 099 addendum 6; the clip's web file `Site/assets/screenshots/shot-gameplay-fire.webp` was missing from that commit and is in this one). The three weapon stills for the A4, A416 and A25 were replaced with new ones that show their magazines; the gallery is 18 items. - **Magazines in the A4, A416 and A25.** The producer saw empty magwells on the site stills. `MAGAZINES` in `Tools/build_adfrc_weapons.py` now seats the pack's `PMAG30` (A4, A416) and `MSS_AR10_pmag` (A25) from the source's `magazine_axis` points, as the A88 already had. Checked in game (`Docs/evidence/s100/magazines_a4_a416_a25.png`). - **Magazines are parts.** `Tools/Blender/adfrc_weapon.py` tags the seated magazine and, besides `SM_.fbx`, exports `SM__NoMag.fbx` (the body, with the same sockets) and `SM__Mag.fbx` (the magazine alone, origin at the top of the well), plus `MagWell` and `MagWellEnd` sockets. `setup_weapons.py` imports them. Only the first-person view uses them; the weapon in the world keeps the whole `SM_`. - **First-person bullpup reload, A88 and A88G** (`USSFirstPersonSubsystem`, `USSHandIKMeshComponent`, new `USSFirstPersonSettings` in `Config/DefaultGame.ini`). The Fab M4 reload clip sent the left hand to an M4's magazine position and tipped the weapon out of view (`reload_a88_before_pack_m4_clip.png`). A weapon with a `LeftHandReload` row now keeps its right hand on the grip in the idle pose, cants and pushes the rig forward on a per-weapon envelope so the underside shows, drives the left wrist along a path through the magwell, the pouch and back (new `SetBoneSpaceTarget` IK override), and moves three magazine components: out of the gun into the hand, falling away, a fresh one in the hand, seated. It is timed to Lyra's reload montage, which stays the gameplay authority. Other weapons are unchanged (the pack's clip). - **Hold, A88 against A88G.** The pack's idle left-hand pose is where an M4's handguard was: at the muzzle of an EF88 and nowhere near the launcher. Both rows now use left-hand IK to their grip socket with the existing per-weapon hold and nudge: the A88 under the handguard just ahead of the trigger guard, the A88G cupping the launcher (`GripNudgeCm` A88G is now -14.0 / 1.7 / 3.5; it was -8.0 / 1.7 / 0.0, judged too far forward and low). `bLeftHandIK` is per weapon; `ss.FP.HandIK 1` still does it for all. - **Recoil from data, and a ballistic zero** (`Docs/WEAPON_BALLISTICS.md`, generated by `Tools/Weapons/weapon_ballistics.py`). The ADFRC configs carry the zero (`discretedistance {100,200,300,400}`, init index 2, which is 300 m; pistols 25 m); sight heights are measured on the meshes; muzzle speeds are the Army F88 page (930) and the HK416 page (790), otherwise typical figures; the A25 and A89 zeros are design values. Recoil per shot is cartridge impulse over ready mass, scaled so the A88 keeps its 0.55 degrees; recovery follows the ADFRC `inertia`. **The recoil rows are applied** (R-103). The **zero is built and off**: `FSSTrajectoryRules` (Core), a second Lyra hook D-10 in `TraceBulletsInCartridge`, `ss.Ballistics 0` by default. ADR-043 is Proposed for the producer. - **HUD concept and 8v8** recorded in `Docs/images/hud_concept_squad_roster.webp`, `NEXT_PRIORITIES.md` sections 6.2 and 10, and the GDD's HUD section. Not built. - `Tools/weapon_import_chain.sh` runs the weapon re-import in the order that keeps materials and sockets (see defects). **FILES CHANGED** `Plugins/SouthernSpearLyraBridge`: `SSFirstPersonSubsystem.h/.cpp` (settings, reload driver, mesh variants, debug cvars `ss.FP.DebugOrbit`, `ss.FP.DebugReloadLog`, and a `Reload@seconds` delay on `ss.FP.DebugClip`), `SSHandIKMeshComponent.h/.cpp`, `SSBallisticsSubsystem.cpp`, `Tests/SSFirstPersonSettingsTests.cpp`. `Plugins/SouthernSpearCore`: `SSWeaponStats.h/.cpp`, `SSTrajectory.h/.cpp`, `Tests/SSTrajectoryTests.cpp`. `Source/LyraGame/Weapons/LyraGameplayAbility_RangedWeapon.h/.cpp` (D-10). `Config/DefaultGame.ini`. `Tools/Blender/adfrc_weapon.py`, `Tools/build_adfrc_weapons.py`, `Tools/Unreal/setup_weapons.py`, `add_sight_sockets.py`, `probe_mag_parts.py`, `Tools/run_anim_sequence.sh`, `Tools/weapon_import_chain.sh`, `Tools/Weapons/weapon_ballistics.py`. `Art/Weapons/{A4,A416,A25,A88,A88G}/ADFRC/*` and the A-series Game Feature weapon assets (re-imported, material chain re-run). `Docs/*` (CHANGELOG, ASSET_REGISTER 4.9s, PROJECT_AUDIT R-101 to R-104, DECISION_LOG ADR-043, LYRA_ADOPTION D-10, NEXT_PRIORITIES, GAME_DESIGN_DOCUMENT, DEVELOPMENT_ROADMAP, WEAPON_BALLISTICS), `Docs/images/*`, `Docs/evidence/s100/*`, `Site/*`. Untracked on purpose: unchanged from Session 099. **TESTING** - `Build.bat SouthernSpearEditor Win64 Development ...`: exit 0, `Result: Succeeded`, after every change including the Lyra D-10 patch. - `UnrealEditor-Cmd ... "-ExecCmds=Automation RunTests SouthernSpear;Quit"`: exit 0, **72 `Result={Success}`, 0 failures** (68 before). New: `SouthernSpear.Bridge.FirstPerson.Settings` reads the shipped A88 and A88G rows back; `SouthernSpear.Core.Trajectory.Zero|Drop|Path` check the crossing at the zero, a 300 m zero's near zero at 28 to 40 m (the service M4's own 36 yard / 300 yard shape), drop against the textbook 4.9 cm, and that no model gives no path. - `python Tools/validate_architecture.py` PASS; `python Tools/check_unity_names.py` PASS; `python Tools/test_architecture_guard.py` 0 failures. - In game (editor `-game`, Wandarra and Dry River, 1280x720 and 1920x1080): reload sequences `reload_a88_first_person.png` and `reload_a88g_first_person.png` (14 frames each), left-side hold views `hold_a88_vs_a88g_left_side.png`, and the magazines on the A4, A416 and A25. The A88G reload ran through the same rows after `SSKit 4`. - `ss.Ballistics 1` with `SSFire 3` on Wandarra: log `LogSSTrajectory: A88: 11 legs to 250 m, path 9.6 cm off the line at the end` (above the sight line, short of the 300 m zero), no errors. - **NOT RUN:** a hit-position check of the ballistic path against a target at range; bots under `ss.Ballistics`; any playtest of the reload, hold or new recoil by the producer; the third-person reload; the reload path for any weapon other than the A88 and A88G; dedicated server (R-09). **ASSETS** Registered in `ASSET_REGISTER.md` 4.9s: WPN-002 (A88 and A88G `_NoMag` and `_Mag` meshes), WPN-003 (magazines on the A4, A416, A25), DATA-001 (ballistics), UI-002 (HUD concept). ANIM-002 (MHBOP Lite) corrected: it holds no animation. **RISKS** R-101 first-person reload tuned by eye; R-102 ballistic zero built and off (ADR-043 Proposed); R-103 recoil rows re-derived; R-104 bots and other players see the old reload. R-99, R-100 and the open Session 099 items stand. **DEFECTS FOUND** - The A4, A416 and A25 shipped with empty magwells (the producer, from the site stills); fixed. - `setup_weapons.py` on its own re-points every weapon material at the colour-only master and drops the `Sight` sockets (found when it re-imported the A-series: the A4 would have lost its detail maps and ADS its sight line). The fix is the chain `setup_weapons`, `upgrade_weapon_materials`, `upgrade_weapon_detail`, `add_sight_sockets`, `fix_visual_regressions`, now `Tools/weapon_import_chain.sh`. It was run for all seven weapons this session; the A89 and A9 were checked in stills. - `Site/assets/screenshots/shot-gameplay-fire.webp` was published but not committed with Session 099 addendum 6; committed now. - `WEAPON_SOURCE_DATA.md` names the A25's recoil class `recoil_ebr` (the SR-25 base class); the item classes use `recoil_spar_dmr`. Noted in `WEAPON_BALLISTICS.md`; the generator was not changed. - Session 099's NEXT ACTION (watchtowers) was already done by its own addendum; this entry replaces it. - An edit script that hard-coded CRLF line endings failed on LF files and changed nothing (its own assertion caught it); the source files are LF in git (`attr eol=lf`). **NEXT ACTION** Producer plays the A88 and A88G (hold, reload, the new recoil) and decides ADR-043 (`ss.Ballistics`); the reload for the other weapons and the HUD squad roster follow that review. ### Session 100 addendum 1 - HUD restyled to the producer's concept; first engine baseline (2026-10-06) Producer: "the next thing is looking at the UI to see how close we can replicate the fonts / style / colours to the concept art, as I think that looks far more premium" (`Docs/images/hud_concept_squad_roster.webp`), and then the engine's optimisation and back end "to improve the quality of graphics". - **Colours** sampled from the concept and added as HUD tokens (`SSPalette`, `SSUIStyle`): slate-blue translucent panels (`HudPanel`, `HudPanelDeep`), near-white text, a gold accent for the heading and the objective, blue squad badges. The brand palette is untouched (menus, website). - **Layout** now follows the concept. Minimap top left, dark and desaturated (post-process plus a tint on the capture), thin slate frame, N/E/S/W on its edges, the map's name and the round clock under it (`SSObjectiveStatusWidget`). Objective panel top right: an OBJECTIVES header strip over a slate body with a check box, the objective's name and its state in plain words, the chip row, a capture bar and the score. Compass: a faint slate band, fine ticks every 5 degrees, labels fading to the edges, the heading in gold between two rules (`SSCompassWidget`); the active objective also gets a gold diamond with its distance in the world. Bottom right: the weapon's white silhouette and name, the magazine count, a divider and the reserve, a bar for the magazine. Bottom left: health, quiet, over a four-row squad roster with blue numbered badges, rank and name over an underline. - **Weapon silhouettes:** `Tools/Blender/render_weapon_icons.py` renders each A-series weapon flat white from the side (`Art/UI/WeaponIcons/`), `Tools/Unreal/import_weapon_icons.py` imports them to `/SouthernSpearUI/Textures/Icons/T_SS_Icon_`. - **The roster is not squads yet.** Until the match has squads (`NEXT_PRIORITIES.md` 6.2) it is the four most senior friendly soldiers from the scoreboard state, best rank first, rank abbreviation from ADR-034 and the player's or bot's name. - **Not matched:** the typeface. The concept's is a squared, semi-bold sans with open tracking, closest to Rajdhani; the HUD still uses Barlow Condensed and Inter. A font file needs the producer's say-so to download (not yet asked for or fetched). The equipment slots (grenades) in the concept's ammo block are not drawn: nothing feeds them. Health is not in the concept; it is kept, small, above the roster. - **Evidence:** `Docs/evidence/s100/hud_before.png`, `hud_after3.png`, `hud_compare.png` (concept above, game below). - **Tests:** build exit 0; `validate_architecture.py` PASS; `check_unity_names.py` PASS. NOT RUN: the full automation suite after the UI change (no UI logic changed that it covers), controller or ultrawide layouts, 4K DPI scaling, the full-screen map and the scoreboard against the new palette. - **Engine baseline (measured, first look only):** editor `-game`, Dry River, 8 bots, 1920x1080 windowed, AMD Radeon RX 9070 XT. `ProfileGPU` frame 7.81 ms (about 128 fps GPU-bound figure). Heaviest passes: RasterPasses (Nanite and virtual shadow maps) 2.03 ms (26 %), volumetric cloud view 0.82 ms (10.5 %), a parallel draw group 0.56 ms, TSR 0.34 + 0.26 + 0.05 ms (the history is 3840x2160: TSR is upsampling from a 1080p render), virtual shadow map projection 0.18 ms. Config: Lumen with hardware ray tracing on, virtual shadow maps, Substrate on, TSR. **Not measured:** game and render thread times (the CSV profiler wrote nothing with `-csvCaptureFrames` alone), VRAM, a lower-end GPU, shader compile hitches, packaged builds. The target hardware and frame budget are not decided; that decides everything else. ### Session 100 addendum 2 - HUD font and the left-hand reach (2026-10-06) - **Font.** Rajdhani Medium and SemiBold (SIL OFL 1.1, from Google Fonts' `ofl/rajdhani`; licence text `Art/Fonts/Rajdhani-OFL.txt`) are now the in-game HUD face (`SSFonts::Hud`, `SSWidgetKit::HudText`): player HUD, compass, minimap, objective panel, round banner. Menus, scoreboard and the website keep Barlow Condensed and Inter. Imported by `Tools/Unreal/setup_fonts.py`. - **Hand on the handguard** (producer: "the hand on the barrel is not aligned"). Cause measured in game: the Fab arms' left shoulder is 44 cm left of the bore, so the forearm crossed the weapon at a steep diagonal, the wrist target was beyond reach (the IK clamped: moving the nudge changed nothing) and the hand lay open beside the handguard. Fix: `USSHandIKMeshComponent::ShoulderShiftCm` (config, component space; `X=-26` pulls the left shoulder toward the centre) applied before the wrist solve in the hold only, plus `FingerCurlDeg=25` (finger joints bent about their own flexion axes, thumb 35 %). The reload path does not use the shift (it pushed the arm out of frame). Evidence `Docs/evidence/s100/hud_compare.png`, `hand_a88g.png`. Tuned by eye on the A88; the A88G hand sits at the launcher, also unjudged (R-101). - **Tests:** build exit 0; 72 `Result={Success}`, 0 failures; architecture and unity-name checks PASS. NOT RUN: A4/A416/A25/A89 holds (they use the pack's pose, not this IK), 4K, controller layouts. ### Session 100 addendum 3 - website sweep before sharing, and PlayTest.bat (2026-10-06) Producer: "do another sweep of the website, add a few more ingame pics ... remove the old weapon models the old AI agent put in there and just give it a nice boost before I share it around"; earlier, "can you have a .bat file created that spins up the game for me to test?" - **Old weapon renders removed.** The Loadout section no longer uses the 28 Sep Blender studio renders (A88, A89, A4, A416, A25 and the AKM reference render). `Site/assets/weapons/`, `build_weapons()` in `Tools/build_site_assets.py` and its README paragraph are gone. Loadout is now seven A-series weapons (A88, A88G, A89, A4, A416, A25, A9) and three MAF weapons from the in-engine captures in `Docs/images/screenshots/weapon-*.png`, each with muzzle speed and zero from `Docs/WEAPON_BALLISTICS.md`. Those captures carry `"gallery": false` in `screenshots.json`, so they appear in Loadout and not twice in the Media gallery. - **New "In the engine" band** under the hero: the new HUD on Red Gum with the A25 (`redgum-a25-hud.png`, a 2000x1119 capture the producer took at 1440p; it replaced my own A88 capture because the A88 hand hold in it looked wrong), the firing clip and three facts. The old Red Gum HUD capture was dropped from the gallery (superseded). - **Two site-script defects fixed.** `sectionOf` in `Site/site.js` only read `#` headings, so the Development cards and the latest session showed "Not recorded" and "None recorded" since the changelog moved to bold `**NEXT ACTION**` labels; it now reads both. The roadmap showed Phase 1 as "Planned" because the dated rows "Phase 1 at 2026-10-05..." overwrote the Phase 1 row; the first row now wins. - **Wording.** "All ... weapons are fictional" removed from the page, footer and 404 (ADR-035: real weapon names are in use); units and forces stay fictional. - **`PlayTest.bat`** in the repo root: opens Dry River windowed 1920x1080 with 8 bots (`PlayTest.bat `), refuses to start while an editor is running. - **Tests:** `python Tools/publish_site.py --build-only` exit 0; page checked in the browser pane at desktop and 375 px (no horizontal overflow, images load, development cards populated). NOT RUN: Lighthouse, Safari/Firefox, real phone. - **Not done:** fresh in-engine captures. The producer's playtest was open (`UnrealEditor.exe`), and `Tools/quick_shot.sh` kills every editor process, so no new capture was taken; the "ingame pics" added are the newest existing ones (HUD, magazines on the A4/A416/A25 already in the gallery). - **Risks / defects:** none new. The Maps section still prints the 2026-10 playability audit (Red Gum "not playable as it stands", Dry River "fails six of nine rules"); it is accurate to the audit but dated, and worth a refresh before wide sharing. - **NEXT ACTION:** unchanged (producer plays the A88/A88G and decides ADR-043).