security-game/docs/3d-model-review.md

3D model review (2026-11-03)

Status: baseline measured; scorecard open. This is the working document for the unversioned 3D model re-evaluation track in TODO.md. It records what was measured, what the Blender skill guidance says the targets are, the ranked candidate gaps, and the decisions that still need the project owner. Nothing here re-opens the accepted 0.9.4 limits by itself.

Related documents: npc-visual-review.md (accepted 0.9.4 limits and before/after sheets), npc-proportion-audit.md (proportion guide), blender-assets.md (asset contract), asset-tooling.md (commands).

Method and skill routing

The review is screenshot- and measurement-backed, following qa-review (screenshot set, defect severity, explicit verdict) rather than an opinion pass:

Family	Skills loaded	Criteria taken from them
Routing and pipeline	blender-director	Plan → blockout → detail → validate → export; MCP steps translated to headless flatpak runs
Characters	character-artist, retopology, rigging	Eye/mouth/jaw/shoulder loops, quad-dominant deformation surfaces, 3-loop joints, no triangles on primary deformers
UVs and textures	uv-workflow, texture-workflow, materials	Texel-density tiers (hero 512–1024, standard 256–512, background 128–256 px/m), consistent density, no overlapping UVs
Budgets and structure	asset-optimization, lod-pipeline	Triangle/vertex budgets, naming, LOD tiers
Props, gear, foliage	prop-artist, hard-surface, vegetation-artist	Silhouette read, bevel/chamfer language, collision proxies, carded foliage
Look and verdict	lookdev, qa-review	Day/night beauty captures, ship / ship-with-notes / no-ship verdict

Project contract and gates stay authoritative; see AGENTS.md §Local Blender skills. The skills assume Blender MCP, which this project does not use: everything below came from the headless tools/blender_assets.py / tools/npc_pipeline.py pipelines plus the Godot capture harnesses.

What was captured and measured (2026-11-03)

Evidence	Command	Result
Contract audit	python3 tools/blender_assets.py audit --json	46/46 assets clean; library 59,012 tris
Imported-GLB inspection	blender_assets.py inspect (all 46)	All clean; transforms applied, helpers ignored
Gear fit	python3 tools/blender_assets.py fit --json	Head width 0.315 m on all eight; clothing radius 0.190–0.260 m
Proportions	python3 tools/npc_proportions.py report --json	All eight inside the stylized guide
Texel density	python3 tools/blender_assets.py texels --json	New measurement, table below
Neutral baseline	capture_npc_baseline.gd → tmp/model-baseline (48 frames) + npc_baseline.py sheet	48 frames, none blank; sheet in tmp/ only
Final context	capture_npc_review.gd → tmp/model-review (96 frames)	96 frames, none blank; raw frames stay local
Performance	smoke_npc_budget.gd	24 avatars, 168 meshes, 131,616 tris, 6.94 ms of 50 ms

Raw captures and JSON stay in the gitignored tmp/; committed review images remain the docs/blender_previews/sheet_*.jpg set from 0.9.4/0.9.5 until a fix actually lands.

Character measurements

Contract numbers from audit + inspect; density from texels (256² slot maps, authored and with the runtime 4× fabric repeat).

Character	Tris / verts	Height / feet	Materials	Authored td (px/m)	Fabric effective (px/m)	Verdict
npc_casual	5,038 / 4,756	1.824 / +0.000	7	106.8–122.2	429–440	Contract pass; face prototype
npc_jacket	5,428 / 5,071	1.824 / +0.000	7	100.3–113.7	409–413	Contract pass; face prototype
npc_coat	6,030 / 5,802	1.824 / +0.000	7	75.3–81.4	307–310	Contract pass; face prototype
npc_vest	5,536 / 5,375	1.819 / +0.000	9	85.5–99.3	362–366	Contract pass; face prototype
npc_dress	5,456 / 5,128	1.825 / +0.000	7	81.9–89.0	338–339	Contract pass; face prototype
npc_hoodie	5,860 / 5,492	1.856 / +0.000	7	89.5–98.6	359–364	Contract pass; face prototype
npc_firefighter	5,900 / 5,902	1.819 / +0.000	9	64.0–75.1	287–290	Contract pass; face prototype (helmet shadow hides the lower lip)
npc_guard	5,602 / 5,684	1.819 / +0.000	11	90.0–100.8	374–374	Contract pass; face prototype
npc_base (fallback)	1,204 / 2,488	1.836 / +0.000	5	70.1–72.9	4× on fabric	Contract pass; far below its 6,000/4,000 cap

Blender-side inspect counts can differ by ±1 vertex from the glTF accessor count in audit (normal-split rounding); both are recorded and neither changes a verdict.

Readings:

• Every character is inside 8,000/8,000 with the farthest at 5,474 tris (68% of budget) and 5,494 verts (69%), so there is headroom for targeted topology detail without a cap change.
• Surface area tracks clothing volume: the firefighter coat is the largest body (4.28 m²) and has the lowest authored density; npc_casual is the smallest (2.44 m²) and the highest.
• Authored density (64–122 px/m) sits below the skill's background tier (128–256 px/m). The runtime tint controls most of the look: skin, hair, accessory and the fixed detail slots are neutral maps tinted at runtime, not painted skin, so density mostly controls subtle weave/noise. Fabric is effective 287–440 px/m (standard tier) because of the 4× repeat. This is a recorded note, not a defect, until a portrait shows texture starvation.
• Material counts 7–11 per character are fixed detail slots (eyes, rubber, hivis, reflective, leather, metal), not extra tint slots.

Scorecard skeleton

Verdicts use qa-review semantics: P0 (blocker: wrong scale, broken normals, budget fail), P1 (major: silhouette/readability), P2 (minor), accept (recorded limit). — means not yet reviewed; this document is the place to fill it.

Characters

Character	Silhouette	Face @1.85 m	Hands/grip	Weights/pose	Cloth/hair	LOD	Verdict
(pose evidence blocked: see "Pose evidence is blocked" below)
npc_casual	accept (0.9.4)	accept (face prototype)	accept (open palm)	blocked	accept	accept (two-quad impostor)	pending owner
npc_jacket	accept	accept (face prototype)	accept	blocked	accept	accept	pending owner
npc_coat	accept	accept (face prototype)	accept	blocked	accept (hem 0.76 stride)	accept	pending owner
npc_vest	accept	accept (face prototype)	accept	blocked	accept	accept	pending owner
npc_dress	accept; performer props dominate	accept (face prototype)	accept	blocked	accept	accept	pending owner
npc_hoodie	accept (hooded criterion)	accept (lower face only)	accept	blocked	accept	accept	pending owner
npc_firefighter	accept	accept (face prototype)	accept	blocked	accept	accept	pending owner
npc_guard	accept	accept (face prototype)	accept	blocked	accept	accept	pending owner

Gear (18)

Gear	Attach	Tris	Verdict
cap	head	328	— (fit measured, visual pending)
hood	head	804	—
sunglasses	head	164	accept (0.9.4 clearance)
firefighter_helmet	head	672	accept (0.9.4 + 3 cm runtime offset)
backpack	back	380	accept (0.9.4 strap refit)
shoulder_bag	back	188	—
scba_tank	back	362	—
lanyard	chest	100	accept (0.9.4 + 4 cm runtime offset)
phone	hand	96	accept (forward offset)
cup	hand	188	accept
clipboard	hand	84	accept
umbrella	hand	134	accept
camera	hand	318	accept
protest_sign	hand	92	accept (performer/protester read)
flashlight	hand	168	accept
wand	hand	244	accept
radio	hand	220	accept

Gear total: 5,002 tris across 18 pieces (all far below the 1,500-per-piece cap).

blender_assets.py fit readings (2026-11-03): shoulder_bag is clear on both npc_coat and npc_dress (the only back-gear clearance the fit test asserts); sunglasses sits inside the head band at 0.64× head width. Three caveats worth recording, none of which changes a verdict: backpack (npc_casual) and scba_tank (npc_firefighter) report CLIP because back gear is measured against the hip/skirt radius, which is wider than the upper back where the bags actually sit; cap and firefighter_helmet report band=OUTSIDE because they sit on the crown above the 1.55–1.81 face/head band (expected); and the lanyard clip is corrected at runtime by GEAR_OFFSETS (+4 cm forward). The unused beanie was retired on 2026-11-03 (removed from GEAR, its builder, npc_avatar.gd and the manifest).

Props (20 library + 6 generated fallbacks)

Library total 13,608 tris; marquee (2,804) and street_tree (2,592) are the two heaviest and are the first candidates for a vegetation-artist/lookdev pass.

Prop	Tris	Verdict
bollard	138	accept (2026-11-03 flared base collar)
bench	560	—
trash_bin	228	—
planter_box	1,324	—
newspaper_box	300	accept (2026-11-03 recessed slot + door seam)
fire_hydrant	610	accept (2026-11-03 rebuild: bolted flange, bonnet cap nut, larger side + front outlets)
street_lamp	428	—
street_tree	2,592	P1 candidate (flat-shaded clumps)
palm_tree	400	P1 candidate (same foliage language)
round_planter	1,240	—
a_frame_sign	212	—
wall_lantern	180	—
traffic_cone	210	—
crowd_barrier	388	—
stanchion	434	—
handrail	268	—
security_podium	184	—
double_door	380	—
marquee	2,804	— (emissive check pending)
sedan	928	—

Vision review 2026-11-03 (tools/describe_image.py on the prop gallery): bench and trash_bin are ship-ready (minor chamfer polish only); bollard passes but leans on its reflective band (add a base collar); planter_box/round_planter fail on the foliage — the _shrub() clumps read as rocks, not plants (same foliage language gap, gated on the owner); newspaper_box is borderline (the slot is a flat decal, wants a recessed inset + door seam); fire_hydrant is now rebuilt (see row). No prop floats or sinks; shadow direction is consistent.

Generated fallbacks (assets/models/generated/): bench, bollard, fire_hydrant, newspaper_box, planter_box, trash_bin. They are only used when the library prop is missing, so they are audited for presence, not re-reviewed for art quality.

Venues

Venue	Measured (map-budget test)	Status
Hollywood Boulevard	58,884 tris / 409 meshes	shape-language review pending
The Joke Shop	26,828 tris / 580 meshes	facade review landed: 6.0 m low hall, white signature band, tied-back curtains, west hotel tower; remaining gap = rounded west corner
TV Studio	—	shape-language review pending

Pose evidence is blocked (2026-11-03)

The weights/pose column cannot be filled yet. The pose logic is verified; the capture is not. src/test/capture_npc_pose.gd stages idle / walk-contact / walk-pass / seated for all eight variants (full-body + shoulder views) and its output does not match the state it reports.

Check	Result
Pose reaches proxies	Seated pose logs rig_y=-0.339, leg0=-1.357
Pose reaches skeleton bones	HipL/KneeL bone rotations differ from rest; headless test test_pose_reaches_the_skeleton_bones locks it (suite 273/273)
Idle restores rest	Same test asserts the hip returns to its rest orientation
Skinning bound	One Skeleton3D, mesh bound to it (same=true); hiding npc_base removes the body
Bone rotations deform the mesh	Direct writes change rendered pixels: Spine 1,719 / HipL 1,052 / KneeL 810 / ElbowL 745 / Head 671 / ShoulderL 640 sampled pixels
Bone translation deforms the mesh	A +0.5 m Head offset moves the head in frame
Standalone window probes render the pose	Three variants (avatar under root, under a Node3D stage, created mid-_process) render the seated crouch (diff ~2,200-2,360 samples)
Offscreen capture renders the pose	A SubViewport + RenderingServer.force_draw() probe renders the crouch reliably (diff 602 at +2/+6/+12 frames), including a re-created second avatar
The harness renders the pose	No in its current form: same stage, same pose, rest frames. One ordering (_frame_view() immediately before force_draw()) produced posed walk frames once, then regressed to rest/garbage on the next run

The rig, the animation math and the offscreen render path are all sound. A 2026-11-03 re-attempt with the recipe (SubViewport + force_draw() + force_update_all_bone_transforms() + disabled process/physics) still produced byte-identical rest frames, and a decisive probe narrowed the failure further: teleporting the avatar changes the render (diff 10), but moving the Rig child (the GLB model's parent) does not (diff 0). The skinned mesh therefore does not follow its Rig Node3D transform, so the rig lowering never reaches the rendered model — the defect is upstream of camera/force_draw() ordering. Next attempt must answer why the skinned model ignores the Rig node, or stage poses via the known-good capture_npc_review.gd fallback.

Working recipe (verified by the probes; keep it when rebuilding):

1. Create a SubViewport (1280x720, own_world_3d = true, UPDATE_ALWAYS) in _initialize.
2. Build the studio (environment, key/fill light, ground, camera) and one avatar inside it.
3. Wait 2-3 frames so the skinned mesh binds.
4. Apply the pose with _animate() ticks, with the avatar's process/physics disabled.
5. Call RenderingServer.force_draw(), then viewport.get_texture().get_image().
6. Free the avatar and repeat for the next pose; a re-created avatar also rendered correctly in a probe.

Ranked candidate gaps

Ordered by expected player-visible gain per unit of work. None is approved for fixing yet; the owner review decides.

1. Face plane structure at portrait distance (P1, all eight). The neutral sheet shows flat, faceted cheek/brow planes and slit-like eyes at 1.85 m. character-artist/retopology guidance asks for concentric eye loops and a 3-loop mouth. This is the largest readability gap and also the most expensive (head rebuild + re-UV + re-AO per variant). Needs a targeted single-variant prototype before committing to eight.

   The eight library characters already carry the 0.9.4 face detail in tools/blender_assets.py._face(): sculpted chin plate + cheekbones (_shape_head_mesh), socketed sclera + thin upper lid, nose bridge/tip/wings, a rounded mouth and ear concha (the plain sphere head in tools/npc_pipeline.py is the frozen npc_base fallback only). So this gap is a readability/scale question on the existing sculpt — the head is seg=18/rings=12 with small offsets — not missing geometry. Concrete next step, if funded: a single-variant npc_casual prototype that raises head subdivision and deepens the cheek/brow offsets, then re-bakes AO and captures the portrait before touching the other seven.

   Prototype landed 2026-11-03/04: npc_casual now has seg=24/rings=16 (smoother), a subtle brow ridge, a fuller lower lip and both eyelids (upper + lower); recessed creases (nasolabial, temple, chin cleft) were rejected because they bake as dark AO patches. Vision-reviewed: the new face reads cleaner and more realistic than the old (set-in eyes, brow break, lip volume, no faceting); tris 4,482 → 5,038 (still under 8,000). The other seven variants keep the old face until this prototype is approved.

2. No pose/weight evidence (unknown). Idle, walk, sit, block and turn deformation has never been captured as a review set, so shoulder/hip/neck collapse is unmeasured. Cheap to add and it gates any claim that the rig is acceptable.

3. Foliage language (P1, street_tree, palm_tree, planter_box, round_planter). Flat-shaded clumps against an otherwise readable set, already deferred once. vegetation-artist carded clusters would fix it inside the existing triangle budget.

4. Texel density below the background tier (note). 64–122 px/m authored vs the skill's 128–256 background tier. Fabric is already standard-tier at runtime. Decide: accept because the maps are neutral tinted detail, or raise only skin/hair to 512² for portrait distance.

5. npc_firefighter density floor (P2). 64 px/m with the largest coat area; split the coat UV shell or raise the map size before the next character pass, whichever the density decision says.

6. Hands at conversation distance (accept, revisit only if face work opens the budget). Open palm with HAND_ITEM_FORWARD; no finger bones exist, so a curl rig is a contract decision, not a modelling tidy-up.

7. Gear never reviewed as silhouette language (P2). 18 pieces pass fit/contract but have no prop-artist verdict (the unused beanie was retired 2026-11-03).

8. Venue shape language (P2). Hollywood and the Joke Shop have budget data but no environment-artist/archviz review; the TV Studio has no recorded map-budget number in this document yet.

Decisions needed from the owner

• Accept the measured texel density as a stylized limitation, or fund a skin/hair resolution pass.
• Approve or reject a single-variant face/eye-loop prototype (and pick which variant to prototype on) before any eight-character rebuild.
• Decide whether the pose/weight capture is required before the crowd-variant roster.
• Confirm foliage rebuild priority versus the map track (both touch the same street trees).

Next actions

Tracked in TODO.md 3D model re-evaluation: the remaining A. Baseline… item is this owner review; B/C/D carry the per-family work, and E adds the pose/weight, collider-coverage and scorecard-coverage gates this document found missing.