feat: asset-quality gate + remodel modular-kit-snap and lightmap-uv-channel assets

[TMHSDigital]

What this is

Codifies "modeled with intent" as a measured gate, then applies it to the two examples that exposed the gap (modular-kit-snap, lightmap-uv-channel — both passed every existing gate and still shipped placeholder-reading assets). Four commits: the gate, two remodels, integration.

The floors (calibrated empirically, measurements reported)

Implemented in examples/gallery_asset_quality.py (render path only, same call pattern as gallery_framing, exit 11, never in check-only paths):

• Naming — no default datablock names on hero parts.
• Material variation — heroes with ≥ 2 parts carry ≥ 2 distinct materials. (Single-part subjects exempt: one honest material is their truth.)
• Edge treatment — right-angle (90° ± 2°) manifold edges ≤ 75%. Calibration: reference assets 0.019–0.288, bmesh-gear's honest machined teeth 0.667, raw boxes 1.0.

Calibration table (measured on Blender 5.1.2 via the survey harness):

asset	parts	mats	edge90	compactness (info)
collision-hull-proxy (ref)	16	6	0.044	65.9
custom-normals-shade (ref)	39	2	0.288	31.0
vertex-weight-limit (ref)	1	4	0.150	34.8
lod-decimate-chain (ref)	3	4	0.019	79.1
depsgraph-export (weak)	2	1	1.000	23.9
text-version-stamp (weak)	1	1	1.000	405.3
bmesh-gear (simple, honest)	1	1	0.667	18.3
soccer-ball-goldberg (simple, honest)	1	2	0.000	10.0

Dropped floors, with evidence (a floor that can't be set without failing a genuinely good asset is not a floor; a floor hand-tuned to pass everything isn't one either):

• Part count — vertex-weight-limit's mech arm is a single skinned mesh and reference-quality.
• Silhouette compactness — soccer-ball-goldberg 10.0, bmesh-gear 18.3, turntable 19.1 are good simple subjects. Kept as printed information only.
• Dominant-material share — the jerry can and the rocket carry accents in per-face assignment (both measure 1.0 dominant). Printed information only.

Floor falsification (blender --background): raw default-named cube → exit 11 (naming + edge90); two-part single-material assembly → exit 11 (materials + edge90); designed restore → exit 0. Watched fail and pass.

Goodhart warning written into docs/VISUAL-STYLE.md: greebles for the metric's sake do not count; detail must serve plausibility.

Reference bar + asset sheet gate

Reference set pinned in CLAUDE.md (canonical): collision-hull-proxy, custom-normals-shade, vertex-weight-limit, lod-decimate-chain. Asset sheet: hero alone, neutral 3/4 studio, composited beside the reference set, committed:

• docs/gallery/asset-sheets/lightmap-uv-channel.webp
• docs/gallery/asset-sheets/modular-kit-snap.webp

Both sheets are [before, after, chp, cns, vwl, lod] with the befores re-rendered from the pre-remodel code in the same neutral rig. Verdicts: both pass — neither remodeled asset is identifiable as the least-designed in the lineup; both befores clearly were.

lightmap-uv-channel remodel

Found and fixed: the wheel DarkWood material was silently never applied (prefix lookup bug — exact per-part mapping now). Rebuilt as 21 named watertight parts: plank-grooved bed + side rails + corner brackets + fasteners, real wheel assemblies (disc + hub + iron band + bolt ring) on a capped axle, ribbed canvas canopy, enlarged atlas board. Closed forms moved deliberately (contracts identical): parts 9 → 21, islands 1114 → 3680, byte-identical on 5.1.2 and 4.5.11. Re-falsified: clobber drift 2.068, overlap 15 SAT hits, margin collapse 2e-5. The MeshUVLoopLayer re-fetch ordering that avoids the documented 4.5.11 segfault is preserved and verified on 4.5.11 (no EXCEPTION_ACCESS_VIOLATION; hazard note still accurate).

modular-kit-snap remodel

Panels → assemblies (backing plate + inset + 4 bolt heads); palette teal-slate (no longer twins the cart); luminance 82.7 → 66.1 (calibration ceiling 77.7) — no longer a fourth over-bright card, so only the documented framing deviation (interior run) remains. Detail parts 12 → 20, hygiene 13 → 21; boundary contract untouched (16 verts/edges, probes exit 3/4 with measured 3 mm/2 mm errors).

Survey worklist (in ROADMAP)

Full 42-example survey: 26 pass, 14 below floors, 2 non-mesh subjects (n/a). Every below-floor entry is a contract-vehicle/display subject (testcard TV, VSE monitors, text bars, display rigs, honest single primitives) with measured numbers — none a game asset, none a remodeling mandate. Worklist committed in ROADMAP.

Visual passes (iteration log)

• luv: 4 iterations (frame margin violation → reframe; giant-triangle geometry bug from bm.verts[n0:] slicing — bolts fanned outward, found by extents probe, fixed with set-difference; bed/iron material tuning) + falsify render + 4.5.11 parity.
• mks: 1 iteration to target (panels/palette/luminance in one pass against measured luminance) + falsify render + 4.5.11 parity.
• Staging contact sheets re-committed for both heroes (mks 66.1, luv 52.8 vs calibration 44.5–77.7).
• In-situ Playwright: both cards read distinctly in compact mode at 1280 and 390; detail pages re-read (3680 islands / twenty detail parts / 66.1 present; alts untruncated).

Test plan (exact binaries)

• E:\Blender-Developer-Tools\.scratch\blender-5.1.2-windows-x64\blender.exe (Blender 5.1.2): all checks, probes, renders, surveys, floor falsification, asset sheets.
• E:\Blender-Developer-Tools\.scratch\blender-4.5.11-windows-x64\blender.exe (Blender 4.5.11 LTS): both check-only paths (byte-identical outputs), both render parity passes, hazard-ordering verification.
• CI: both smoke jobs run the two example steps on their floating series.

Live run vs inspection-only

Live run: everything above. Inspection only: the 4.5 series smoke on CI runs 4.5.12 (local binary is 4.5.11 — stated exactly), Cycles engine paths (renders used EEVEE), Firefox/Safari rendering (Chromium only).