fix(examples): redesign degenerate-bevel-weld as rugged cases with a traced collapse seam

[TMHSDigital]

What changed and why

The gallery review judged the degenerate-bevel-weld hero weak. It showed two near-identical teal slabs, shot front-on, which hid the one face the collapse changes. The tiny orange dots and floor labels could not be read at card size, so at thumbnail the two slabs looked like the same object.

This is a redesign:

• Designed prop whose bevel matters. Each shell is one of the check's meshes, beveled_box(DIMS, 0.10) and beveled_box(DIMS, 0.20), built out as a rugged hard-shell equipment case. Fittings: a D-handle on pivots, twin draw latches with anodized levers, a lid seam, hasps, a molded reinforcement frame, ribs, a purge valve, an ID plate and rubber feet. That is 22 named parts in 6 materials. Every fitting sits inside the front land that survives both offsets, so the bevel is the only difference between the two cases.
• The collapse reads as silhouette. Both cases turn their end panels to the camera. On the left, a flat end panel is framed by clean chamfer bands. On the right, the end and top lands are gone and the band rolls into a knife ridge at mid-depth.
• Hot highlight from live data. A red emissive seam traces every edge of every face under 1e-6 area on the degenerate shell: 16 thin faces, 32 segments. An orange bead sits on each of the 12 zero-area faces the check counts. Change the offset and the overlay moves or disappears.
• Labels removed.
• Stage turns, not the props. Every root object is rotated by -CASE_YAW after staging. The image does not change (mean abs diff 0.019/255 against the un-turned build), and the cases stay axis-aligned in world space, which is what the asset-sheet panel's three-quarter camera expects. A first attempt with rotated cases produced a cropped end-on panel.
• The render path now also runs gallery_asset_quality.check_asset_quality (exit 11) on the clean case. scripts/asset_sheet.py gains the degenerate-bevel-weld SELECT row (^CaseSafe\.).
• Updated: README render-as-proof section and exit-code table, the examples/gallery.json teaches/alt, and the root README gallery-row alt. Gallery regenerated.

The check path is unchanged: check(), the constants and the exit codes 3–8 are untouched.

Gates (Blender 5.2.1 LTS, E:\Blender-Developer-Tools\.scratch\blender-5.2.1-windows-x64\blender.exe)

• Framing: fill x=0.778, y=0.689 (band 0.70–0.90). Margins: left 0.081, right 0.141, bottom 0.111, top 0.200. framing_ok.
• Asset quality: naming ok, materials n=6 (dominant [email protected]), edge90 0.008, compactness 17.5 (informational). aq_ok.
• Hero drift (measure_hero_drift.py --only degenerate-bevel-weld): mean_abs 0.00345, gt2pct 0.95 %, verdict matches.
• Hero webp is 29.8 KB and preview webp 27.2 KB, both written by render_hero.py.

Contact sheet: docs/gallery/contact-sheets/degenerate-bevel-weld-contact-sheet.webp

tile	luma	stage	warmth	sat
degenerate-bevel-weld	0.1937	0.2066	−0.2607	0.2487
armature-bend	0.2787	0.2132	−0.4279	0.1436
damped-track-aim	0.1511	0.1448	+0.2772	0.2305
bmesh-gear	0.2358	0.1772	+0.3027	0.2495

Verdicts:

• Stage darkness: pass. Stage is 0.207, within the calibration range 0.145–0.213.
• Wedge warmth: pass, on the cool side. The warm wedge pool is on the back wall. The metric is negative because the brightest tenth of pixels is dominated by the blue polymer highlights, as with armature-bend's cool subject (−0.43).
• Subject fill: pass. Measured fill is 0.778 in x.
• Saturation: pass. 0.249 matches bmesh-gear (0.250).
• Thumbnail legibility: pass. At 427×240 the left case reads as a box with a flat end, and the right reads as a rounded case wrapped by a red seam.
• Mean luminance: 0.194, against a calibration mean of 0.222 (range 0.151–0.279).

Asset sheet: docs/gallery/asset-sheets/degenerate-bevel-weld.webp

Lineup: the candidate (the clean case, CaseSafe.*), then collision-hull-proxy, socket-attach-points, vertex-color-ao, wheelbarrow and apothecary-shelf, read from CLAUDE.md on main.

Verdict: ships, but it is among the simpler objects in the lineup. It reads as a manufactured object with assembled hardware, treated edges and a real material split. It is on a par with the hydrant for form complexity and below the drone, wheelbarrow and shelf. It is not clearly identifiable as the least-designed object in the lineup.

Check runs (python tests/smoke/run_example.py, check-only path)

binary	reported version	result
.scratch\blender-5.2.1-windows-x64\blender.exe	5.2.1 LTS	PASS: collapse 12f/16v ships 32 degenerate tris in the GLB
.scratch\blender-5.1.2-windows-x64\blender.exe	5.1.2	PASS (same counts)
.scratch\blender-4.5.11-windows-x64\blender.exe	4.5.11 LTS	PASS (same counts)

The render path (--output) also ran clean on 4.5.11 LTS: render_defects thin_faces=16 zero_area_faces=12 seam_segments=32, framing_ok, aq_ok.

Falsifiers

• Existing check (unchanged): --both-safe on 5.2.1 exits 4 with ERROR: degenerate bevel produced 0 zero-area faces, closed form 12.
• Overlay: no new gated check was added. The seam and the beads are counted from the same mesh data. At the safe offset the thin-face set is empty, so the overlay disappears; --both-safe stops at exit 4 before any render.

Proven by live run

• All three check-only runs listed above, and the --both-safe exit 4.
• On 5.2.1: the framing and asset-quality gates, the render_hero.py regeneration, the contact sheet, the asset sheet and the drift check.
• The render path on 4.5.11.
• build_gallery.py after the rebase onto main. The generated <img alt> in docs/gallery/index.html and docs/gallery/degenerate-bevel-weld/index.html was read back and is the new one-sentence alt (181 chars).

Established by inspection only

• That the turned stage is equivalent to rotated props: measured as a pixel diff on one pair of renders, not proven in general.
• The asset-sheet verdict and thumbnail legibility are visual judgements.

🤖 Generated with Claude Code