fix(examples): redesign car-mirror-symmetry as a smooth-shaded stylized hatchback

[TMHSDigital]

What and why

A gallery review judged the car-mirror-symmetry hero weak. It named four problems: a stray pink streak on the hood near the windshield, headlamps that looked like flat white stickers, a heavily faceted body that read as unfinished, and dull windows.

The streak came first. It is not a mirror-seam defect. I cast camera rays through the streak pixels on the committed code. Every hit landed on one face: authored face #37 (station 4, ring segment 5), seen on the mirrored half.

• It sits at |x| 0.11–0.38, well away from the symmetry plane.
• It is the thin panel where the old 9-point ring collapsed its greenhouse points onto the hood.
• It was flat-shaded and faced the key light almost head on (n · −key = 0.967), so the clear-coated red blew toward pink.
• All weld checks passed on that code (2n−c, on-plane count, manifold, Euler 2).

So there was no seam bug and no weld check to add. The redesigned ring cannot produce that thin panel.

Changes

• Body loft, rebuilt. 52 placed stations × 13-point rings. Each ring blends from a hood layout into a cabin layout, which removes the thin panels. Other additions: round wheel arches around the wheels, flared black arch cladding and side skirts, a character line, black A- and B-pillars and drip rail, and a hatch.
• Smooth shading. Sharp edges are set on the authored half by dihedral angle (35°) and material border. Edges on the plane take the angle to their own mirror image. There is no bevel or weighted-normal modifier, so the stack is still Mirror only and every count stays closed-form.
• Parts. Headlamps, taillamps, grille, door mirrors and door handles are closed superellipse pods. Each has a chrome bezel and an inset lens, and a BVHTree ray cast seats it on the body surface. The wheels have five recessed spokes.
• Glass. A dark, glossy dielectric whose colour ramps toward steel blue at grazing angles (Layer Weight → Color Ramp). On a near-black stage there is little for glass to reflect, so plain dark glass read as dull paint.
• Grille welded across the plane (check extended). The grille is authored as a half whose end points lie exactly on x = 0, and the merge welds it into one part. The part check now takes a weld count per part: authored on-plane verts must equal weld, evaluated verts must equal 2n − weld, and evaluated on-plane verts must equal weld. Offset parts use weld = 0, so their check is unchanged (2n).
• Render gates added. The render path now runs gallery_framing and gallery_asset_quality; it previously ran neither. Their exit codes are remapped to 17 / 18, because 10 and 11 are already check codes in this script.
• Camera. It sits higher on the key side and looks down the hood centerline, so both headlamps, both mirrors, both A-pillars and the single centered grille read in the same view.

Before / after

• Contact sheet: docs/gallery/contact-sheets/car-mirror-symmetry-contact-sheet.webp. The before is the same path on main.
• Asset sheet: docs/gallery/asset-sheets/car-mirror-symmetry.webp. Built with the committed python scripts/asset_sheet.py --blender <5.2.1> car-mirror-symmetry against the reference set chore(gates): raise the asset-sheet reference bar and commit its renderer #257 pinned in CLAUDE.md. Top row: new car, collision-hull-proxy, socket-attach-points. Bottom row: vertex-color-ao, wheelbarrow, apothecary-shelf. This PR adds the car's SELECT row (body, wheels and fitted parts; stage excluded).

Gate numbers (5.2.1 LTS, EEVEE)

• Framing: fill x=0.734, y=0.789 (band 0.70–0.90). Margins: left 0.138, right 0.128, bottom 0.072, top 0.139. framing_ok.
• Asset quality: no default names; 8 materials; edge90 0.035 (112/3174); compactness 16.5 (informational only). aq_ok.

Contact sheet, per criterion:

Criterion	Candidate	Calibration	Verdict
Mean luminance	0.1730	armature-bend 0.2787, damped-track-aim 0.1511, bmesh-gear 0.3024	pass (within band)
Stage darkness	0.1652	0.145–0.236	pass
Wedge warmth	+0.395	bmesh-gear +0.374	pass, warm pool behind the car
Saturation	0.262	highest of the four	appropriate for a red hero
Subject fill	see framing above		pass
Thumbnail legibility (320×180)			reads instantly as a red car with lamps, grille and wheels

• Asset sheet verdict (new reference set): pass. Alone under neutral light, the car shows panel breaks, arch cladding, black pillars, bezelled lamps, one welded grille, mirrors, handles and spoked wheels. It is not the least-designed object in the lineup. It clearly clears the hydrant, and it is comparable to the wheelbarrow. The drone, the stone well and the apothecary shelf carry more fine-grained detail and material variety; the car's detail is broader and lower-frequency, by stylized game-asset design.
• Hero drift: measure_hero_drift.py --only car-mirror-symmetry reports matches (mean_abs 0.00306, luma 0.1730/0.1728).

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

Binary	Reported version	Result
E:\Blender-Developer-Tools\.scratch\blender-5.2.1-windows-x64\blender.exe	Blender 5.2.1 LTS	PASS
E:\Blender-Developer-Tools\.scratch\blender-5.1.2-windows-x64\blender.exe	Blender 5.1.2	PASS
E:\Blender-Developer-Tools\.scratch\blender-4.5.11-windows-x64\blender.exe	Blender 4.5.11 LTS	PASS

The check lines (body half=… / mirrored parts | …) are byte-identical across all three: body half 676/1289/614, 104 centerline verts; evaluated 1248/2474/1228, Euler 2, manifold, sym_dev 0.000e+00.

Falsifiers (5.2.1, each probe breaks one thing then runs check())

Probe	Exit	Measured
--no-mirror	5	676 verts ≠ 1248
body use_mirror_merge=False	5	1352 verts ≠ 1248
Mirror applied into data	3	datablock (1248, 2474, 1228) ≠ (676, 1289, 614)
one centerline vert moved to x=0.01	4	103 ≠ 104
Headlamp origin at x=0.01	12	origin x=0.01
new: grille merge off	14	(72, 52, 16) ≠ (64, 52, 8)
new: grille mirror axis off	14	(36, 26, 8) ≠ (64, 52, 8)
new: one grille end point at x=0.004 (beyond merge threshold)	13	on-plane 7 ≠ 8
front wheel mirror axis off	14	(390, 361) ≠ (780, 722)
new: framing band forced to 0.95–0.99	17	fill 0.789 FAIL
new: a hero part renamed Cube	18	default datablock names ['Cube']

Proven by live run

• The streak diagnosis: ray-cast face mapping, x-range and normal·key on the committed code.
• Check PASS on 5.2.1, 5.1.2 and 4.5.11, with byte-identical output.
• Every falsifier above.
• Framing and asset-quality numbers.
• Hero and preview regenerated via render_hero.py, and the hero-drift match.
• Contact sheet via contact_sheet.py, with its luma, stage, warmth and saturation.
• Asset sheet rendered on 5.2.1 with the committed scripts/asset_sheet.py.
• An unwelded body seam renders within 12/255 of the welded one (62 px > 10/255). This is recorded in the README as "not a pixel witness"; the count is what catches it.
• build_gallery.py, with the new alt read in the generated docs/gallery/index.html and docs/gallery/car-mirror-symmetry/index.html.

Established by inspection only

• The render path (--output) was run on 5.2.1 only, not on 5.1.2 or 4.5.11. It uses Coat Weight, Coat Roughness, ShaderNodeLayerWeight and ShaderNodeValToRGB, all present in 4.x Principled/node APIs. CI smoke does not pass --output.

🤖 Generated with Claude Code