fix(examples): rebuild triangulate-tangents buckler as a machined shield

[TMHSDigital]

What and why

A gallery review judged the triangulate-tangents hero weak:

• Wobbly face. The face was covered in wavy ring ripples that read as a shading bug. That is the worst look for an example about tangent-space correctness. The root cause was the old material: a ShaderNodeTexWave on Generated coordinates with Distortion 3.0. The ripples were never UV-driven.
• Wrong read. The object looked like a dish or speaker cone with a murky grey interior.
• Hole at the apex. The profile's innermost ring was left open and rendered as a black hole.

Changes

• Geometry (checked mesh). The buckler is one closed lathe mesh, 24 rings × 96 sides:
  • a pointed turned boss, a flat flange, a cosine dome, a cut groove and a rolled rim band;
  • a new apex fan closes the top, and the closed-form triangle count includes it (2*((RINGS-1)*SIDES + 2*SIDES) + 2*SIDES = 4992);
  • the back plate is thinner;
  • the mesh is smooth-shaded, with sharp edges where faces turn more than 30°. As a result mikktspace now really welds frames across shared vertices.
• Materials (render only). Assigned per lathe ring: turned brass boss and rim, blued-steel flange, turned-steel dome, leather back, and a separate Buckler.Rivets part with 16 rim rivets and 8 flange rivets.
• Grooves. Lathe grooves are |sin(pi*40*v)| bumps read straight off UV v, which is the radius. They are exact circles with no distortion, masked by a constant ramp into two decorative bands.
• Highlight. The anisotropic highlight follows a UV-map ShaderNodeTangent. The old code looked up a nonexistent "Anisotropic IOR Level" input first; it now uses Anisotropic directly.
• Staging. The shield is turned three-quarter (yaw +48°), so the boss projects and the dome reads convex. The stand is walnut with beveled edges. The kickstand foot now angles away from the camera, so it no longer peeks out under the rim. A warmer key replaces the old Glint light, a Top softbox is added, and the Fill is trimmed.
• Asset-quality gate. The render path now runs gallery_asset_quality.check_asset_quality (exit 11) after check_framing.
• Docs.
  • The docs/VISUAL-STYLE.md numeric-only precedent sentence now states the actual mechanism: grooves are a bump off UV v, and the streak follows the renderer's own UV-map tangent. So a break in the calc_tangents() loop data the check reads does not move pixels.
  • The README numbers, render description and exit-code table (11) are updated.
  • The gallery.json alt and the README gallery-row alt are updated.

Evidence

• Contact sheet: docs/gallery/contact-sheets/triangulate-tangents-contact-sheet.webp
• Asset sheet: docs/gallery/asset-sheets/triangulate-tangents.webp. It was regenerated with the committed scripts/asset_sheet.py (from chore(gates): raise the asset-sheet reference bar and commit its renderer #257) against the new reference set on 5.2.1.
  • Candidate first, then collision-hull-proxy, socket-attach-points, vertex-color-ao, wheelbarrow and apothecary-shelf.
  • The new SELECT row is ^Buckler(\.Rivets)?$: the lathe mesh plus the rivets. The walnut stand and strut are display staging, not part of the asset, so they are excluded.
  • Panel floors: parts=2, materials=4, edge90=0.012.
• Before: the committed hero on main (docs/gallery/assets/triangulate-tangents-hero.webp at 15ba0a3).

Gate numbers (5.2.1)

• Framing: fill x=0.381, y=0.828 (band 0.70–0.90) ok. Margins: left 0.300, right 0.319, bottom 0.028, top 0.039 (min 0.02). No touch or cross.
• Asset quality: no default names; materials n=4; edge90 0.012 (96/8224), no degenerate faces; compactness 10.2 (informational only). aq_ok.
  • The line prints [email protected] ... FAIL, but that metric is informational only and is not gated (see VISUAL-STYLE § Dropped floors).
• Clipping: 0.09 % of pixels at full white. The old hero had a blown dome.
• Drift: measure_hero_drift.py --only triangulate-tangents gives mean_abs 0.00307, gt2pct 0.33 %, verdict matches.

Contact-sheet metrics

tile	luma	stage	warmth	sat
candidate (new)	0.2147	0.1891	+0.0186	0.1018
old hero	0.2454	0.1996	−0.0074	0.0553
armature-bend	0.2787	0.2132	−0.4279	0.1436
damped-track-aim	0.1511	0.1448	+0.2772	0.2305
bmesh-gear	0.3024	0.2357	+0.3743	0.1953

Per-criterion verdicts

• Stage darkness: pass. Stage 0.189 sits inside the calibration range 0.145–0.236.
• Wedge warmth: marginal. It is +0.019, up from −0.007. The brightest tenth of the lower half is the steel dome, which is neutral by material. The warm wedge is visible on the back wall at upper right, and the key is warm.
• Subject fill: pass. 0.828 y.
• Saturation: improved but below calibration. 0.102, up from 0.055. A steel subject is inherently low-saturation; the brass boss and rim and the walnut stand carry the colour.
• Mean luminance: pass. 0.215 falls within the calibration spread of 0.151–0.302.
• Thumbnail legibility: pass. At 320×180 it reads as a round shield with a pointed boss, riveted rim and concentric grooves. There is no wobble or moiré.

Asset-sheet verdict (new reference set)

Pass, but close. The buckler is on par with the hydrant (collision-hull-proxy). Both are lathe-turned bodies with distinct material zones and designed profile breaks. The buckler adds a pointed boss, a groove-banded dome and 24 rivets, so it is not identifiable as the least-designed object.

It is the simplest silhouette in the lineup, though. The drone, well, wheelbarrow and shelf are assemblies of many parts with more structure. A round shield is inherently a single turned plate, so it will not match their part count.

Checks per binary (check-only path)

binary (reported version)	exit	tris	unit / dot / btn / uv err	weld_err	seam / chart flips
E:\Blender-Developer-Tools\.scratch\blender-5.2.1-windows-x64\blender.exe (Blender 5.2.1 LTS)	0	4992	1.34e-07 / 9.89e-08 / 0 / 0	3.27e-02	150 / 12
...\blender-5.1.2-windows-x64\blender.exe (Blender 5.1.2)	0	4992	identical	3.27e-02	150 / 12
...\blender-4.5.11-windows-x64\blender.exe (Blender 4.5.11 LTS)	0	4992	identical	3.27e-02	150 / 12

Falsifiers (re-proven on the new geometry, all three binaries unless noted)

probe	how	5.2.1	5.1.2	4.5.11	measured
--zero-uv	README flag	exit 4	exit 4	exit 4	UV deviation 1.405
dropped profile ring	delete one rim ring from PROFILE	exit 3	exit 3	exit 3	4800 vs 4992 triangles
swapped du/dv	break the independent tangent formula	exit 7	exit 7	exit 7	weld 1.460
no re-fetch after calc_tangents	stale MeshUVLoopLayer handle	exit 0, correct	exit 0, correct	exit 0, corrupt	4.5 shows 2829 phantom seam flips over 2497 phantom seam positions
asset-quality gate (new call site)	rename Buckler.Rivets to Sphere (5.2.1, --output)	exit 11, no still written	—	—	—

The stale-handle row matches the documented hazard: the handle is silent on 4.5 and survives by luck on 5.x.

Proven by live run vs established by inspection only

Proven by live run (headless binaries):

• all check-path numbers above;
• all falsifier exits;
• framing and asset-quality gate output;
• the hero and preview regenerated via scripts/render_hero.py;
• contact-sheet metrics;
• asset-sheet panels, via scripts/asset_sheet.py against the chore(gates): raise the asset-sheet reference bar and commit its renderer #257 reference set;
• the drift measurement;
• build_gallery.py output, with the generated <img alt> read back in docs/gallery/index.html and docs/gallery/triangulate-tangents/index.html.

Established by inspection only:

• the thumbnail-legibility and asset-sheet verdicts are eye judgments of the committed images;
• the VISUAL-STYLE claim that the render does not depend on calc_tangents() loop data. It rests on the material using only the UV-map Tangent node and UV coordinates; no render was made with corrupted loop tangents.

🤖 Generated with Claude Code