feat: attribute-domain-shear example + count/numeric-contract housekeeping

[TMHSDigital]

Attribute domain shear + two housekeeping commits

New example witnessing POINT vs CORNER semantics on Mesh.color_attributes — plus the two housekeeping commits that opened the branch (count-prose fix, numeric-contract exception codified).

Housekeeping

1. docs: count-prose fix. AGENTS.md and CLAUDE.md said "23 examples" while README, the manifest, and CI all agree on 38 (derived from .cursor-plugin/plugin.json; every other count in both files verified correct as written). Example-shipping commits keep the lines current — this PR moves them to 39.
2. docs: numeric-contract exception in VISUAL-STYLE Layer 1. Codifies what vertex-weight-limit (pose-preserving prune) and triangulate-tangents (mikktspace field) already declared honestly: a contract with no visual signature is declared numeric in the README, the render depicts subject and setup legibly, and falsification names why. Anchored to those two precedents so the exception stays narrow.

The example

examples/attribute-domain-shear/ — an 8-wedge pinwheel around one raised hub vertex shared by every wedge (plus a shared outer ring). The contract, all closed form:

• Storage sizes: CORNER attr data == len(loops) == 3K; POINT attr == len(vertices) == K+1.
• CORNER authoring exact: hub corner of wedge i reads palette[i], err ≤ 2.6e-08 — K faces may disagree at one vertex.
• Naive POINT authoring shears by construction: a per-wedge "paint each wedge its color" loop rewrites every shared vertex once per neighbor; last write wins — hub reads palette[K−1], ring vert i reads palette[i] except the wrap-around vert 0 (also palette[K−1]). Measured shear 0.751031 == palette closed form 0.751031.

Companion to color-attribute-wheel (domain sizing, active_color, shader Attribute node) — this covers the meaning of the domains on shared topology, not a repackage.

Falsification evidence

probe	expected	observed
wedge 3 miscolored in CORNER pass	exit 4	exit 4 (hub corners fail exact color)
naive writes reversed → hub reads palette[0]	exit 5	exit 5 (last-write-wins broken)
ring vert 2 corrupted	exit 6	exit 6 (overwrite-ordering witness)
constant palette → shear collapses	exit 7	exit 7 ("shear is ~0 — cannot distinguish naive from correct")

All restored; final check exits 0 with attribute-domain-shear OK corner_exact point_shear=0.751031.

Framing (measured, gallery_framing silhouette)

fill 0.837x / 0.478y, margins L0.081 R0.081 B0.089 T0.133 — in band on 5.1.2 and identical on 4.5.11. Gate wired via check_framing (exit 10). Draft 1 was gate-rejected (fill 1.000, margins 0) — the gate caught it before any artifact shipped.

Visual-pass log (4 render iterations)

1. Full size + helper numbers: d1 gate-rejected (over-fill); d2 edge-on camera hid the pinwheel pattern; d3 good but 0.887x near the ceiling; d4 final (camera pull-back + thicker/lighter stems).
2. Contract visible: the dual panel carries working and broken in one frame — CORNER fan crisp, naive POINT fan smeared. The right fan is the falsification variant, generated from the same closed-form palette the check asserts.
3. Thumbnail (preview webp at 400 px): crisp-vs-smeared distinction and both placards read in under a second.
4. Playwright (full-page captures): index at 1280 and 390 plus the detail page — new card, teaches, witnesses callout, and detail page all render correctly; generated HTML read character by character. One real finding: the alt guard correctly rejected my first teaches draft (a bare decimal "0.751031" sat past the 160-char cap, tripping the dotted-path assert — exact numbers moved to README per the WITNESSES-one-claim rule).
5. Contact sheet: composite vs armature-bend 70.8 / damped-track-aim 44.5 / bmesh-gear 77.7 — candidate 73.6, inside the calibration range (not a fourth bright card). Stage darkness ✓, wedge warmth ✓, fill comparable ✓, saturation is the palette data by design ✓, thumbnail legibility ✓.

Runtime

Local binaries: .scratch/blender-5.1.2-windows-x64/blender.exe (reports Blender 5.1.2), .scratch/blender-4.5.11-windows-x64/blender.exe (reports Blender 4.5.11 LTS). Check-only: exit 0 on both, byte-identical output, sub-second runtime (new smoke step is check-only, so smoke runtimes are unaffected). CI floats within series per CLAUDE.md.

Live-run vs inspection-only

• Live run (5.1.2): check runs, 4 falsification probes, 4 render iterations + gate, webp conversion, contact sheet, gallery regen, Playwright captures.
• Live run (4.5.11): check-only + render path (identical numbers).
• Inspection only: none beyond the standard assumption that CI smoke runtimes track local timings.

Test plan

1. blender --background --python examples/attribute-domain-shear/attribute_domain_shear.py -- → exit 0, point_shear measured=0.751031 closed_form=0.751031 (4.5 LTS and 5.1).
2. -- --output out.png → framing lines print framing_ok, exit 0, artifact written.
3. Break any of the four probed paths → the documented exit code.