fix: final gallery quality run — hero-material redesigns for bmesh-gear, gn-sdf-remesh, depsgraph-export

[TMHSDigital]

Final gallery quality run. Scope: exactly three examples — bmesh-gear, gn-sdf-remesh, depsgraph-export — the three that survived two passes without reaching the house standard. No other example is touched.

The acceptance gate this run is the contact sheet: each candidate placed at equal size between armature-bend and damped-track-aim, passing only if the three cannot be sorted into different eras or studios. The final sheets are committed with this PR:

• bmesh-gear contact sheet
• gn-sdf-remesh contact sheet
• depsgraph-export contact sheet

Why previous passes under-delivered

Two rounds of "re-stage to the style doc" adjusted lighting around unchanged subjects. The defect was the subject treatment: a bare bright-steel gear face renders as one featureless gradient under area lights (reads as white plastic), the torus was lit frontally so the SDF facets — the API evidence — never glinted, and the grey-cube/green-sphere pair was never designed at all. This run changed the subjects.

Per-example before/after

• bmesh-gear — before: flat pale disc. after: machined brass — a spherical RINGS wave drives roughness (0.14–0.36) so the flat face breaks into concentric turning marks that each catch the key at a different angle; steeper tilt, small hot glint light, dark stage with the warm wedge. Two iterations: the first used the Wave default X-axis rings direction, which rendered as zebra banding; fixed to rings_direction='SPHERICAL'. Material name changed Steel → MachinedBrass (render-path only; the check is topology-only). Alt text updated.
• gn-sdf-remesh — before: frontal, softly lit, facets invisible. after: glossier crimson (roughness 0.16), low raking key from the left, off-axis low camera — every remesh facet now reads as a distinct glint, which is exactly the evidence the example exists to show. Passed the gate on the first candidate.
• depsgraph-export — before: default-grey cube and green sphere, the least designed card in the gallery. after: one saturated teal material on both objects, so the only visible difference between the two silhouettes is the thing the check measures — base-mesh geometry versus the depsgraph-evaluated subsurf. Cube turned 32° off camera-parallel and camera raised so it presents three faces. Three iterations (18° turn was imperceptible). Alt text updated.

Rules compliance

• Check-only path exit 0 on Blender 4.5.11 LTS and 5.1.2 for all three after modification (live runs, logged per version below in CI as well). No tolerance touched; no version-gated path disturbed (none of the three carries a shim beyond the EEVEE engine-id helper, which is exercised by both smoke jobs).
• gn-sdf-remesh's check asserts the Ceramic material rides through the GN Set Material node — the material keeps its name and slot; only color/roughness values changed, and the check passes on both versions.
• Gallery regenerated with scripts/build_gallery.py; no hand-edits to generated pages or release-owned fields; per-example commits.

Whole-gallery confirmation

The regenerated gallery was served locally (python -m http.server from docs/), loaded in a real browser with lazy-loading defeated, and scrolled end to end: no remaining card reads as old-era. No out-of-scope defect was found that needs flagging.

Proven by live run: all check paths (both Blender versions), all renders (5.1.2 EEVEE), every candidate inspected at full size, card crop, and thumbnail, plus the contact-sheet gate per candidate. Inspection only: Cycles render paths and 4.5 EEVEE-Next render paths (engine-agnostic scene code behind the existing gate; CI smoke is check-only by design).

🤖 Generated with Claude Code