feat: add modular-kit-snap example

[TMHSDigital]

What this is

New example modular-kit-snap: a tiling corridor kit where the snap is the contract — a 4 × 3 × 3 m corridor segment whose open-end boundary vertices sit exactly on the declared tile grid, so instances at 4 m multiples join with zero gap and zero overlap. Asset-first entry in the game-asset arc: built for reuse (pivot at the start-edge connection point, identity transforms, Kit.CorridorSeg.* datablocks, manifold except the two intentional open ends, 12 beveled detail parts contained strictly inside the tile).

Also on this branch, per the task's opening commit: ROADMAP candidate-pool reconciliation (docs: commit) — removed the stale "Gallery coverage follow-ups from the GPv3 review" line (everything it names shipped; its light-linking subject duplicates the light-link-studio line) and struck through the gltf-skin-roundtrip entry, which had shipped but was left unstruck. Every remaining unstruck pool entry verified genuinely unbuilt.

Contract witnessed (check-only path, all closed form or re-derived)

Check	Measured (5.1.2 / 4.5.11)	Exit
Snap: 16 boundary verts (8/end) on x ∈ {0, TILE}	max_dev 0.000e+00, tol 1e-6 m	3
Loop coincidence (nearest (y,z)-key match)	8/8 ring, max_dev 0.000e+00	4
Tiling: linked duplicate at (4,0,0) joint positions	yz_dev 0.000e+00, x_dev 0.000e+00	5
BBox == declared tile (shell and whole asset)	dev 0.000e+00, tol 2e-6	6
Manifold except open ends	16 boundary edges, 0 nonmanifold, 0 off-plane	7
Detail watertight + contained in tile	12 parts, all inside	8–9
Hygiene: identity scale, kit names, pivot origins	13 parts	11

Check output byte-identical on 4.5.11 LTS and 5.1.2.

Falsification evidence (live run)

Probe (--background, same check() against the unsnapped variant built by the same code minus the snap pass):

• 3 mm end-ring skew → exit 3, worst 3.000e-03 m off the end planes; cascades to bbox dev 3.000e-03 and 8 off-plane boundary edges.
• 2 mm y-nudge, no skew → exit 4, ring displacement 2.000e-03 measured.
• Authoring hazard found and fixed: sorted-zip ring matching mispairs displaced verts and reported a bogus 3.002 cascade for a 2 mm defect — the matcher now pairs by nearest key (in code comments and ROADMAP).

Render and visual passes (iteration log)

• Pass 1 (5 iterations): draft1 too dark/no focal/dead right wall → fixtures relit; draft2 lit bulkhead door added but clipped white; draft3 door tamed (amber, no clip) + cool mouth rim; draft4 reframed symmetric into the mouth; draft5 panel/rim polish = final.
• Pass 2 (2 iterations): falsify1 at 40 mm/15 mm was invisible behind joint ribs → falsify2 at 120 mm gap + 50 mm jogs + 40 mm floor steps: floor plates split with dark seams, trim rails visibly break. Snapped-vs-broken comparison captured.
• Pass 3: 320px thumbnail reads as corridor-with-destination in under a second.
• Pass 4 (in situ, Playwright): card reads instantly in compact mode (the gallery default) at 1280 and 390; detail page verified; 41-card index intact. Screenshots reviewed.
• Pass 5 (contact sheet): docs/gallery/contact-sheets/modular-kit-snap.webp — holds the calibration lineup (armature-bend, damped-track-aim, bmesh-gear): dark base, warm focal accents, designed materials. Mean luminance 82.7 vs calibration 44.5–77.7 — the lit-interior deviation class, not the flat-bright outlier class.
• Framing: documented deviation (interior run; the envelope surrounds the camera, joints are the proof). check_framing reports fill 1.000/1.000 with all-edge bleed, gate in report mode, reason at call site and in the README's Framing deviation section.

featured_rank: yes — takes rank 8 from color-attribute-wheel (the flattest of the previous eight); the corridor is the only interior/envelope composition in the landing set. Landing build verified locally after the swap.

Generated output read

docs/gallery/index.html card alt re-read character by character — complete sentence, no truncation, no dotted-path issues (teaches shortened after the first regen capped it mid-phrase). WITNESSES copy verified against measured check output: 16 verts, 1e-6 m, deviation 0.0, probe 3e-3 m — all exact.

Test plan (exact binaries)

• E:\Blender-Developer-Tools\.scratch\blender-5.1.2-windows-x64\blender.exe — reports Blender 5.1.2: check exit 0, render, falsification renders, both probes.
• E:\Blender-Developer-Tools\.scratch\blender-4.5.11-windows-x64\blender.exe — reports Blender 4.5.11 LTS: check exit 0 (byte-identical output), render parity confirmed visually.
• CI: both blender-smoke jobs run the new step (check-only) on their floating series releases.

Live run vs inspection-only

Live run: all checks, probes, renders, framing numbers, in-situ captures, contact sheet. Inspection only: glTF/engine import behavior (not claimed — README explicitly does not claim engine compatibility), Cycles engine path (--engine cycles flag exists; renders here used EEVEE).