fix: re-stage the two newest heroes to the calibration-set bar

[TMHSDigital]

Summary

• Re-stages the vse-gamma-cross and light-link-studio heroes to the calibration-set bar. A card-scale review found both sorted as the odd one out in their own committed contact sheets — the failure mode the CLAUDE.md contact-sheet gate exists to prevent. The witness engineering reviewed at or above bar; the renders did not.
• Render-path and docs only, with two deliberate check-side exceptions: vse-gamma-cross gains the final-frame sample (frame 32, t=31/32) so its endpoint-convention witness is pixel-backed end to end, and strips_coll() now calls .strips directly instead of papering over the rename with a .sequences fallback.
• light-link-studio ships no dead code: IS_5X, eevee_engine_id(), the one-choice --engine flag, and the ignored engine parameter are deleted, and the previously unreproducible EEVEE figures are replaced with fresh probe measurements.
• Measured values shifted with the re-lit scene and are updated consistently across READMEs, gallery.json, and the regenerated gallery pages.

Per example

vse-gamma-cross — the gamma-dip proof was a 0.115 red-channel delta between two non-adjacent ~95×115 px swatches: invisible at card scale, so a naive-lerp Blender would render a nearly identical image. The bench is recomposed around four large fade panels (t=0, 1/4, 1/2, 3/4) with the authentic-mid and naive-lerp-mid swatches large and adjacent at frame center — a naive-lerp cross would make the pair identical, so the still now visibly breaks with the contract. Stage re-darkened to spec; pair labels plinth-mounted and legible at 400 px; 0 clipped pixels.

light-link-studio — the wall carried no wedge warmth (measured R−B −0.1 against +8…+58 across the pinned set) and ~40% of the frame was featureless grey. Re-lit to a measured +20…+50 warm wall pool; re-framed to ~75% subject fill with a TRACK_TO camera; LINKED/UNLINKED plaques upright and legible at card scale; rim-graze horizon stripe and clipped specular core eliminated; decoy stays legible-but-unlit (fill is deliberately 26 W, not the recipe's ~100 W — the decoy's fill-floor brightness is the ratio denominator, and the 3.0x gate needs it).

Contact-sheet gate (measured)

• docs/gallery/contact-sheets/vse-gamma-cross-contact-sheet.webp
• docs/gallery/contact-sheets/light-link-studio-contact-sheet.webp

Criterion	vse-gamma-cross	light-link-studio	Calibration (armature-bend / damped-track-aim / bmesh-gear)
Mean luminance	139.0	131.2	139.4 / 100.1 / 143.4
Wall wedge warmth (R−B)	+12.2 (pool)	+22.1 / +20.5	+8…+58
Subject fill	~76% w, ~85% h	~75% w	~70–90%
Clipped pixels	0 (max 238)	0 near-white	—
Thumbnail legibility	pair contrast + labels read at 400 px	pairing + plaques read at 400 px	—

Neither candidate sorts as the odd one out in its composite.

Falsification evidence (live run, Blender 5.1.2)

• vse naive-lerp expectation → exit 6, deviation 0.1138 at mid (matches README)
• vse endpoint-inclusive t convention → exit 5, t 0.129 vs 0.125 (matches README)
• vse swapped cross inputs → exit 4, GC inputs are not T1 -> T2 (matches README)
• light-link inverted to the decoy collection → exit 6, hero 0.288, hero/decoy 0.39x — caught by the ratio leg (README updated from the pre-relight 0.179/0.22x)

Test plan and evidence labels

Proven by live run (local headless binaries):

• .scratch/blender-5.1.2-windows-x64/blender.exe — self-reports Blender 5.1.2 (hash ec6e62d40fa9)
• .scratch/blender-4.5.11-windows-x64/blender.exe — self-reports Blender 4.5.11 LTS (hash 4db51e9d1e1e)
• vse check-only, exit 0 on both: samples=9 q_err=2.93e-03 (tol 0.005) lerp_mid_dev=0.115 (min 0.05)
• light-link check-only, exit 0 on both: 5.1 ratio=3.6x (min 3.0x), rise=206% (min 50%), hero_drift=0.0%; 4.5 ratio=3.4x, rise=213%, drift=0.0%
• light-link EEVEE luminance probe (the shipped scene, only render.engine swapped): 3.7x linked ratio on both 4.5.11 EEVEE Next (BLENDER_EEVEE_NEXT) and 5.1.2 EEVEE (BLENDER_EEVEE)
• All four falsification probes above, exit codes captured

Established by inspection only:

• Generated docs/gallery/index.html cards and both detail pages read after regeneration: alt texts clean, witnesses sections render as <li> bullets (both READMEs converted from ordered lists the generator flattened), no truncation or duplicated words
• blender-smoke.yml steps unchanged; neither step passes --engine

CI: the two blender-smoke legs (4.5 LTS, 5.1) on this PR remain the authoritative cross-version evidence.