fix: recompose the four genuine framing defects (worklist drained)

[TMHSDigital]

Recompose the four genuine framing defects

Final pass on the framing worklist. The corrected survey (#109) left four examples whose subjects were genuinely mis-framed — not camera-nudge cases. All four are recomposed, gated with check_framing, and verified on both binaries. The worklist is drained; no example needed a deviation classification (none is a radiating/world/atmosphere subject — all are objects on stages).

Corrected worklist consumed

example	defect (corrected survey)
depsgraph-export	left cube amputated at left+bottom edges; fill 0.984x
vertex-weight-limit	base cropped at bottom; off-center dead space; 0.247x/0.861y
triangulate-tangents	under-filled buckler; 0.334x/0.661y
text-version-stamp	under-filled; dead lower third; 0.663x/0.567y

Before → after (gated render numbers, 5.1.2; 4.5.11 identical unless noted)

example	before	after	what changed
depsgraph-export	0.984x/0.822y, touch L+B	0.812x/0.622y, L0.069 R0.119 B0.178 T0.200	aim-based camera (0,-9.6,2.6)→(0,0,0.9); pair recentered (subject was 5.9 units wide vs a 5.3-unit frame)
vertex-weight-limit	0.247x/0.861y, touch B	0.206x/0.867y, L0.378 R0.416 B0.050 T0.083	more frontal camera (4.6,-8.2,2.3), aim 1.5→1.25; base fully in frame. x stays narrow by subject geometry; band met on y
triangulate-tangents	0.334x/0.661y	0.459x/0.833y, L0.278 R0.262 B0.039 T0.039	2×-class move-in, 5 iterations (4.1/4.8/5.6/6.2/6.35 + aim split at 1.42) around the leaning disc's span. More than reframe (disclosed): Key 500→320, Glint 400→150 to clear a pre-existing dome blowout the larger subject exposed (30.7 % of the dome region ≥ 252 → max 231, near-white 0.00 %; old hero clipped in the same region at smaller scale)
text-version-stamp	0.663x/0.567y	0.719x/0.478y, L0.025 R0.025 B0.039 T0.144	aim-based camera (0,-6.0,1.5)→(0,0,1.05); dead third gone. Side margins bind at 0.025 via the underline bar (widest element). 4.5.11: 0.716x — the constant-width scaling absorbs the longer "4.5.11 LTS" body

No check code or closed forms changed; every example's printed check values are unchanged (depsgraph 8/98/98; vwl influence/pose/LBS numbers; tt tris=720 + error suite; tvs body/extents/faces=2917). No closed form needed re-falsification.

Contract legibility (Pass 2 falsification variants, all rendered live)

• depsgraph-export — broken (subsurf cleared) renders two faceted cubes vs working cube + smooth dome (mean |Δ| 4.53, 2.91 % pixels differ): the evaluated-vs-base witness is visible in the reframe.
• vertex-weight-limit — broken (no prune) is pixel-identical by design: pruning preserves the pose, so the influence-cap contract is numeric; the still witnesses the rig and pruned-weight zones. Named, not hidden.
• triangulate-tangents — broken (back cap dissolved to an ngon, no calc_tangents) is pixel-identical: the concentric sweep is UV/bump-driven, so the mikktspace contract is numeric; the still witnesses the UV layout/rings. Named, not hidden.
• text-version-stamp — broken (stale "9.9.9" body) is unmistakable at the reframed size; the live-version witness strengthened.

Contact sheets (Pass 5)

Calibration luminance: armature-bend 70.8 / damped-track-aim 44.5 / bmesh-gear 77.7.

• depsgraph-export — 59.2 ✓ in range
• vertex-weight-limit — 52.7 ✓ in range
• triangulate-tangents — 66.3 ✓ in range (the relight pulled it into range)
• text-version-stamp — 87.3, above range: its brighter warm-backdrop studio predates the current standard (accepted look); camera-only reframe, flagged as information

Per-criterion verdicts: stage darkness ✓ all; wedge warmth ✓ all; saturation ✓ all (materials untouched); thumbnail legibility ✓ all four (rings, arm, stamp, cube/dome all read at cell scale). tt dome no longer clips; no visible light shapes anywhere.

Visual-pass log

• Pass 1 (full size + helper numbers): dgx 1 iteration; vwl 2 (first stop still cropped the base and overshot 0.917y); tt 5 (the leaning disc+stand span fights the frame — distance/aim split needed); tvs 2 (first stop left hero fill at 0.609 — the bar, not the stamp, was filling the frame).
• Pass 2: variants above — visible difference named for each.
• Pass 3 (100 % crops): no clipping after the tt relight (dome max 231); no visible helpers; dgx cube edges clean; vwl base fully in frame; tvs bar crisp.
• Pass 4 (Playwright, full-page captures): index at 1280/390 plus all four detail pages — new heroes, teaches, witnesses, and the framing-gate README line all correct; generated HTML read character by character (alts complete, no truncation).
• Pass 5: sheets above.

Runtime

Check-only (smoke invocations) for all four, both binaries: exit 0; 753–857 ms (5.1.2), 614–636 ms (4.5.11) — unchanged paths. 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). CI floats within series per CLAUDE.md.

Live-run vs inspection-only

• Live run (5.1.2): all reframe iterations, gate runs, falsification probes (lens breaks, exit 10 ×4), Pass-2 variant renders, webp conversion, contact sheets, gallery regen, Playwright captures, dome blowout measurements.
• Live run (4.5.11): all four render paths + check-only.
• Inspection only: the luminance-range judgments on the two pre-existing brighter stages (shader-node-group from fix: corrected framing survey + marginal camera-nudge sweep (7 examples) #109, text-version-stamp here), and the assumption that CI smoke runtimes track local timings.