fix: corrected framing survey + marginal camera-nudge sweep (7 examples)

[TMHSDigital]

Corrected framing survey + floor sanity-check + mechanical marginal sweep

Follow-up to #107/#108. Three parts: fix the survey's scene resolution and re-measure everything, sanity-check the 0.02 margin floor against the calibration set, then camera-nudge only the genuinely marginal cases into the band and wire check_framing into them as the acceptance gate. Recomposition cases deferred to the ROADMAP worklist.

Part 1 — the survey bug and its fix

Bug: the #107 survey measured each example against bpy.context.scene after main() returned. Any example that stills from a non-context scene was measured against the wrong scene. vse-cut-list renders its hero from a dedicated Bay scene via bpy.ops.render.render(..., scene=sc.name) — it was reported "unsurveyable" when it was actually mis-surveyed.

Fix: the corrected survey registers a persistent bpy.app.handlers.render_pre handler (persistent because examples call read_factory_settings, which clears plain handlers), records every (scene, camera) render each example issues, and measures the last render — the gallery still. A shipped helper bug fell out of the same root cause: _matte_render rendered the active scene, so a passed non-context scene would have been matte-measured wrong too. gallery_framing._matte_render now renders the scene it was handed (same scene= idiom vse-cut-list uses); both current callers measure unchanged (0.803/0.733 on 5.1.2 and 4.5.11).

Old vs corrected table (silhouette, union-of-renderables): all 38 examples re-measured; 37 of 38 identical to the old table (they render from the context scene). The one change:

example	old	corrected	note
vse-cut-list	unsurveyable (context scene had no camera)	fill 0.772x/1.000y, margins L0.113 R0.116 B0.000 T0.000, touch bottom+top	measured against the wrong scene before; genuinely out of band — monitor top cropped, desk bleeding bottom

World/atmosphere example (sky-texture-sun-elevation) measures identically (0.369x/0.489y) — the #108 artifact classification stands. No previously-measured classification changed.

Part 2 — floor decision: keep 0.02

Calibration evidence (corrected survey): armature-bend min margin 0.039, bmesh-gear min 0.033; damped-track-aim sits at 0.0 but is a documented deviation (#108), not an enforcing reference. Across the accepted in-band corpus the tightest margin anywhere is 0.028 (shape-key-blend, gn-instance-grid). The enforcing reference standard clears the floor by ≥ 65 %, the corpus by ≥ 40 %, and nothing accepted lives in the 0.013–0.019 gap — 0.02 cleanly separates accepted work from curve-bevel-arc's 0.012 near-contact. Unlike damped-track-aim (which tripped the fill rule by design), no reference member trips the margin floor, so the floor stands. MARGIN_MIN stays 0.02; no constant change, no re-floored survey needed.

Part 3 — mechanical sweep (7 examples, camera-only)

Per-example before → after (corrected-survey numbers → gated render numbers, 5.1.2; 4.5.11 identical unless noted):

example	before	after	nudge
curve-bevel-arc	R margin 0.012; fill 0.784x/0.683y	fill 0.738x/0.633y, margins L0.219 R0.044 B0.083 T0.283	cam (2.4,-3.4,1.7)→(2.55,-3.62,1.78)
driver-wave	fill 0.969x over; margins 0.016	0.869x/0.333y, L/R 0.066	cam y -13.5→-15.0
shader-node-group	fill 0.969x over; margins 0.016	0.869x/0.650y, L/R 0.066	cam y -7.8→-8.7
swatch-grid	fill 0.944y over	0.769x/0.878y, L/R 0.116 B/T 0.061	ortho_scale 7.8→8.4
png-exr-alpha	touch B 0.000	0.725x/0.628y, B 0.028	aim z 1.65→1.38 (a pull-back overshot under-band; pitch only)
gp-lineart-contour	fill 0.978y over; touch B	0.463x/0.867y, B 0.044	cam (3.4,-4.2,2.4)→(3.95,-4.9,2.62)
vse-cut-list	fill 1.000y; touch B+T	0.506x/0.850y, B 0.067 T 0.083	BayCam (2.0,-5.2,1.72)→(3.05,-7.9,2.0), aim z 1.28→1.20

All seven are wired to check_framing (exit 10, before the beauty render) as the acceptance gate. No check code or closed forms changed: swatch-grid's verify_png third/half cell sampling tolerates the new ortho_scale (confirmed distinct_regions=6 at both the default and the exact smoke invocation --engine cycles --samples 8 --width 640, both versions); every other example's printed check values are unchanged. vse-cut-list was the most-broken (frame-filling before), so its composition change is the largest — still camera-only; anything needing recomposition was deferred.

Deferred to the ROADMAP recomposition worklist: depsgraph-export (0.984x, cube amputated), vertex-weight-limit (0.247x/0.861y, off-center, base cropped), text-version-stamp (0.663 max, dead lower third), triangulate-tangents (0.334x/0.661y, 2×-class move-in). ROADMAP updated accordingly.

Falsification (each new gate proven to fail, then restored)

curve-bevel-arc lens 65 → margins FAIL, exit 10. driver-wave lens 55 → fill 0.994, exit 10. shader-node-group lens 72 → fill 1.000, exit 10. swatch-grid ortho_scale 6.5 → fill 1.000, exit 10. png-exr-alpha lens 64 → fill 0.916, exit 10. gp-lineart-contour lens 62 → fill 1.000, exit 10. vse-cut-list lens 66 → fill 1.000, exit 10. All: no artifact written on violation; all restored (diff-verified).

Contact sheets (Pass 5)

Each pairs the candidate with the pinned set (armature-bend 70.8, damped-track-aim 44.5, bmesh-gear 77.7 mean luminance; reading order TL→BR, candidate last):

• curve-bevel-arc — 48.6 ✓ in range
• driver-wave — 55.7 ✓ in range
• shader-node-group — 89.2, above range: its lighter studio variant (0.055 floor vs 0.03 house) predates this camera-only pass — flagged as information, not altered here
• swatch-grid — 82.1, marginally above range (bright swatches by design; pre-existing)
• png-exr-alpha — 68.2 ✓ in range
• gp-lineart-contour — 47.7 ✓ in range
• vse-cut-list — 54.3 ✓ in range; subject presence lower than calibration but in family with the bay/monitor examples

Per-criterion verdicts: stage darkness, wedge warmth, saturation, and thumbnail legibility ✓ for all seven (no lighting or material touched); subject fill now band-consistent for all seven.

Visual-pass log (per touched example)

• Pass 1 (full size + helper numbers): one nudge iteration each except png-exr-alpha (2 — first pull-back overshot under-band), gp-lineart-contour (2 — first stop landed y at exactly 0.900, on the ceiling), vse-cut-list (3 — desk's near corner needed distance and aim drop). All final renders reviewed at full size; compositions preserved.
• Pass 2 (thumbnail, 400 px): all seven read instantly; vse-cut-list is the weakest but legible and in family.
• Pass 3 (100 % edge crops): no clipping, no blown highlights, no visible helpers on any of the seven.
• Pass 4 (Playwright, full-page captures per the scroll-anchoring note): gallery index at 1280 and 390 + all seven detail pages — new heroes, teaches, witnesses, and the new framing-gate README line all render correctly; generated HTML read character by character (img alts complete, no truncation artifacts).
• Pass 5: sheets above.

Runtime

• Check-only (the smoke invocations) for all touched examples, both binaries: exit 0; timings sub-second to ~2.6 s (vse-cut-list --check-pixels renders tiny frames as before) — unchanged code paths.
• swatch-grid's smoke step already renders; the gate adds two 320×180 8-sample mattes (~2 s) to that one existing render step. All other smoke steps are check-only and untouched.
• 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); its jobs may run a newer 4.5.x point release.

Live-run vs inspection-only

• Live run (5.1.2): corrected 38-example survey, all seven nudge renders + gate, all seven falsification probes, webp conversion, contact sheets, gallery regen, Playwright captures.
• Live run (4.5.11): all seven render paths + check-only; swatch smoke-style render; framing numbers identical unless noted (gp-lineart 0.469x vs 0.463x — GP stroke rasterization; inside band).
• Inspection only: that the corrected survey's union-of-renderables fill approximates per-example hero fill (placards are small); that CI smoke runtimes track local timings.