feat(videos): Level 2 post-render QA — per-beat vision-judge frame check

[jjackson]

Context

PR #452 ships Level 1 of post-render QA: deterministic stream sanity + sentinel-string detection + beat-midpoint frame extraction + PIL variance check for all-black/white frames. Catches the major classes of "render exited ok but the video is broken" failures that have shipped this week.

What Level 1 can't catch (because there's no vision intelligence in the pipeline):

• Text overflows the frame edges (rendered glyphs are wider than the layout assumed)
• Wrong cycle step is highlighted given the audible narration
• Stat number is unreadable due to color contrast with background
• Asset failed to load (frame is rendered but contains a Remotion error overlay or broken-image glyph)
• Captions visually overlap a clip's lower-third
• B-roll thumbnail is the wrong shot (someone swapped the gdrive id but the alias is still @field-walking-towards-house)

These are all "is this frame acceptable" judgments. A vision-capable model can answer them with a small rubric per beat-kind.

Proposal

Add scripts/qa_render_vision.py (or extend qa_render.py --vision) that:

1. Reuses Level 1's per-beat midpoint frame extraction.
2. For each beat, builds a small rubric prompt parameterized by the beat's kind:
   • intro_hook → does the tagline text fit and read cleanly?
   • intro_cycle → which of {Learn, Deliver, Verify, Pay} appears highlighted? Does it match the caption text under it?
   • intro_handoff → is the program name visible?
   • body_scene → does the lower-third overlay obscure subject of the b-roll?
   • body_problem_stat, body_impact_stats → does the big number fit, is the caption readable, is the source line visible but de-emphasized?
   • body_product_beats → are the three phone-frame screencaps positioned correctly, with the right caption beneath?
   • outro_cta → logo + tagline + "Request a demo" link all present?
3. Sends each frame + rubric to a vision-capable model (claude-opus-4-7 or similar), gets a structured verdict (OK / WARN / FAIL + short explanation).
4. Surfaces results in the same OK/WARN/FAIL summary as Level 1.

Open questions

• Cost: 8 beats × per-render. If we run on every render that's ~$0.05–$0.20 of vision tokens. Acceptable but worth capping to opt-in (--vision) initially.
• Determinism: vision judges drift between model versions. Tag the model id in the report so a future "this used to pass" investigation has a known baseline.
• False positives: model might flag a legitimate stylistic choice (e.g. low-contrast Headline by design). Need a per-program override path that says "ignore the contrast warning on impact[1]".
• Frame selection: midpoint might miss the worst frame in a beat (e.g. a stat that's mid-animation). Maybe sample at midpoint + last-2s + first-2s and judge the worst of the three.

Acceptance

• --vision flag added to qa_render.py
• Vision judge runs per beat against the midpoint frame
• Token spend ≤ $0.25/render at default rubric size
• README / CLAUDE.md note on when to use vision vs. deterministic-only
• One regression case captured as a golden fixture: a chc render with the original "E2E-85648Z%" stat — vision judge should fail it on the contrast/overflow rubric, not just the sentinel regex

Blocks: nothing — Level 1 stands alone. This is purely additive.