Add AUC explainer film; size score decode buffer by duration

[calvin-archastro]

What

1. videos/auc-ml: an intuition-first narrated explainer, AUC in machine learning (11:48, 1080p30, 22 chapters).
   • It follows one six-payment example through thresholds, ROC, the pairs/ranking view, AP, and the PR conventions. It then surveys partial AUC, multiclass AUC, and other curves called "AUC".
2. packages/cli/src/music-render.ts: one-line fix. The reverb decode buffer had a fixed 256 MB cap, so any score longer than about 11 minutes failed with Cannot decode <track> for room processing. It is now sized from the score duration.

How the film is built

• Sentence-synced builds.
  • Each scene is one natural Kokoro take.
  • prepare-timing.mjs and sentence-bounds.mjs find the sentence onsets from the take's own pauses. Each boundary has a plausibility check.
  • Scenes animate on lineStarts(id)(k, offset, fraction), so each diagram step lands on the sentence that explains it.
• Score. src/score.ts is a quiet 60 BPM bed (piano, viola, cello, harp).
  • It ducks under speech using automation derived from the same timing.
  • The piano figure changes per lesson part.
  • It ends on an A→D cadence onto the end card.
• Bookends. A title card follows the cold-open hook, and an end card closes the film.
• Fonts. Schibsted Grotesk, Fragment Mono and Instrument Serif, all with full Latin coverage.
  • Note: videos/showcase/src/fonts/fonts.css labels a Space Grotesk latin-ext file as "latin". That file has no A–Z, so films importing it silently fall back to a system font. This PR does not touch that shared file.

Verification

• pnpm test passes: 51 tests.
• CLI and film typecheck pass.
• biome check is clean on the changed files.
• python3 videos/auc-ml/verify_math.py passes. It checks every on-screen number with exact fractions (7/9, 29/36, 55/72, McClish ≈ 0.649, and the rest).
• Four rounds of independent review on the rendered MP4 (visual, math/narrative, audio/music). The final round (r4):
  • Visual: SHIP.
  • Content: SHIP.
  • Mix: SHIP on measurements (−17.1 LUFS, −2.7 dBTP, no dead air except the final fade).
  • Music: SHIP on measurements.
• Listening is not verified. No human audition or ASR was available. Pronunciation ("A U C", "R O C", "Coco", "twenty-nine over thirty-six") and the music taste still need an ear.
• The rendered film and review kits are in the ignored out/ folder, so they are not part of this PR. To regenerate: pnpm --dir videos/auc-ml render.

🤖 Generated with Claude Code