Spot-check auto-caption quality across eras; add assessments
New stage-2b tool scripts/transcribe_video.py (direct download with
decodo-proxy fallback, S3 audio archive to s3://joedistilled/media/audio/,
butler transcription, full videos.db registration) and
scripts/compare_transcripts.py (token-alignment metrics between
transcript versions). Schema v2 adds the assessments table recording
head-to-head comparisons (metrics / llm-judge / human) so canonical
promotions cite evidence.
Butler re-transcribed 7 videos (4 from 2015-2016, 3 from 2024-2026; 8
pairs assessed counting the channel's first video). Verdict: no
systematic era problem -- 7 of 8 pairs agree at 0.97-1.00 similarity.
Old-era captions garble occasional key terms ("oh man" for OMAD,
"ay hoffler" for Ori Hofmekler); recent ones are near-verbatim.
Canonical flipped to butler for the 3 videos where it was judged
strictly better (plus 4gcAjbAlcxk earlier); ties keep youtube-auto.
Audio for all 7 archived in S3 and registered in media.
