Record exact transcript provenance for distillations (schema v4)
principles_files.transcript_id points at the transcript version each
distillation was actually derived from, immune to later canonical flips
or text regeneration; a distillation is stale when it no longer matches
videos.canonical_transcript_id (currently one: smD91VwGVSY, distilled
from youtube-auto before today's flip to butler). Text regeneration now
upserts text_files rows in place so references never dangle.
