Fetches HPR's own metadata, audio, and official transcripts live; falls
back to local speech recognition (tiny model, matching HPR's own
hpr-get-and-transcribe.bash tooling) only when needed, and to
correlation against vendored reference clips or sibling episodes
otherwise. License-aware, resumable, and never guesses -- flags
anything it isn't confident about for manual review rather than
producing a plausible-looking wrong cut.