Returns human-written captions when they exist, otherwise the
auto-generated track. Output is truncated at `max_chars`; when that
happens the response ends with the timestamp to resume from, which
you pass back as `start_time` to read on.

Cheaper than reading the whole transcript: returns matching passages
with timestamps and deep links. Follow up with `get_subtitles` and a
time range to read a hit in full.