- detect.py: _best_sibling_estimate() replaces a median/consensus across
neighbouring episodes with the best match found. Found directly on a
user-reported case (hpr0400): its three closest neighbours correlated
at ~0 since they don't share this episode's specific jingle variant,
but a fourth matched cleanly at 29.25s -- the median discarded that
real evidence because it was outvoted by neighbours that simply aren't
relevant, not because the intro genuinely doesn't match anything. Only
the case where nothing nearby correlates at all is still flagged
inconclusive.
- cli.py: trim --episode N with no prior detect for that episode used to
silently do nothing -- no output, no error, exit 0. Now errors clearly
and tells the user to run detect first; a batch trim over a range still
treats this as normal and reports a count instead. Also made trim's
metadata connection lazy, so the error path doesn't need network access
at all (a test that exercises it dropped from ~4s to instant).
- docs/DESIGN.md: documented both fixes, and the format-default/
--format/precision fixes from immediately before this round.
New tests: test_sibling_aggregation.py, test_cli_trim_errors.py.