v0.23.1 — fix: two people diarized as one in a short mono interview

An interview came back with a single speaker. Not a regression: the
Android capture path is unchanged since 0.16.1, pyannote 4.0.4 /
speaker-diarization-community-1 has been installed since July, and the
current stack reproduces the speaker splits of two earlier Android sessions. Two things went
wrong in turn:

1. pyannote produced one turn for the first 43 s of quick question/answer
   exchanges and gave the interviewee only one segment (43.4-47.8 s), even
   with num_speakers=2. The voices were easy to tell apart: per-segment
   embeddings put the two speakers' centroids at cosine 0.42.
2. label --auto's tiny-noise fold (absorb_tiny_speakers, <= 5 s and
   <= 3 segments) then treated that one 4.4 s, 21-word segment as noise and
   folded it into the interviewer.

Changes:

• millet/speaker_repair.py (new), wired into transcribe()'s mono /
  single-source path after diarization. Every ASR segment of at least
  0.5 s is embedded on its own (the WeSpeaker model voiceprints already use).
  Each cluster with at least 6 embedded segments gets one average-linkage
  2-way split. The split is accepted only if the smaller half has at least
  3 segments and 4 s of speech, its own segments are coherent (mean cosine

  = 0.5), and the two halves' centroids are at most 0.35 alike. The
  smaller half then joins an existing speaker it matches at cosine >= 0.6.
  Otherwise it becomes a new SPEAKER_NN, which needs stronger evidence
  (>= 8 s, centroids <= 0.30). A failure (no model, any error) leaves the
  transcript as diarized. Pure numpy (no scipy); the bipartition matches
  scipy's average linkage on 300 random cases. --no-speaker-repair /
  TranscriptionConfig.speaker_repair=False turns it off.
  Replay over the 100 diarized mono sessions on the server:

  • the incident session: 9 answers (8.3 s) move into the interviewee's
    cluster;
  • 56 single-speaker sessions: untouched;
  • 15 group meetings: 11 moves into an existing speaker and 8 new speakers,
    not yet checked by ear.
• absorb_tiny_speakers keeps a small cluster that holds a real sentence
  (voiceprint.is_dense_utterance: >= 1.5 s, >= 6 words, >= 1.5 words/s, the
  test voiceprint learning already used). Fillers still fold. Replaying
  704 sessions, 6 tiny clusters stay raw for a human instead of being folded.
• Relabeling dedupes frontmatter participants. The find-and-replace path
  turned SPEAKER_01 into Nora next to the Nora the model had already
  extracted, so Nora was listed twice. Duplicates now collapse into one
  entry (first wins, missing role/channel filled from the others).