fix(stt): don't clear the primary speaker on an unattributed segment

[chuenchen309]

Summary

_PrimarySpeakerDetector._update_primary_speaker clears _primary_speaker whenever a final transcript arrives without a speaker_id, because that case is folded in with detect_primary_speaker=False:

if sd.speaker_id is None or not self._detect_primary:
    self._primary_speaker = None
    return

Those are different situations. The flag being off means "never track a primary"; a missing speaker_id just means the diarizer couldn't attribute this one segment.

Clearing matters because None is also the "no primary yet" sentinel — the next attributed segment takes the it's the first speaker branch and is crowned with no loudness comparison. One unattributed segment is enough to hand primary to a quiet background speaker:

	primary
A speaks (amplitude 3000)	A
unattributed segment (no speaker_id)	None
B speaks (amplitude 50, 60x quieter)	B — skipped the RMS threshold

Without the unattributed segment in the middle, B is correctly rejected and A stays primary.

on_stt_event already guards speaker_id is None before touching _primary_speaker, so the unattributed segment was never going to be labelled either way — the only effect of clearing was losing the tracked state.

Scope

The detect_primary_speaker=False branch keeps its current behaviour; only the missing-speaker_id path changes. This is reachable through MultiSpeakerAdapter (exported from livekit.agents.stt), which needs a diarization-capable STT — it has no in-repo callers, so nothing inside the framework changes.

Testing

New tests/test_multi_speaker_primary_speaker.py, three unit tests: the quiet-speaker control, the bug case, and the detect_primary_speaker=False behaviour. The bug test fails on main and passes with this change; the other two pass on both. Existing stt tests (test_speaker_id_grouping, test_stt*) — 49 passed. ruff check/ruff format clean; mypy reports the same 14 pre-existing errors in other files with and without this diff, none in the file I touched.

Disclosure: written with Claude Code. I verified the reproduction, the reachability, and the test red/green myself.

[devin-ai-integration]

✅ Devin Review: No Issues Found

Devin Review analyzed this PR and found no bugs or issues to report.