feat(bot): say so, during the meeting, when a speaker's audio has no level

[TheMeinerLP]

Two live sessions produced transcripts of one hallucinated "Thank you." and of nothing at all, from WAV files exactly as long as the meetings. Bytes flowed in the right volume; the amplitude in them was at the noise floor. Nobody could tell whether that was a broken decode path or somebody who never spoke — and by the time anyone looked, the meeting was over and the recording was worthless.

Silence on its own is never the signal

People are quiet for most of a meeting, and Discord sends no packets while they are. Warning on "silence" would fire constantly and be ignored within a week. The condition here is narrower:

packets arrived for this speaker, decoded successfully, and every sample in them stayed at the noise floor, for a meaningful amount of received audio.

Three states exist and only the third is a fault:

	Reported
No packets at all	No — a participant not speaking
Packets that will not decode	No — already EndReason.DECODE_FAILURE
Packets that decode into nothing audible	Yes

The evidence threshold is counted in bytes of PCM, not wall-clock, and that is the point: somebody who transmits nothing for half an hour has produced no evidence about their microphone, and warning them would be the false positive that makes the feature unusable. Any audible packet discards that speaker's evidence and starts over.

Where, and when

In the bot at capture time — the decoded PCM is already there, and it is the only process that can say anything. RecordingService.voice_packet, last, after the write and the state machine: a report about the recording, never a step in making it.

During the meeting, once per speaker, never repeated. At the end the information is worthless because the recording is already lost.

Three outputs, each guarded separately so one failure cannot swallow another: a log.warning, a public message through the existing Announcer port, and session_participant.silent_audio_detected_at (migration 0006, nullable, not backfilled).

Only amplitude is ever read

No sample is buffered, logged or handed on. A peak is a number about loudness, not about content — which is what keeps this outside the consent model entirely: it is exactly as true of somebody who has not consented as of somebody who has.

Living in domain meant no numpy (the architecture test forbids it); stdlib array with max/min is a C-level pass over ~1920 samples and not the bottleneck.

Verification

610 tests pass (was 580; nothing weakened or removed), mypy and ruff clean. Twenty mutations were applied one at a time, each caught by the test claiming that behaviour.

Two mutants initially survived and both were real test defects, which is the part worth reading:

1. The threshold test asserted against SILENCE_PEAK_AMPLITUDE itself — self-referential, so it passed for any value including 0. Rewritten against literals, pinning the constant from both sides.
2. A _make_announcer seam existed only so a test could override it, so the production line never ran. The seam was deleted rather than tested; the test now stands in for the channel.

Independently re-checked here on two mutations not in that list: raising the threshold to 20000 and dropping the evidence window to zero each fail the right tests and only those.

Worth second-guessing

• SILENCE_PEAK_AMPLITUDE = 32 (~-60 dBFS) cannot be validated without real recordings. Not 0, because Opus does not always decode digital silence back to exact zeros — an equality test would find nothing on the very recordings this was written for. If a working microphone's room tone stays below that for 30 continuous seconds, that speaker gets one factual sentence and the recording continues.
• The channel message is posted before the database write. The in-meeting half is the time-critical one; the opposite order is defensible.
• announcer is a required constructor argument, which cost a _UnusedAnnouncer in recovery.py and four test fakes. A defaulted None would have been less churn and would let a future call site silently wire nothing.
• The wording is a module constant, English, not per-guild configurable. It renders through the same sandboxed environment as the link announcement, so that can be added without a code change.