Meetings: record the diarizer and voiceprint model that actually ran

[r3dbars]

What was wrong

Every meeting saved since Nemotron became the default (#1887) says diarization_engine: pyannote_offline. The formatter had that value hardcoded. The pipeline log Using local Parakeet V2 + PyAnnote pipeline and the raw footer + PyAnnote (local) were hardcoded too.

I checked the 2026-09-29 meetings in the local logs: both runs (14:08 and 14:24 UTC) used Nemotron fast128 with the redimnet2-b4 embedder. No pyannote fallback happened, so the files are wrong about their own engine.

What changed

• DiarizationEngine gets activeRunDescriptor (DiarizationRunDescriptor: backend + voiceprint model id). The protocol default is pyannote with no model, so existing conformers and test stubs keep working. DiarizationService builds it from activeBackend, so a Nemotron load failure reads pyannote. The voiceprint model is the injected embedder's id, else Nemotron's fallback embedder, else pyannote's native wespeaker.
• The pipeline reads the descriptor right before it diarizes (system track, or mic when local speaker split is on) and stores it in TranscriptionResult.diarization.
• The formatter writes diarization_engine: nemotron_offline | pyannote_offline | none. none means no diarizer ran, e.g. a mic-only meeting without split. It also writes an optional quoted voiceprint_model: "redimnet2-b4".
• The log line and the raw footer name the real backend. The pyannote path's "Offline diarization starting" log now carries backend too, so a fallback shows up in the logs.
• CaptureKit parser: voiceprintModel: String?. MCP engines gains voiceprint_model, left out of the JSON when absent.
• QA TranscriptValidator accepted only pyannote_offline, so it would have failed every correct Nemotron file. It now accepts all three values and still rejects unknown ones.
• docs/capture-format.md documents the values and the new key. Files saved before 2026-09-29 always say pyannote_offline, so in older files that value should be read as "unknown backend".
• Goldens regenerated: two-person call = Nemotron + ReDimNet2; imported = pyannote fallback + wespeaker; mic-only room = none.
• Hand-kept smoke source lists get DiarizationBackend.swift.

Older transcripts parse as before: no key is required, and unknown values are kept.

Tests

Behavior tests, no source reads:

• Pipeline end to end with a stub diarizer: Nemotron is recorded; Nemotron swapped to pyannote during model load records pyannote; a diarizer with no model name writes no voiceprint_model.
• Parser reads the new keys; the legacy file has no voiceprint.
• MCP JSON includes or omits voiceprint_model.
• QA validator accepts current and legacy engines and rejects unknown ones.
• Writer/kit contract and goldens updated.

Checks

• bash check.sh: swift test, CaptureKit/CLI/MCP/QA packages, integration smoke, source pins, test shape, and doc paths all pass. In that run, build.sh failed in 0.4 s because it ran before build-deps; run-tests.sh and the e2e smoke failed on the missing source-list entry (fixed in a1386a8). Rerun after the fix: build.sh --no-open ✅, run-tests.sh 19963/19963 ✅, run-e2e-smoke.sh ✅.
• concurrency-census --check fails with Sources/UI at 54 against a baseline of 51. This diff touches nothing in Sources/UI (the warnings are in MenuBar/Overlay files), so this was already on main.
• Review: a /code-review pass found nothing, but it ran in the same session that wrote the code. An independent review (codex review or a separate thread) is still owed per AGENTS.md.

🤖 Generated with Claude Code