Add automatic speaker diarization

[cemoody]

Summary

• Diarize the mixed system-audio track locally with FluidAudio’s offline Community-1/VBx Core ML pipeline and attribute transcript segments to anonymous session-local speakers.
• Use Parakeet word timestamps to split transcript turns at diarization boundaries; only emit a speaker-N label when the time-alignment confidence clears a configurable threshold, otherwise safely retain them.
• Record speaker source, alignment confidence, and overlap metadata in canonical transcript JSON.
• Add model-cache readiness to quill doctor and safe fallback behavior for unavailable diarization.

Quality and evaluation

• Adds an opt-in Core ML benchmark over 12 pinned WAV/RTTM smoke fixtures.
• Adds portable RTTM/DER scoring, fixture integrity, and word-level attribution tests.
• Documents fixture attribution and a consented meeting-corpus/release-benchmark checklist. The committed fixtures are explicitly documented as read-speech smoke coverage, not meeting-quality evidence.

CI

• Ubuntu 24.04 runs portable Foundation-only tests using pinned swift:6.0 in Docker.
• macos-15 builds Quill and runs deterministic tests.
• The model-backed benchmark remains opt-in (QUILL_RUN_DIARIZATION_BENCHMARK=1) on a separately provisioned macOS runner with a preloaded FluidAudio model cache; normal CI never downloads models.

Validation

docker run --rm -v /home/coder/quill:/src -w /src swift:6.0 swift test

Result: 12 tests passed (fixture checksums, RTTM parsing/DER semantics, anonymous-label mapping, collar behavior, and word-level attribution fallback/splitting).