feat: speaker-turn segmentation for the system source (streaming Sortformer)

[m5d215]

What & why

The system source is a mix of whatever the call's far side outputs. When two
or more remote people talk back-and-forth, SpeechAnalyzer merges the exchange
into one long final — a single record that scrambles multiple speakers and
loses meaning downstream.

This adds speaker-turn segmentation to the system source: when the speaker
changes, the current utterance is cut so each turn becomes its own record. The
goal is segmentation, not speaker identification — no labels are emitted, so
imperfect speaker tracking (4-speaker cap, label permutation) is tolerable.

This reverses an earlier decision to ship without diarization. That call was
about labeling (low-accuracy spk_N hurts a downstream LLM) and about cost
(67s startup, ANE contention). Reframing the requirement to boundary detection
removes those objections — see CLAUDE.local.md for the full rationale.

How

• SpeakerSegmenter runs FluidAudio's Streaming Sortformer (CoreML) on the
  system tap audio, in parallel with the transcriber.
• On a detected turn boundary it calls Transcriber.finalize(through:), letting
  SpeechAnalyzer itself cut the record at the turn — so no token-time
  alignment or text re-splitting is needed, and the jsonl schema is unchanged.
• The diarizer loads with .cpuAndGPU so it stays off the ANE that the two
  SpeechAnalyzer instances use (avoids the E5RT crash, and lets a second build
  run in parallel).
• Boundary detection uses the live (tentative) speaker with a small debounce.

Validation

Ran a full work day on real calls, alongside the previous build for A/B:

• ~31% fewer ≥100-char multi-speaker merges (the target problem).
• Startup ~1.3s (model cached; first run downloads once), no 67s regression.
• No E5RT / crash, coexisted with the parallel build all day.

Known rough edges (follow-ups, not blockers)

• ~1s boundary bleed — boundary detection lags ~1s, so the next speaker's
  first ~1s can land in the previous record. Tightening needs
  finalize(through: boundaryTime) with diarizer↔transcriber clock alignment.
• Back-channel over-segmentation — rapid "はい"/acknowledgement exchanges get
  cut into many tiny records (system ≤5-char records ~13% → ~20%). A
  min-cut-interval would suppress sub-~1.5s speaker flips.

Config

• Default on. Set STENO_DIAR=0 to opt out (also degrades gracefully to
  plain transcription if the model fails to load).
• First run downloads FluidInference/diar-streaming-sortformer-coreml from
  HuggingFace, then caches locally. No other new network use.

Test plan

• swift build clean
• make install, production app runs with diarization enabled (no env needed)
• full-day real-call run, A/B vs prior build