feat(diar): min-cut-interval to suppress spurious rapid turn cuts

[m5d215]

What

Don't fire a turn boundary within 1.5s of the previous cut (and don't update
the tracked speaker while suppressed). A short cross-talk excursion is absorbed
when the original speaker returns before the interval elapses. Debounce already
guards single-update flicker; this guards rapid real flips and diarizer
flicker.

Why / validation

Same-audio A/B — the production build (no min-cut) vs this — over a full day of
real calls:

• sub-1.5s inter-boundary gaps: 14% → 0%
• diarizer flicker no longer chops a single speaker's phrase into a tiny
  standalone record (observed: a phrase that production cut into its own 5-char
  record stays attached to its sentence here)
• normal turns (>1.5s apart) unaffected

Honest caveat

This does not move the overall share of ≤5-char records. Measurement shows
those are mostly SpeechAnalyzer-side short finals, not diarizer over-cuts —
trimming them is a separate lever (the Transcriber min-length filter), not in
scope here. So this is a clean, low-risk quality fix (removes spurious
fragmentation, no downside on normal turns) rather than a fix for the
micro-fragment count.

minCutInterval is a single tunable constant in SpeakerSegmenter.

Test plan

• swift build clean
• make install, production app runs with it
• full-day same-audio A/B vs the no-min-cut production build