[Epic] Diarization & speaker-tagging consistency

[fmasi]

Tracking issue for the diarization fix work. Addresses the two main user-facing symptoms: "too many speaker boxes" and "speaker names not remembered across sections".

Tier 1 — fix the broken reconciliation path

• [Tier 1] Cross-chunk speaker reconciliation is never applied (key mismatch + missing local embeddings) #64 — Cross-chunk speaker reconciliation is never applied (key mismatch + missing local embeddings)

Tier 2 — reduce over-segmentation

• [Tier 2] Merge tiny/adjacent same-speaker diarization fragments before labeling #65 — Merge tiny/adjacent same-speaker diarization fragments before labeling
• [Tier 2] Expose FluidAudio diarizer tuning knobs (clustering threshold / min duration / numSpeakers hint) #66 — Expose FluidAudio diarizer tuning knobs (clustering threshold / min duration / numSpeakers hint)
• [Tier 2] Optional expected-speaker-count hint in session dialog #67 — Optional expected-speaker-count hint in session dialog

Tier 3 — persistence across recordings

• [Tier 3] Cross-session speaker profiles + persistent renames #68 — Cross-session speaker profiles + persistent renames

Tier 4 — tuning & coverage

• [Tier 4] Make reconciliation cosine threshold and EMA alpha configurable #69 — Make reconciliation cosine threshold and EMA alpha configurable
• [Tier 4] Apply cross-speaker reconciliation to single-file / crash-recovery run() path #70 — Apply cross-speaker reconciliation to single-file / crash-recovery run() path

Recommended sequencing

1. Tier 1 first ([Tier 1] Cross-chunk speaker reconciliation is never applied (key mismatch + missing local embeddings) #64) — the reconciliation pipeline is currently a no-op; fixing it is the highest-leverage change.
2. Tier 2 next, starting with the fragment merge ([Tier 2] Merge tiny/adjacent same-speaker diarization fragments before labeling #65), then the diarizer knobs ([Tier 2] Expose FluidAudio diarizer tuning knobs (clustering threshold / min duration / numSpeakers hint) #66, gated by SDK research) and the optional speaker-count hint ([Tier 2] Optional expected-speaker-count hint in session dialog #67).
3. Tier 3 ([Tier 3] Cross-session speaker profiles + persistent renames #68) is likely a v0.8 feature.
4. Tier 4 ([Tier 4] Make reconciliation cosine threshold and EMA alpha configurable #69, [Tier 4] Apply cross-speaker reconciliation to single-file / crash-recovery run() path #70) as follow-up tuning/coverage.