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

[fmasi]

Symptom: Too many speaker boxes; speaker identities are not consistent across chunks/sections.

Root cause

In the chunked path, ChunkProcessor stores segment speakers already source-prefixed (Local Speaker 1 / Remote Speaker 1) at TranscriberApp/Services/ChunkProcessor.swift:112 (tagging) and :132 (stored into ProcessedChunk.Segment). But SpeakerReconciler builds its cross-chunk mapping keyed on BARE names (Speaker 1) from chunk.speakerDatabase. So in TranscriberApp/Services/TranscriptionRunner.swift:219, chunkMapping[seg.speaker] (a prefixed key) misses every time and falls back to the raw per-chunk label — the embedding matching runs and is then discarded.

Second cause: ChunkProcessor.swift:143-144 stores only systemResult.speakerDatabase (remote/system audio); mic/local embeddings are never stored, so local speakers cannot be reconciled across chunks even if the key matched.

Proposed fix

• Store a per-chunk speaker DB keyed on source-prefixed names (Local → mic embedding, Remote → system embedding).
• Reconcile local and remote in SEPARATE pools (they are different streams).
• Ensure finalize()'s chunkMapping[seg.speaker] lookup matches the stored keys.
• Add a Core-layer regression test (SpeakerReconcilerTests) proving a person appearing across 3 chunks collapses to one global ID through the finalize mapping.

Acceptance criteria

• Reconciliation actually remaps speaker labels.
• The same person across chunks gets one stable label.
• Regression test added.
• Full test suite green.