[Tier 2] Merge tiny/adjacent same-speaker diarization fragments before labeling

[fmasi]

Symptom: Too many speaker boxes.

Root cause

TranscriberCore/SpeakerAssignment.swift assigns speakers purely by max temporal overlap (around :150-166) with no merging of short/adjacent same-speaker fragments. When the diarizer splits one person into fragments, this yields multiple Speaker N labels. buildSpeakerMap (:231) numbers speakers by first appearance with no fragment collapsing.

Proposed fix

• Before buildSpeakerMap, collapse diarization turns shorter than a configurable threshold (e.g. 0.5s) into the temporally-dominant neighbor.
• Optionally merge adjacent turns of the same speaker.
• Add a unit test in Core.

Acceptance criteria

• Short fragments no longer create standalone speakers.
• Unit test added.