As an admin, I can configure the diarization clustering threshold so that speaker merging is reduced

[julien731]

User Story

As an administrator, I can configure the diarization clustering threshold via environment variable so that the system can be tuned to favor splitting over merging speakers.

Spec: docs/specs/transcription-quality-improvements.md — US-5

Size: L

Acceptance Criteria

• DIARIZATION_THRESHOLD env var controls the clustering sensitivity
• Default value (0.715) preserves current diarization behavior
• Lower values reduce speaker merging (at the cost of potential over-splitting)
• Given the threshold is set too low, then speakers may be over-split — this is documented with a recommended range (0.4–0.8)
• Requires replacing WhisperX's built-in diarization wrapper with direct pyannote Pipeline usage

Truths

• T1: This is an admin-only setting (env var), not exposed in the upload form
• T2: This story involves the largest refactor — replacing the WhisperX diarization wrapper with direct pyannote pipeline calls
• T3: All existing diarization functionality must continue to work identically at the default threshold

Business Rules

None directly. Addresses Goal #2 (reduce speaker merging errors).

Edge Cases

• DIARIZATION_THRESHOLD set too low → over-splitting (one speaker appears as multiple). Documented with recommended range (0.4–0.8).
• Default threshold (0.715) → behavior identical to current

Dependencies

None strictly, but recommended to implement last due to refactor scope (replaces WhisperX diarization wrapper).