fix: stabilize transcription session state and diarization fallback

[Newarr]

What

This PR stabilizes transcription session model selection, makes diarization failures degrade visibly instead of silently, keeps renamed settings rollback-safe, and aligns the re-transcription notification copy.

Why

This is the second PR in a 2-PR stack, and it depends on #7.

The main issue here was not one isolated bug. The engine was still rereading mutable settings after startup, which meant the requested transcription model could drift during the same session. On top of that, diarization failures could disappear silently, and renamed settings keys were only safe in one direction.

Separating this from the logger cleanup keeps the review focused on behavior.

How

• captures the requested transcription model at session start and uses it for backend preparation, flush interval selection, and prepare-failure cache clearing
• keeps the active model session-scoped instead of rereading mutable settings mid-run
• surfaces the first diarization feed/finalize failure, then continues transcription in best-effort mode
• dual-writes the renamed settings keys for downgrade safety
• updates the batch-complete notification wording to Re-transcription
• includes the small teardown follow-up to nil out vadManager and remove dead engine error code

Testing

• swift test
• Result: 310 tests passed, 0 failures
• adds regression coverage for active-model stability, diarization failure handling, settings rollback compatibility, and notification wording

Migration / rollback

No schema migration.

Rollback is straightforward, and the settings changes explicitly preserve downgrade compatibility by continuing to write the legacy keys.

Stack

• Base PR: refactor: unify logger categories and reduce logging noise #7
• This PR targets stack/logging-cleanup in the fork to keep the stacked diff small until the upstream base work is ready