feat: recover dropped dictation commands via blank-gap TDT repair

[timlenardo]

Summary

• TDT suppresses dictation commands ("period", "comma", "end quote", etc.) when LM context makes blank more probable than the token — producing a silent ≥960ms blank gap with real speech inside it
• After the main TDT inference, scan for inter-token gaps ≥12 frames, confirm speech via Silero VAD, run a focused re-inference on the gap window, and insert the result if it matches a known dictation command
• VadManager is initialized once in AsrManager.initialize() and reused; repair is a no-op if VAD is unavailable

Validation

Validated on 536 eval recordings and 2,757 real history audio files:

• 7.9% of real utterances have a qualifying gap (≥960ms with VAD-confirmed speech)
• 3 dictation command hits found in history audio (all "end quote" / "unquote")
• Window TDT inference measured at 72ms median (not a full 15s pass)
• Expected average latency overhead: ~6ms per transcription

Files

• `Sources/FluidAudio/ASR/DictationRepair.swift` — repair logic as an `AsrManager` extension
• `Sources/FluidAudio/ASR/AsrManager.swift` — adds `vadManager` property, initializes in `initialize(models:)`
• `Sources/FluidAudio/ASR/AsrTranscription.swift` — calls `repairDictationGaps` in the single-chunk path after `processTranscriptionResult`
• `Tests/.../EvalHistoryRepairTests.swift` — eval harness: runs repair against all eval recording audio files, writes `/tmp/repair_comparison.json`
• `Tests/.../HistoryAudioGapScanTests.swift` — scan harness: measures trigger rate and recovery on history audio

Test plan

• Build passes with no new errors
• `swift test --filter EvalHistoryRepairTests` produces repair output at `/tmp/repair_comparison.json`
• Transcribing audio with a spoken "period" or "end quote" gap correctly recovers the command
• Audio without qualifying gaps shows no change in output or latency