fix(app): gate on-device transcription hallucinations

[moni11811]

Theory
On-device transcription accepted low-quality local Whisper/Speech output as final text. Short hallucinations, repeated filler, and weak local transcripts could leak into conversations instead of being rejected or falling back.

Scope

• Adds on-device transcript quality gate.
• Applies gate in socket transcription flow.
• Tightens on-device provider handling.
• Sets iOS speech request dictation hint and punctuation.

Cut from #9084

• Meta glasses runtime/UI.
• Multi-device saved-device work.
• Native BLE/device integration churn.
• Docs/plugin examples/auth changes.

Tests

• Red first: flutter test test/unit/on_device_transcription_upgrade_test.dart failed on missing OnDeviceTranscriptQualityGate.
• Green: flutter test test/unit/on_device_transcription_upgrade_test.dart passed 5 tests.

Notes

• Draft PR. Surgical transcription slice from the latest coding spree.
• Low overlap with feat(desktop): microphone picker with Ray-Ban Meta glasses support #9181; this branch does not touch app/lib/pages/home/page.dart.

[Git-on-my-level]

Thanks for extracting this into a focused draft — the scope is much easier to reason about than the larger branch, and the local Whisper quality gate is a useful direction.

A couple of things I think should be addressed before this is ready for maintainer review:

• The new gate is only wired into OnDeviceWhisperProvider. On iOS, SttProvider.onDeviceWhisper returns OnDeviceAppleProvider, so the native Apple/Speech path still accepts short/filler results as final text. The PR description says it gates on-device Whisper/Speech output, but the iOS change currently only sets dictation/punctuation hints. Please either apply the same quality gate to OnDeviceAppleProvider too, or narrow the PR description/tests to make it clear this only gates the non-iOS local Whisper path.
• The regression tests are mostly static string checks for the iOS path. It would be stronger to add behavior-level coverage for the provider path that should be gated, especially the iOS/native path if that remains in scope.

Not approving because this is still a draft and there is no passing check summary available. Once the iOS path/scope mismatch is resolved and checks are available, this should be a good candidate for another pass.

[moni11811]

Addressed draft review in 160d818.

Changes:

• Wired OnDeviceTranscriptQualityGate into OnDeviceAppleProvider, not only local Whisper.
• Apple/native iOS path now drops low-energy filler transcripts before returning final STT text.
• Added behavior-level mocked iOS speech-channel test for Apple provider.
• Cleaned analyzer warnings in the on-device provider files.

Local checks:

• flutter test test/unit/on_device_transcription_upgrade_test.dart
• dart analyze lib/services/sockets/on_device_apple_provider.dart lib/services/sockets/on_device_whisper_provider.dart lib/services/sockets/on_device_transcript_quality_gate.dart test/unit/on_device_transcription_upgrade_test.dart

[Git-on-my-level]

Good progress since the last pass — the quality gate is now wired into OnDeviceAppleProvider (not just local Whisper), and the new mocked iOS speech-channel test gives much better confidence than the earlier static-string checks.

A few observations for the next iteration or for the human maintainer:

1. Dead parameter: OnDeviceTranscriptQualityGate.filter() accepts duration but never uses it. Either add a TODO for future duration-based heuristics or remove it.
2. Magic thresholds: The low-energy threshold (avg |sample| < 180) and the six-token filler list are uncalibrated. They're reasonable starting points, but real-device audio validation would strengthen confidence that legitimate short utterances aren't dropped.
3. State lifecycle: _lastAcceptedFiller is per-provider-instance. If providers are short-lived the duplicate-filler suppression won't persist across polling cycles; if long-lived it may suppress fillers across unrelated conversation contexts. Worth a comment documenting the expected lifecycle.
4. Polling window change: The iOS buffer change (5s → 3s, minBufferSizeBytes doubled) affects the transcription latency/quality tradeoff. Real-device testing would help confirm shorter windows don't regress accuracy.

Not approving: still a draft with no CI checks available. Once CI confirms the new test suite passes, the needs-tests label can be lifted. Keeping needs-maintainer-review — the thresholds and polling-window changes warrant human judgment before merge.

Thanks for the focused extraction from the larger branch — the scope is clean and the quality gate direction is valuable for Omi's transcription reliability.