Build Speech-to-Text Adapter with Permission Handling

Define SpeechToTextAdapter as an abstract class and provide FlutterSpeechToTextAdapter as the concrete implementation. This allows tests to inject a fake adapter and allows future swapping of the underlying plugin. Use a StreamController internally; expose only the stream, not the controller. Subscribe to WidgetsBindingObserver for lifecycle hooks — register the adapter as an observer in its constructor and deregister in a dispose() method.

The Riverpod provider for this adapter should be a Provider so any feature can read it via ref.watch without triggering unnecessary rebuilds. For the no-speech timeout, use a Timer that starts when recording begins and resets on each interim result; on timeout, cancel the timer, emit the error, and call stop(). Important: Blindeforbundet workshop notes explicitly state that recording during the actual peer-mentor visit is unwanted — ensure the adapter's start() method cannot be called from any screen shown during an active session (enforce this via a guard in the upstream report screen, not in the adapter itself).

Write unit tests using flutter_test with a MockSpeechToText class that simulates the speech_to_text plugin interface. Test matrix: (1) requestPermission granted → start() succeeds, (2) requestPermission denied → start() throws SpeechPermissionException, (3) locale nb_NO selected when device locale is Norwegian, (4) locale en_US selected as fallback for non-Norwegian locale, (5) caller-specified locale overrides detection, (6) 5-second silence → noSpeechDetected SpeechError emitted and recording stops, (7) AppLifecycleState.paused → stop() called automatically, (8) interim results streamed correctly, (9) final result emitted with isFinal=true. Integration test: on a physical iOS device, confirm the permission dialog appears, granting it allows recording, and the transcript appears in the stream.