fix(engines): restart capture on mid-session input-device config change (WhisperKit, Apple Speech, SpeechAnalyzer)

[initcore0]

Summary

PR #224 fixed the mid-dictation word-loss dropout for Parakeet by observing AVAudioEngineConfigurationChange and rebuilding capture onto the same feed stream. The same hole existed in the three other streaming engines — each owns a capture tap and none observed the notification, so a mid-session input-device disconnect/switch/format renegotiation left the session showing "Listening…" while capturing nothing.

This ports the Parakeet pattern (selector-based observer, makeRoutedEngine/startCapture split, restart-then-loud-error, shared CaptureConfigChangePolicy gates) to all three:

• AppleSpeechEngine — rebuilds the engine + tap appending to the same SFSpeechAudioBufferRecognitionRequest; the recognizer sees one continuous audio stream and keeps everything already heard.
• SpeechAnalyzerStreamingEngine — rebuilds the engine + tap yielding into the same raw-buffer AsyncStream. The analyzer-format converter is now adaptive (rebuilt when the incoming buffer format changes): a converter pinned to the start-time mic format would silently drop every buffer from the replacement device — capture restarted, transcript still dead.
• WhisperKitStreamingEngine — the mic belongs to WhisperKit's AudioProcessor, so the rebuild lives on WhisperKitStreamHandle.restartCapture (mirrors the internal setupEngine via public API — native-format tap → resample to 16 kHz mono → processBuffer — appending to the same decode buffer the realtime loop polls). The engine arms the observer on the processor's engine object at first state diff and re-arms on the replacement engine after a restart.

Restart failure (e.g. no input device left after the change) surfaces loudly through the session error path via CaptureConfigChangePolicy.restartFailedMessage.

Testing

• swift test — 1904 tests green; the shared gates stay pinned by CaptureConfigChangePolicyTests (header updated to cover all four engines).
• ./build.sh — clean; no new warnings in the touched files (the adaptive converter is @unchecked Sendable with the sequential-consumption invariant documented).
• ⚠️ Still needed: a live-mic device-change pass per engine (unplug/switch the input mid-dictation on WhisperKit, Apple Speech, and SpeechAnalyzer streaming) — same manual pass Parakeet's restart got. The wiring can't be driven from fixture audio in swift test because the recovery hangs off a real AVAudioEngine notification.

🤖 Generated with Claude Code