fix(ios): dictation race/lifecycle bugs found in review

ios/ScribaKeyboard/AudioRecorder.swift

@@ -72,7 +72,11 @@ final class AudioRecorder: ObservableObject {
         do {
             try engine.start()
         } catch {
+            // Tear down everything we set up so a retry starts clean (otherwise
+            // the observers leak and the session stays active).
             input.removeTap(onBus: 0)
+            removeSessionObservers()
+            try? session.setActive(false, options: [.notifyOthersOnDeactivation])
             throw RecorderError.engineFailed(error)
         }
         publish(isRecording: true, level: 0)

ios/ScribaKeyboard/DictationController.swift

@@ -29,6 +29,11 @@ final class DictationController: ObservableObject {
     private let impact = UIImpactFeedbackGenerator(style: .medium)
     private let notify = UINotificationFeedbackGenerator()
 
+    // Guards against a double-start: `state` only flips to `.recording` after the
+    // async `recorder.start()`, so a second quick tap would otherwise also start
+    // and install a second audio tap (an uncatchable crash).
+    private var isStarting = false
+
     init() {
         // If the system cuts the recording short (call, another app, AirPods
         // removed), finalize what we captured rather than losing it.
@@ -73,6 +78,11 @@ final class DictationController: ObservableObject {
     }
 
     private func startRecording() async {
+        // Ignore a second start that races the first (state is still .idle until
+        // the await below completes).
+        if isStarting || state == .recording { return }
+        isStarting = true
+        defer { isStarting = false }
         do {
             try await recorder.start()
             live.start() // live preview; no-op if speech permission isn't granted
@@ -87,7 +97,9 @@ final class DictationController: ObservableObject {
     private func finishRecording() async {
         live.stop()
         let audio = recorder.stop()
-        guard !audio.isEmpty else {
+        // A header-only WAV (no captured samples) isn't worth a round-trip — it'd
+        // just come back as "no speech". Treat it as a silent no-op.
+        guard audio.count > WAVEncoder.headerSize else {
             state = .idle
             return
         }

ios/ScribaKeyboard/LiveTranscriber.swift

@@ -29,12 +29,14 @@ final class LiveTranscriber: ObservableObject {
     /// Begins a live recognition session. No-op (server path still works) if
     /// permission isn't granted or a recognizer isn't available.
     func start() {
+        // Clear any leftover preview first, so an unavailable recognizer (early
+        // return below) can't leave a previous session's words on screen.
+        publish("")
         guard SFSpeechRecognizer.authorizationStatus() == .authorized,
             let recognizer = SFSpeechRecognizer(locale: Self.locale()),
             recognizer.isAvailable
         else { return }
 
-        publish("")
         let request = SFSpeechAudioBufferRecognitionRequest()
         request.shouldReportPartialResults = true
         // Keep the live preview on-device (private, no network) when supported.
@@ -63,6 +65,7 @@ final class LiveTranscriber: ObservableObject {
         }
         task?.cancel()
         task = nil
+        publish("")
     }
 
     private func publish(_ text: String) {