feat(voice V5): turn-taking + barge-in + .record(STT)<->.playback(TTS) session handoff

The integration-risk piece. STT forces the shared AVAudioSession to .record; TTS needs .playback.
Handled by MUTUAL OVERRIDE — each side stops the other and re-asserts its own category before its
operation, so they never fight or deadlock:
 - startVoiceInput() stops TTS first (trinityVoice.stop()) → SpeechTranscriber sets .record+active.
   This is also BARGE-IN: tapping the mic while Trinity speaks cuts the reply and starts listening,
   and the mic never captures Trinity's own voice.
 - speakIfEnabled() stops STT first (if isListening) → TrinityVoice.speak() now re-asserts .playback
   via configureAudioSession() before speaking, flipping the session back from .record.
isSpeaking is published for UI. The shared stop()/CarPlay path is untouched (the re-assert lives in
speak()). Money-isolated: 0 signing calls. App build green. Held. DEVICE-TEST gate applies — session
transitions (mic-after-TTS, TTS-after-mic, barge-in) only fully verify on a real device.

Author: Neo

Core/Trinity/TrinityVoice.swift

@@ -99,6 +99,11 @@ final class TrinityVoice: NSObject, @unchecked Sendable {
     /// Speak the given text using Trinity's voice profile.
     /// - Parameter text: The text to speak.
     func speak(_ text: String) async {
+        // V5 — re-assert the .playback session before speaking. The mic (STT) switches the shared
+        // AVAudioSession to .record; this flips it back to .playback so the reply plays on the right
+        // route. Idempotent for back-to-back replies; the mic side stops TTS before it records, so
+        // the two never fight over the session.
+        configureAudioSession()
         // Stop any current speech
         if synthesizer.isSpeaking {
             synthesizer.stopSpeaking(at: .word)

UI/Views/Agent/AgentConversationViewModel.swift

@@ -54,6 +54,10 @@ final class AgentConversationViewModel: ObservableObject {
 
     func startVoiceInput() {
         voiceError = nil
+        // V5 — barge-in + clean handoff: stop any in-progress speech before recording, so the
+        // session moves .playback → .record and the mic never captures Trinity's own voice.
+        trinityVoice.stop()
+        isSpeaking = false
         transcriber.requestAuthorization { [weak self] granted in
             guard let self else { return }
             guard granted else {
@@ -93,6 +97,7 @@ final class AgentConversationViewModel: ObservableObject {
 
     // MARK: - Voice output (V4 — on-device TTS, Tier 1)
     @Published var voiceOutputEnabled = false
+    @Published var isSpeaking = false
     private let trinityVoice = TrinityVoice()
 
     func toggleVoiceOutput() {
@@ -110,7 +115,14 @@ final class AgentConversationViewModel: ObservableObject {
         guard !spoken.isEmpty else { return }
         let lang = NaturalLanguageProcessor.shared.languageProfile(for: spoken)
         guard lang.tier1Supported else { return }
-        Task { await trinityVoice.speak(spoken, languageCode: lang.code) }
+        // V5 — clean handoff: stop listening before speaking so the session moves .record → .playback
+        // without a clash, and we never transcribe Trinity's own voice back into the input.
+        if isListening { stopVoiceInput() }
+        Task {
+            isSpeaking = true
+            await trinityVoice.speak(spoken, languageCode: lang.code)
+            isSpeaking = false
+        }
     }
 
     /// Strip markdown so TTS doesn't read "asterisk asterisk" etc.