Fix dictation startup races and migrate tap defaults

What changed:
- treat an in-flight dictation start as stoppable so a rapid second tap queues a stop instead of being ignored
- preserve the requested stop mode through the before-start stop path and clear pending mode state on cancel/cleanup
- migrate dictation toggle and auto-send preferences to v2 storage keys so devices with legacy values pick up the new tap-to-toggle defaults
- extend manual verification steps with rapid-tap and touch-viewport coverage

Why:
- PR #48 received review feedback about startup race conditions in useDictation
- phones with previously saved dictation settings could stay on hold-to-talk even after switching the product default to tap-to-toggle

Impact:
- desktop and mobile dictation now follow the same tap-to-toggle flow by default
- rapid double taps during recorder startup stop and transcribe instead of leaving recording running

Validation:
- pnpm run build
- Playwright verification on 375x812 and 768x1024, including legacy-pref mobile load and rapid second-tap startup handling

Author: SHAREN

src/App.vue

@@ -1003,8 +1003,8 @@ const accountActionError = ref('')
 const SEND_WITH_ENTER_KEY = 'codex-web-local.send-with-enter.v1'
 const IN_PROGRESS_SEND_MODE_KEY = 'codex-web-local.in-progress-send-mode.v1'
 const DARK_MODE_KEY = 'codex-web-local.dark-mode.v1'
-const DICTATION_CLICK_TO_TOGGLE_KEY = 'codex-web-local.dictation-click-to-toggle.v1'
-const DICTATION_AUTO_SEND_KEY = 'codex-web-local.dictation-auto-send.v1'
+const DICTATION_CLICK_TO_TOGGLE_KEY = 'codex-web-local.dictation-click-to-toggle.v2'
+const DICTATION_AUTO_SEND_KEY = 'codex-web-local.dictation-auto-send.v2'
 const DICTATION_LANGUAGE_KEY = 'codex-web-local.dictation-language.v1'
 
 const GITHUB_TRENDING_PROJECTS_KEY = 'codex-web-local.github-trending-projects.v1'

src/composables/useDictation.ts

@@ -151,6 +151,7 @@ export function useDictation(options: {
     if (state.value !== 'idle' || !isSupported.value || isStartingRecording) return
     isStartingRecording = true
     stopRequestedBeforeStart = false
+    pendingStopMode = 'insert'
 
     try {
       mediaStream = await navigator.mediaDevices.getUserMedia({ audio: { channelCount: 1 } })
@@ -171,10 +172,13 @@ export function useDictation(options: {
       mediaRecorder.start(250)
       state.value = 'recording'
       if (stopRequestedBeforeStart) {
-        stopRecording()
+        const stopMode = pendingStopMode
+        stopRequestedBeforeStart = false
+        stopRecording(stopMode)
       }
     } catch (error) {
       cleanup()
+      stopRequestedBeforeStart = false
       state.value = 'idle'
       options.onError?.(error)
     } finally {
@@ -203,6 +207,7 @@ export function useDictation(options: {
 
   function cancel() {
     stopRequestedBeforeStart = false
+    pendingStopMode = 'insert'
     cancelTranscription()
     cleanup()
     state.value = 'idle'
@@ -274,6 +279,7 @@ export function useDictation(options: {
   function cleanup() {
     stopWaveformCapture()
     resetWaveformDisplay()
+    pendingStopMode = 'insert'
     if (mediaRecorder) {
       mediaRecorder.ondataavailable = null
       mediaRecorder.onstop = null
@@ -291,7 +297,7 @@ export function useDictation(options: {
   })
 
   function toggleRecording() {
-    if (state.value === 'recording') {
+    if (state.value === 'recording' || (isStartingRecording && state.value === 'idle')) {
       stopRecording()
       return
     }

tests.md

@@ -2242,17 +2242,21 @@ Toggle "Free mode" in settings to use free OpenRouter models without an OpenAI A
 
 #### Steps
 1. Open Settings and confirm `Click to toggle dictation` is enabled by default and `Auto send dictation` is disabled by default.
-2. In the composer, click the microphone once and verify recording starts without holding the button.
+2. On desktop, click the microphone once and verify recording starts without holding the button.
 3. Speak a short phrase, then click the microphone again to stop recording.
-4. While transcription is processing, confirm the microphone button shows a loading spinner.
-5. Confirm the transcript is inserted into the composer without being auto-sent.
-6. Start dictation again, speak another short phrase, and click the send arrow instead of the microphone stop button.
+4. Start dictation again and immediately click the microphone a second time before the recorder fully starts.
+5. While transcription is processing, confirm the microphone button shows a loading spinner.
+6. Confirm the transcript is inserted into the composer without being auto-sent.
+7. Start dictation again, speak another short phrase, and click the send arrow instead of the microphone stop button.
+8. Repeat steps 1-7 in a touch viewport or on a phone and verify a tap starts recording and the next tap stops it without requiring a press-and-hold.
 
 #### Expected Results
 - Dictation starts on first click and keeps recording hands-free until the next click.
+- A rapid second tap during startup still queues a stop request instead of leaving recording running.
 - Stopping with the microphone inserts the transcript into the composer and leaves it editable.
 - The microphone shows a transcribing spinner while `/codex-api/transcribe` is still in progress.
 - Clicking the send arrow during recording transcribes the audio and submits the message automatically.
+- Devices with previously saved dictation preferences pick up the new tap-to-toggle default after the preference-key migration.
 
 #### Rollback/Cleanup
 - Clear the draft or delete the test thread message if needed.