fix(#3169): pin capture backend at mic start (_activeCaptureMode) — toggle-mid-record safety (Codex review #2)

Codex round-2: _stopMic and mediaRecorder.onstop read the CURRENT _rawAudioMode
to choose backend/dispatch, but the recording was started on the OLD mode — so
toggling Settings→Sound mid-recording could stop the wrong backend (orphaning
the other) or dispatch raw-vs-transcribe wrongly. Pin _activeCaptureMode
(speech | media-raw | media-transcribe) at start; _stopMic + onstop use it.
Adds front-end source-invariant regression tests.

Co-authored-by: lucasrc <lrclucas@gmail.com>

static/boot.js

@@ -444,6 +444,10 @@ $('btnAttach').onclick=e=>{if(e&&e.preventDefault)e.preventDefault();$('fileInpu
 
   // Raw audio mode preference: send audio file instead of transcribing
   let _rawAudioMode = localStorage.getItem('hermes-raw-audio-mode') === 'true';
+  // Capture backend pinned at recording start ('speech' | 'media' | null) so
+  // _stopMic / onstop act on the backend that actually started, even if the
+  // raw-audio toggle changes mid-recording (#3169 Codex review).
+  let _activeCaptureMode = null;
 
   const btn=$('btnMic');
   const status=$('micStatus');
@@ -549,7 +553,11 @@ $('btnAttach').onclick=e=>{if(e&&e.preventDefault)e.preventDefault();$('fileInpu
 
   function _stopMic(){
     if(!window._micActive) return;
-    if(recognition && !_rawAudioMode){
+    // Stop the backend that was ACTIVE WHEN RECORDING STARTED — not whatever
+    // _rawAudioMode says now. The user can toggle Settings → Sound mid-recording,
+    // which would otherwise make us stop the wrong backend and orphan the other
+    // (#3169 Codex review). _activeCaptureMode is pinned at start.
+    if(recognition && _activeCaptureMode==='speech'){
       recognition.stop();
       return;
     }
@@ -630,6 +638,7 @@ $('btnAttach').onclick=e=>{if(e&&e.preventDefault)e.preventDefault();$('fileInpu
     _finalText='';
     _prefix=ta.value;
     if(recognition && !_forceMediaRecorder && !_rawAudioMode){
+      _activeCaptureMode='speech';
       recognition.start();
       _setRecording(true);
       return;
@@ -659,7 +668,7 @@ $('btnAttach').onclick=e=>{if(e&&e.preventDefault)e.preventDefault();$('fileInpu
         _setRecording(false);
         _stopTracks();
         if(blob.size){
-          if(_rawAudioMode){
+          if(_activeCaptureMode==='media-raw'){
             await _sendRawAudio(blob);
           }else{
             await _transcribeBlob(blob);
@@ -669,6 +678,7 @@ $('btnAttach').onclick=e=>{if(e&&e.preventDefault)e.preventDefault();$('fileInpu
           window._micPendingSend=false;
         }
       };
+      _activeCaptureMode=_rawAudioMode?'media-raw':'media-transcribe';
       mediaRecorder.start();
       _setRecording(true);
     }catch(err){

tests/test_raw_audio_upload.py

@@ -141,3 +141,36 @@ def test_raw_audio_upload_different_formats():
         except KeyError:
             pass
         assert handler.status is not None, f"Failed for {filename} ({mime})"
+
+
+# ── Front-end: raw-audio mic backend pinning (boot.js source invariants) ──────
+# Regression guard for the #3169 Codex-review fix: _stopMic / onstop must act on
+# the capture backend that was ACTIVE WHEN RECORDING STARTED (pinned in
+# _activeCaptureMode), not whatever _rawAudioMode says now — otherwise toggling
+# Settings → Sound mid-recording orphans the wrong backend. Also: an explicit
+# Send click (_micPendingSend) must send even with text in the composer.
+import pathlib as _pathlib
+
+_BOOT_JS = (_pathlib.Path(__file__).parent.parent / "static" / "boot.js").read_text(encoding="utf-8")
+
+
+def test_stop_mic_uses_pinned_active_capture_mode_not_current_rawmode():
+    assert "let _activeCaptureMode" in _BOOT_JS, "_activeCaptureMode must be declared"
+    # _stopMic decides backend from the pinned mode, not _rawAudioMode.
+    assert "_activeCaptureMode==='speech'" in _BOOT_JS
+    # onstop dispatches raw-vs-transcribe from the pinned mode too.
+    assert "_activeCaptureMode==='media-raw'" in _BOOT_JS
+    # the mode is pinned at both start branches.
+    assert "_activeCaptureMode='speech'" in _BOOT_JS
+    assert "_activeCaptureMode=_rawAudioMode?'media-raw':'media-transcribe'" in _BOOT_JS
+
+
+def test_send_raw_audio_honors_explicit_pending_send():
+    # An explicit Send-button click (sets _micPendingSend) must send the raw
+    # audio even when the composer already has text.
+    assert "if(window._micPendingSend){" in _BOOT_JS
+    # and it lives inside _sendRawAudio (before the empty-composer fallback).
+    idx = _BOOT_JS.index("async function _sendRawAudio")
+    end = _BOOT_JS.index("function _commitTranscript", idx)
+    body = _BOOT_JS[idx:end]
+    assert "window._micPendingSend" in body and "send()" in body