fix(voice): drop playback-phase barge transcripts that echo Hermes' own TTS

The full-duplex barge-in listener added in 5081551f0 stays active during
TTS playback with no acoustic echo cancellation. On some speaker/mic
combinations, TTS bleed alone crosses the barge threshold, gets
transcribed, and is queued as the next user turn -- whose reply is then
spoken, captured, and queued again, producing an unbounded TTS -> STT ->
TTS feedback loop (#75780).

Add a fail-closed transcript-level guard: when a barge trip happens during
the playback phase, compare the captured transcript against the TTS text
Hermes just spoke (tools/voice_mode.is_tts_echo, a language-agnostic
character-level similarity ratio). A close match is dropped instead of
queued, and the mic is handed back to the normal continuous-listening
loop. Generation-phase trips (no TTS playing, so no bleed is possible)
are unaffected.

cli.py

@@ -4723,6 +4723,8 @@ class HermesCLI(CLIAgentSetupMixin, CLICommandsMixin, CLIBillingMixin):
         self._voice_tts_done.set()
         self._voice_tts_stop = None  # active streaming pipeline's stop event
         self._voice_barge_capture = threading.Event()  # barge monitor is capturing the interruption
+        self._voice_last_tts_text = ""  # most recently spoken TTS text (echo guard, #75780)
+        self._voice_barge_phase = None  # "generation" or "playback" phase of the last barge trip
 
         # Status bar visibility (toggled via /statusbar)
         self._status_bar_visible = True
@@ -12566,6 +12568,7 @@ class HermesCLI(CLIAgentSetupMixin, CLICommandsMixin, CLIBillingMixin):
                 tts_text = tts_text.strip()
             if not tts_text:
                 return
+            self._voice_last_tts_text = tts_text
 
             # Use MP3 output for CLI playback (afplay doesn't handle OGG well).
             # The TTS tool may auto-convert MP3->OGG, but the original MP3 remains.
@@ -12676,6 +12679,7 @@ class HermesCLI(CLIAgentSetupMixin, CLICommandsMixin, CLIBillingMixin):
                 # Latch BEFORE cutting anything: suppresses process_loop's
                 # auto-restart until the capture is submitted.
                 self._voice_barge_capture.set()
+                self._voice_barge_phase = phase
                 if phase == "playback":
                     logger.debug(
                         "TTS CUT: full-duplex listener tripped during playback"
@@ -12733,6 +12737,20 @@ class HermesCLI(CLIAgentSetupMixin, CLICommandsMixin, CLIBillingMixin):
                     _cprint(f"\n{_DIM}Stop phrase detected — ending voice chat.{_RST}")
                     self._disable_voice_mode()
                     return
+                # Fail-closed echo guard (#75780): a playback-phase capture
+                # has no acoustic echo cancellation, so speaker bleed alone
+                # can trip the barge trigger. If the transcript is a close
+                # match for what Hermes just spoke, treat it as self-capture
+                # instead of queuing it as a user turn.
+                if getattr(self, "_voice_barge_phase", None) == "playback":
+                    from tools.voice_mode import is_tts_echo
+                    if is_tts_echo(transcript, getattr(self, "_voice_last_tts_text", "")):
+                        logger.debug(
+                            "Dropping playback-phase barge transcript as TTS echo: %r",
+                            transcript,
+                        )
+                        _cprint(f"\n{_DIM}Ignored likely TTS echo (not queued).{_RST}")
+                        return
                 self._pending_input.put(_VoiceInputMessage(transcript))
                 submitted = True
             elif not result.get("success"):
@@ -13892,6 +13910,7 @@ class HermesCLI(CLIAgentSetupMixin, CLICommandsMixin, CLIBillingMixin):
             # playback (speech cuts TTS), and disarms itself when the turn
             # is fully done. See _voice_full_duplex_listener.
             if self._voice_mode and self._voice_continuous:
+                self._voice_last_tts_text = ""
                 threading.Thread(
                     target=self._voice_full_duplex_listener, daemon=True
                 ).start()
@@ -13963,6 +13982,10 @@ class HermesCLI(CLIAgentSetupMixin, CLICommandsMixin, CLIBillingMixin):
                 def stream_callback(delta: str):
                     if text_queue is not None:
                         text_queue.put(delta)
+                    # Track what's actually being spoken so a playback-phase
+                    # barge capture can be checked against it (echo guard,
+                    # #75780).
+                    self._voice_last_tts_text = (self._voice_last_tts_text or "") + delta
 
             # When voice mode is active, prepend a brief instruction so the
             # model responds concisely. The prefix is API-call-local only —
@@ -15230,6 +15253,8 @@ class HermesCLI(CLIAgentSetupMixin, CLICommandsMixin, CLIBillingMixin):
         self._voice_tts_done.set()  # Initially "done" (no TTS pending)
         self._voice_tts_stop = None  # active streaming pipeline's stop event
         self._voice_barge_capture = threading.Event()  # barge monitor is capturing the interruption
+        self._voice_last_tts_text = ""  # most recently spoken TTS text (echo guard, #75780)
+        self._voice_barge_phase = None  # "generation" or "playback" phase of the last barge trip
 
         if os.environ.get("HERMES_DEFER_AGENT_STARTUP") != "1":
             self._install_tool_callbacks()