Merge branch 'main' of https://gitea.xiaowang.eu.org/wx44wx/ai-video-fullstack

The CallEndCoordinator.observe() decremented _tracked_speeches on every
BotStoppedSpeakingFrame, not just the one corresponding to tracked
(end-node) speech.  When a preceding LLM utterance finished between
track_speech() and the actual fixed-speech playback, the counter was
consumed prematurely, ending the call before the end-node message played.

Gate the decrement behind a _current_speech_is_tracked flag that is set
only when BotStartedSpeakingFrame follows a pending tracked-speech start.

Co-Authored-By: Claude <noreply@anthropic.com>

backend/services/pipecat/call_lifecycle.py

@@ -27,6 +27,12 @@ class CallEndCoordinator:
         self._finish_after_tracked_speech = False
         self._finished = False
         self._reason = "completed"
+        # Only BotStoppedSpeakingFrame that follows a tracked BotStartedSpeakingFrame
+        # decrements the tracked-speech counter.  This prevents a stale or cross-talk
+        # stop frame (e.g. from a preceding LLM utterance) from consuming the counter
+        # meant for a fixed end-node message.
+        self._pending_tracked_starts = 0
+        self._current_speech_is_tracked = False
 
     @property
     def ending(self) -> bool:
@@ -58,6 +64,7 @@ class CallEndCoordinator:
         completion = asyncio.get_running_loop().create_future()
         self._tracked_speech_completions.append(completion)
         self._tracked_speeches += 1
+        self._pending_tracked_starts += 1
         return completion
 
     async def arm_after_tracked_speech(self) -> None:
@@ -87,14 +94,19 @@ class CallEndCoordinator:
             self._speaking = True
             self._speech_stopped.clear()
             self._response_speech_started = True
+            if self._pending_tracked_starts > 0:
+                self._pending_tracked_starts -= 1
+                self._current_speech_is_tracked = True
         elif isinstance(frame, BotStoppedSpeakingFrame) and self._speaking:
             self._speaking = False
             self._speech_stopped.set()
-            if self._tracked_speeches > 0:
-                self._tracked_speeches -= 1
-                completion = self._tracked_speech_completions.popleft()
-                if not completion.done():
-                    completion.set_result(None)
+            if self._current_speech_is_tracked:
+                self._current_speech_is_tracked = False
+                if self._tracked_speeches > 0:
+                    self._tracked_speeches -= 1
+                    completion = self._tracked_speech_completions.popleft()
+                    if not completion.done():
+                        completion.set_result(None)
             if (
                 self._finish_after_tracked_speech
                 and self._tracked_speeches == 0