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>
Ubuntu