fix/preserve-stt-eos-timestamp-for-metrics

[dhruvladia-sarvam]

Problem

When turn_handling.turn_detection = "stt" is used with an STT plugin that emits SpeechEventType.END_OF_SPEECH (e.g. Sarvam, AssemblyAI) and no external VAD, EOUMetrics.transcription_delay and EOUMetrics.end_of_utterance_delay are always reported as 0.0.

Root cause

AudioRecognition._on_stt_event unconditionally overwrites self._last_speaking_time with time.time() whenever a FINAL_TRANSCRIPT (or PREFLIGHT_TRANSCRIPT) arrives and self._vad is None — even though the STT plugin already populated that field with the actual end-of-speech moment via END_OF_SPEECH. The framework's own TODO comment acknowledged this:

the correct way is to ensure STT fires SpeechEventType.END_OF_SPEECH and using that timestamp for _last_speaking_time

Fix

Skip the overwrite when turn_detection_mode == "stt", since in that mode the STT plugin is the authoritative source of speech boundaries.

Behavior matrix

External VAD	turn_detection	Before	After
Yes	any	unaffected	same
No	"vad" / "manual"	falls back to transcript time	same
No	"stt"	overwrites STT timestamp → metrics = 0	preserves STT timestamp

[devin-ai-integration]

✅ Devin Review: No Issues Found

Devin Review analyzed this PR and found no potential bugs to report.

View in Devin Review to see 3 additional findings.