fix(voice): skip end-of-turn metrics on stale/out-of-order speaking anchor

[anshulkulhari7]

Summary

Fixes #6093.

For some user turns, the reported transcription_delay / end_of_turn_delay metrics on ChatMessage are extremely large (often >200s) even though the session recordings show no real delay, and stopped_speaking_at can precede started_speaking_at. In other turns the fields are missing entirely.

Root cause

The metrics are computed in _bounce_eou_task (audio_recognition.py) from three captured anchors:

started_speaking_at  = speech_start_time
stopped_speaking_at  = last_speaking_time           # the internal _last_speaking_time anchor
transcription_delay  = max(last_final_transcript_time - last_speaking_time, 0)
end_of_turn_delay    = time.time() - last_speaking_time

The guard around this block only checked that the three values are not None. When the turn detector commits a user turn whose _last_speaking_time was never refreshed for that segment — e.g. consecutive same-role turns split from one continuous utterance, with no VAD speech-stop/start cycle between them — the anchor is left over from an earlier point in the session and can predate the start of the current turn.

In that case the not-None guard still passes, so end_of_turn_delay = now - last_speaking_time becomes ~200s and stopped_speaking_at ends up before started_speaking_at, exactly the payload reported in the issue.

This is the same class of bug noted in #2361 / #5669 / #4388 (stale/0 anchor), now manifesting as an out-of-order anchor on adjacent turns within one long utterance.

Fix

An anchor that predates the start of the turn (last_speaking_time < speech_start_time) is logically impossible — you cannot stop speaking before the turn started. The existing code already has a policy for unreliable timing (see the in-code comment): skip the calculation and report the metrics as None, because that is better than emitting a likely-wrong value. This change extends that same policy to the out-of-order case.

The computation is extracted into a small pure helper, _compute_end_of_turn_metrics, which:

• returns None for all four metrics when any anchor is missing or when last_speaking_time < speech_start_time (stale/out-of-order), and
• otherwise returns the same values as before (with end_of_turn_delay now clamped to >= 0, consistent with the existing transcription_delay clamp).

This makes the behaviour directly unit-testable without audio/STT/VAD.

Testing

New unit test module tests/test_end_of_turn_metrics.py exercises the pure helper with crafted timestamps (no audio):

• test_normal_turn_produces_small_bounded_delays — well-ordered turn yields the expected sub-second delays.
• test_stale_anchor_predating_turn_start_is_skipped — regression for this issue, using the exact ~220s numbers from the reported payload; all four metrics must be None.
• test_anchor_equal_to_start_is_accepted — boundary (last_speaking_time == speech_start_time) stays valid.
• test_missing_anchor_is_skipped — any missing anchor skips the calculation.

$ uv run pytest tests/test_end_of_turn_metrics.py --unit -q
......                                                                   [100%]
6 passed in 0.02s

Confirmed RED before the fix (reverting the ordering guard): test_stale_anchor_predating_turn_start_is_skipped failed with started_speaking_at=1781342804.815377, end_of_turn_delay=220.28458189964294 — i.e. the bogus >200s value. The existing tests/test_speech_start_time_persistence.py still passes.

ruff check, ruff format --check, and mypy are clean on the changed files.

AI disclosure

This change was AI-assisted; all logic, tests, and verification were reviewed by the author.

[CLAassistant]

All committers have signed the CLA.

[devin-ai-integration]

✅ Devin Review: No Issues Found

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

[chenghao-mou]

Thanks for the PR! I have one small comment.

[chenghao-mou]

since _EndOfTurnInfo is internal, I would go one step further to replace the four variables in _EndOfTurnInfo with this directly so we don't have to unpack or pass around those values individually.

[anshulkulhari7]

Done — folded the four metric fields on _EndOfTurnInfo into a single metrics: _EndOfTurnMetrics. The computed object is now passed straight through; _user_turn_completed_task, _init_metrics_from_end_of_turn, and the turn span read info.metrics.*. mypy strict (593 files) and the unit tests pass.