fix(voice): deadline-based user_away_timeout ignores transcript-less noise (#6030)

[dorukdumlu]

Summary

• Track an absolute _user_away_deadline for user_away_timeout so transcript-less user speaking ↔ listening flips re-arm with remaining time (telephony noise can no longer defer "away" indefinitely) — user_away_timeout getting triggered for SIP participant on telephony calls even when user is not speaking and taking background noise and conversations as user state speaking. #6030.
• Add user_away_on: "presence" (default, backcompat) vs "conversation" (ignore user VAD speaking; only pause while the agent is active / refresh on final transcript). Prefer conversation for noisy SIP + turn_detection="stt".
• Drop redundant away-recovery _set_user_away_timer(reset=True) that could arm the timer while the agent is speaking (Devin review).

Test plan

• test_noise_user_state_flips_do_not_defer_away_deadline
• test_final_transcript_and_agent_activity_refresh_away_deadline
• test_away_recovery_transcript_does_not_arm_during_agent_speech
• test_conversation_user_away_ignores_user_speaking
• test_conversation_user_away_still_cancels_on_agent_speaking
• Existing test_final_transcript_resets_away_timer_when_not_speaking

[longcw]

I think we can add an option to only cancel the user away timer when agent speaking or user input transcribed. right now it's cancelled when agent or user speaking that may too sensitive to the noises.

[dorukdumlu]

@longcw thanks — agreed the cancel policy was the missing piece (deadline alone only fixes flip-storms, not continuous noise stuck in speaking).

Pushed:

1. user_away_on — "presence" (default) keeps current cancel-on-user-speaking + deadline preservation; "conversation" cancels only while the agent is active and ignores user VAD speaking. Telephony/user_away_timeout getting triggered for SIP participant on telephony calls even when user is not speaking and taking background noise and conversations as user state speaking. #6030 should use user_away_on="conversation".
2. Devin bug — removed the redundant away-recovery reset=True that could arm the timer mid-agent-speech.

Happy to rename the option or flip the default if you'd rather.

[dorukdumlu]

Follow-up for the Devin finding on stale remaining=0 after long genuine speech in presence mode:

• Track _user_away_speech_had_transcript across the current speech segment (finals count even while VAD still says speaking).
• On speaking→listening: reset=True if a final was seen (full window); reset=False if transcript-less (noise / user_away_timeout getting triggered for SIP participant on telephony calls even when user is not speaking and taking background noise and conversations as user state speaking. #6030).
• Added test_presence_long_speech_with_transcript_refreshes_away_deadline.

conversation mode unchanged.

[dorukdumlu]

Follow-up for Devin's conversation-mode stall:

When the agent returns to listening while the user is stuck in VAD speaking (noise), _update_agent_state now re-arms the away timer in user_away_on=""conversation"" — presence mode unchanged (still requires user listening).

Added test_conversation_rearms_away_after_agent_while_user_speaking + presence regression guard.

[dorukdumlu]

Rebased onto current main (resolved conflicts with transcription_timeout / session options).

user_away_on + deadline-based tracking still in place on top of #6478's final-transcript refresh. Away unit tests green locally.

@longcw ready for another look whenever you have a minute.

[dorukdumlu]

Addressed Devin's conversation-mode finding:

Final non-empty transcripts now refresh the away deadline mid-utterance when user_away_on=""conversation"" (agent idle). Prevents a long answer from flipping to away before EOS while VAD still says speaking.

Added test_conversation_final_transcript_refreshes_away_during_speech. Away suite: 10 passed.

[dorukdumlu]

@longcw rebased onto current main and squashed the follow-ups.

Your cancel-policy request is user_away_on: presence (default) still pauses on user/agent speaking; conversation ignores user VAD speaking and only pauses while the agent is active. Telephony / #6030 should use user_away_on=conversation.

Deadline is preserved across transcript-less noise flips. Final transcripts refresh the window (including mid-utterance in conversation mode). Ready for another look.

[longcw]

closing in favor of #6880