fix: resolve audio dropout after AgentTask completion

[tomc98]

Fixed critical bugs causing agent to go silent after nested workflow agents (AgentTasks) complete:

1. Audio recognition not recreated on resume:

   • When activity is paused for AgentTask handover, _close_session() closed audio_recognition but didn't set it to None
   • On resume, _start_session() skipped recreation because the check if self._audio_recognition is None was False
   • Fix: Set audio_recognition, rt_session, and realtime_spans to None after closing them in _close_session()

2. Atomic handover with prestart pattern:

   • Added prestart() to create audio_recognition before activity swap
   • Added complete_start() to finish startup after swap
   • Reordered _update_activity() to prestart new activity before pausing old one, eliminating ~100-200ms audio gap

3. Transcript buffering during handover:

   • Buffer transcripts received during prestart mode
   • Replay them after complete_start() finishes

4. Additional race condition fixes:

   • Fixed generation futures not marked done on interruption
   • Fixed async generator cleanup race conditions
   • Added auth_lock for authorization operations on SpeechHandle

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you all sign our Contributor License Agreement before we can accept your contribution.
1 out of 2 committers have signed the CLA.

:white_check_mark: tomc98
:x: TC

---

TC seems not to be a GitHub user. You need a GitHub account to be able to sign the CLA. If you have already a GitHub account, please add the email address used for this commit to your account.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[Hormold]

Could you please share a minimal PoC agent to reproduce this issue? Thanks!