fix: prevent _cancel_speech_pause from poisoning subsequent user turns

[giulio-leone]

Summary

Fixes a race condition where _cancel_speech_pause can permanently break user-turn processing, causing the assistant to silently stop replying.

Problem

When a user interrupts paused speech that has no active generation (e.g., the generation was never authorized or was cancelled before starting), _wait_for_generation() raises:

RuntimeError: cannot use wait_for_generation: no active generation is running.

This error bubbles up from _cancel_speech_pause and poisons the _cancel_speech_pause_task. Since subsequent tasks chain on the prior task via await old_task, the RuntimeError re-raises in every subsequent call, permanently breaking user-turn completion.

Impact: User speech is transcribed (STT events arrive) but the turn is never committed to chat context and no assistant reply is generated — for the current turn and all future turns.

Root Cause

In _cancel_speech_pause:

1. await old_task — re-raises any error from a previously failed task
2. await self._paused_speech._wait_for_generation() — raises RuntimeError if self._generations is empty

A paused speech handle with an empty _generations list triggers (2), which then triggers (1) on every subsequent call.

Fix

Two changes to _cancel_speech_pause:

1. Guard _wait_for_generation(): Only call it when self._paused_speech._generations is non-empty. This prevents the RuntimeError from being raised in the first place.

2. Catch errors from await old_task: Wrap it in try/except so a single failed task doesn't permanently poison the task chain. The error is logged at debug level and subsequent turns can proceed normally.

Testing

The fix is defensive — it prevents an uncaught exception from a known race condition. The scenario requires specific timing between STT events, pause/resume, and generation authorization that is difficult to unit-test deterministically. The fix follows the same defensive pattern used in _mark_done() (which uses contextlib.suppress for similar edge cases).

Closes #5100

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.

---

giulio-leone 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.}

[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 2 additional findings.