fix(voice): force-interrupt outdated reply in _user_turn_completed_task

[mpg-kaoru-hotate]

Fixes #4443

Root cause

When a new user turn starts while a previous _user_turn_completed_task is still running, the older task detects it is outdated and discards its reply:

if self._user_turn_completed_atask != asyncio.current_task():
    # If a new user turn has already started, interrupt this one since it's now outdated
    await speech_handle.interrupt()

speech_handle was just created by _generate_reply() and inherits the session's allow_interruptions. When interruptions are disabled (turn_handling={"interruption": {"enabled": False}}, or Agent(allow_interruptions=False)), interrupt() without force=True raises:

RuntimeError: This generation handle does not allow interruptions

Consequences beyond the noisy error log:

• the remainder of the task is aborted, so EOU metrics for that turn are never emitted
• the outdated reply handle is not cancelled and stays scheduled, so the agent can play a reply to a partial/stale transcript alongside the reply of the newer turn

This is internal cleanup of a reply that should never play — not a user-initiated interruption — so it should bypass the allow_interruptions guard with force=True. All other internal interrupt() call sites in agent_activity.py already either pass force=True or check allow_interruptions first; this was the only unguarded one.

Reproduction

We hit this in production (telephony agent with interruptions disabled to suppress line-echo false positives) whenever the user produced consecutive turns in quick succession — e.g. speaking again right after end-of-turn was detected, or STT splitting one utterance into multiple finals. Traceback from Sentry matches the one reported in #4443:

  File ".../livekit/agents/voice/agent_activity.py", line 1983, in _user_turn_completed_task
    await old_task
  ...
  File ".../livekit/agents/voice/agent_activity.py", line 2127, in _user_turn_completed_task
    await speech_handle.interrupt()
  File ".../livekit/agents/voice/speech_handle.py", line 151, in interrupt
    raise RuntimeError("This generation handle does not allow interruptions")

The added regression test (test_consecutive_user_turns_when_interruptions_disabled) reproduces exactly this traceback on main without the fix, and passes with it. It mirrors the existing test_interrupt_during_on_user_turn_completed scenario (second user turn arriving while on_user_turn_completed of the first turn is still running) with interruption.enabled=False.

Testing

• pytest tests/test_agent_session.py — 55 passed (including the 2 new parametrized cases)
• ruff check / ruff format clean

🤖 Generated with Claude Code

[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.
_{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 1 additional finding.

[longcw]

thanks for the pr! could you sign t he CLA before we merge it