fix(interruption): tolerate websocket teardown race on session.close write

[BhaskarKapri07]

Summary

Fixes an unhandled ClientConnectionResetError: Cannot write to closing transport that escapes the adaptive interruption (barge-in) WebSocket during teardown, surfacing as spurious error reports and silently killing barge-in detection for the affected activity.

Problem

InterruptionWebSocketStream.send_task writes a final session.close frame once the audio input channel closes. If the transport is already closing at that moment — the inference server won the close race, or the local shutdown path called ws.close() first — aiohttp's writer guard raises:

aiohttp/_websocket/writer.py:79 in send_frame
  raise ClientConnectionResetError("Cannot write to closing transport")
aiohttp/client_ws.py:245 in send_str
livekit/agents/inference/interruption.py:819 in send_task

Because ClientConnectionResetError is not an APIError, neither _main_task's retry/emit-error path nor AudioRecognition._interruption_task's except APIError: return swallows it. It escapes as an unhandled asyncio task exception whenever an agent activity is torn down around the interruption stream (embedded agent handoffs, call end) — and the adaptive barge-in capability for that activity silently stops.

Observed in production on LiveKit Cloud: 3 occurrences in ~3 weeks, all at the session.close write during teardown, all with zero impact on the call itself — pure crash-report noise plus lost false-interruption filtering.

Fix

• send_task: skip the session.close write when ws.closed, and swallow the connection reset when the peer wins the close race. The close message is best-effort — if the transport is already gone, there is nothing worth sending.
• _run finally: cancel the send/recv/forward tasks before ws.close(), so the local close can no longer arm the writer guard while send_task is mid-write. This removes the local half of the race; the try/except covers the remaining peer-initiated half.
• Mid-stream audio send failures intentionally still surface as an unrecoverable detector error (new test documents this) — only the final benign teardown write is suppressed.

Tests

• tests/test_interruption/test_interruption_failover.py::TestWsConnectionReset::test_session_close_reset_is_ignored_during_teardown — the teardown close race produces no error event and no task exception.
• tests/test_interruption/test_interruption_failover.py::TestWsConnectionReset::test_audio_send_reset_fails_instead_of_hanging — a mid-stream send reset surfaces as exactly one unrecoverable error.

Validated locally: full tests/test_interruption/ suite (17 tests) green, ruff check and ruff format --check clean, scripts/check_types.py (strict mypy) reports no issues.

[devin-ai-integration]

✅ Devin Review: No Issues Found

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