Drop the stale-client close message in sync runs

[kelkalot]

Follow-up to #96. This commit was pushed after #96 merged.

After several run() calls in one process, asyncio printed "Task exception was never retrieved … RuntimeError: Event loop is closed". This showed up in the live run that #96 came from. Results were not affected.

Cause. Each sync run() calls asyncio.run(), so each run has its own event loop. OpenAI SDK clients sit in reference cycles and are freed whenever the cyclic GC runs. If that happens while a later run's loop is running, the client's __del__ schedules aclose() on that loop for connections opened on the earlier, closed one, and the task fails. Reusing one auditor across runs is fine: three consecutive run() calls on one auditor worked live.

Change. ModelAuditor.run, AuditExperiment.run, CrossJudgeExperiment.run and reframing's _run_sync (rejudge and friends) now go through simpleaudit._event_loop.run_sync. It works like asyncio.run, but sets an exception handler on the loop that drops exactly one case: an AsyncClient.aclose() task failing with "Event loop is closed". Everything else still goes to asyncio's default handler.

The real fix is to close the clients before their loop ends, but any-llm has no public close method. Asked upstream: mozilla-ai/any-llm#1439.

Not covered: run_async() called from your own event loop, for example in a notebook, uses that loop's handler, so the message can still appear there.

Testing

• pytest on Python 3.11 (-n auto) and 3.13: 1366 passed, 1 skipped.
• tests/test_event_loop.py, 6 tests:
  • The filter drops only the targeted case and passes everything else to the default handler.
  • A regression test reproduces the message without a network, using a real OpenAI SDK client against a local HTTP server, used in one run and collected during the next. The message appears under asyncio.run and not under run_sync.
  • With the filter removed, the two tests that depend on it fail.