unavoidable spurious "exception in shielded future" logged if exception raised after shield cancelled

Bug report

Bug description:

test case:

import asyncio

async def fail(t):
    await asyncio.sleep(1)
    raise ValueError("should be caught, not logged")

async def main():
    fut = asyncio.create_task(fail(1))
    try:
        await asyncio.wait_for(asyncio.shield(fut), timeout=0.5)
    except asyncio.TimeoutError:
        print("timeout, as expected")
    
    try:
        await fut
    except Exception as e:
        print("error caught and handled:", e)
    

if __name__ == "__main__":
    asyncio.run(main())

#134331 appears to have introduced a regression where _log_on_exception is added unconditionally and permanently on a future that has been shielded at some point, when the shield future is cancelled, rather than removing the handler when the shield future is done.

In the above code, a long-running future is created. shield It is awaited with a timeout using wait_for(shield(fut)). Since the future is not done, this completes and should leave us in a state just like we had an asyncio.sleep(0.5). However, instead when the future resolves, the error is logged, even though the exception is caught and handled:

> python3 test.py
timeout, as expected
ValueError exception in shielded future
future: <Task finished name='Task-2' coro=<fail() done, defined at /tmp/test.py:3> exception=ValueError('should be caught, not logged')>
Traceback (most recent call last):
  File "/tmp/test.py", line 5, in fail
    raise ValueError("should be caught, not logged")
ValueError: should be caught, not logged
error caught and handled: should be caught, not logged

as far as I can tell, the behavior is more correct if _log_on_exception is removed entirely, but I don't fully track #134331, so it may be solving a problem I don't see.

Maybe I'm missing something? Is there some assumption that the inner future should always be finished before the outer one unless something is wrong? That's what the logging behavior suggests to me.

This is a regression in 3.14. 3.13 and prior appear to have correct logging behavior for failure after shielding.

CPython versions tested on:

3.14

Operating systems tested on:

macOS

Linked PRs

• gh-156521

Start in Lib/asyncio/tasks.py around the _log_on_exception handling referenced in the issue, then run the provided shield, wait_for, and failing-task example on Python 3.14. Compare the behavior with Python 3.13 and prior versions. Done means an exception raised after the shield is cancelled is not logged when the original future is subsequently awaited and handled.