Cancellation leaks out of asyncio.TaskGroup on 3.12 when using eager tasks

[graingert]

Bug report

Bug description:

import sys
import asyncio

if sys.platform == "win32":
    EventLoop = asyncio.ProactorEventLoop
else:
    EventLoop = asyncio.SelectorEventLoop


async def demo():
    class MyError(Exception):
        pass

    async def throw_error():
        raise MyError

    try:
        async with asyncio.TaskGroup() as tg:
            tg.create_task(throw_error())
    except* MyError:
        pass


def loop_factory():
    loop = EventLoop()
    loop.set_task_factory(asyncio.eager_task_factory)
    return loop

asyncio.run(demo(), loop_factory=loop_factory)

results in:

Traceback (most recent call last):
  File "/home/graingert/projects/anyio/demo.py", line 29, in <module>
    asyncio.run(demo(), loop_factory=loop_factory)
  File "/usr/lib/python3.12/asyncio/runners.py", line 194, in run
    return runner.run(main)
           ^^^^^^^^^^^^^^^^
  File "/usr/lib/python3.12/asyncio/runners.py", line 118, in run
    return self._loop.run_until_complete(task)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/usr/lib/python3.12/asyncio/base_events.py", line 687, in run_until_complete
    return future.result()
           ^^^^^^^^^^^^^^^
asyncio.exceptions.CancelledError

CPython versions tested on:

3.12

Operating systems tested on:

Linux

Linked PRs

• gh-128588: gh-128550: remove eager tasks optimization that missed and introduced incorrect cancellations #129063
• [3.13] gh-128588: gh-128550: remove eager tasks optimization that missed and introduced incorrect cancellations (GH-129063) #129089
• [3.12] gh-128588: gh-128550: gh-128552: fix refcycles in eager task creation (#128553) and remove eager tasks optimization that missed and introduced incorrect cancellations (python#129063) #128586

[graingert]

I think related to #116720 but a lot easier to trigger by accident.

[kumaraditya303]

This seems fixed on main. Perhaps we missed backporting a fix for it?

[graingert]

I believe it was intentionally not backported because the conditions to trigger it were thought to be a bit complex and it needed a change to how Task.cancel worked (re unsetting must_cancel)

[kumaraditya303]

cc @gvanrossum