Loop factory parametrization tears down async fixtures used by sync tests

[tjkuson]

I encountered an issue with defining pytest_asyncio_loop_factories in test suites where non-async tests use async fixtures in shared loop scopes.

To reproduce:

# conftest.py

import asyncio
from collections.abc import Mapping, Callable

import pytest


def pytest_asyncio_loop_factories(
    config: pytest.Config,
    item: pytest.Item,
) -> Mapping[str, Callable[[], asyncio.AbstractEventLoop]]:
    return {"default": asyncio.new_event_loop}

# test_mre.py

from collections.abc import AsyncGenerator

import pytest
import pytest_asyncio


@pytest_asyncio.fixture(scope="session")
async def parent() -> AsyncGenerator[str]:
    yield "parent"


@pytest_asyncio.fixture(scope="session")
async def child(parent: str) -> AsyncGenerator[str]:
    yield "child"


@pytest.mark.asyncio(loop_scope="session")
async def test_async(parent) -> None:
    assert parent == "parent"


def test_sync(child: str) -> None:
    assert child == "child"

Running test_mre.py on the above using pytest-asyncio 1.4.0 and pytest 9.1.1 shows one pass and one failure with the following errror:

AssertionError: The fixture value for "parent" is not available.  This can happen when the fixture has already been torn down.

If pytest_asyncio_loop_factories is not defined in the conftest file, both tests pass.

If test_async is after test_sync, both tests pass.

The expected behaviour is that both tests pass regardless of order and pytest_asyncio_loop_factories hook implementation.

I think the fix would involve handling the loop factory for any test item that uses a pytest-asyncio-managed async fixture, not just async tests themselves.

[hiSandog]

This appears to exercise the teardown boundary between a parametrized loop factory and async fixtures consumed by a synchronous test. A regression case could cover both function- and module-scoped async fixtures, since loop ownership may differ by scope. It would also help to assert that each parametrized loop closes only after its dependent fixture finalizers complete.

[TLNing260310]

I reproduced this on current main and am working on a collection-time fix for the cache-key mismatch identified in the review of #1502. The approach parametrizes a synchronous test only when its static fixture closure contains a pytest-asyncio-managed fixture, using the widest effective fixture loop scope; ordinary synchronous tests remain unparametrized. The regression coverage includes both test orderings, transitive session fixtures and teardown order, two loop factories across multiple scopes, an explicit synchronous managed fixture, and auto mode. I will open a PR after completing local and human review.

[tjkuson]

This issue goes away if factory selection is no longer transported through a pytest fixture, which is the ambition for a v2.