…1705)

This fixes memory unsafety when there is a race condition related to
attribute get/set operations on a free-threaded build, for simple
reference-based attributes only (`PyObject *`). This includes attributes
with primitive types like `list`, `set` and `str` and `dict`, and native
class types.

The main idea is to use delayed decref for the old attribute value when
assigning to an attribute. There are detailed comments explaining the
approach in detail.

This causes a ~5% slowdown in self check when using 3.14t. Currently it
looks like a significant perf cost is unavoidable if we want to fix
memory unsafety, but we can further optimize mypy to reduce the impact
by introducing final attributes, for example.

Remaining work includes fixing attributes with fixed-length tuple types
and tagged integer types (in follow-up PRs).

I used coding agent assist heavily.

Work on mypyc/mypyc#1203.

…21726)

Mypyc now gives additional thread safety guarantees.

Release tracking issue: #21717

Fixes #21824
Fixes #19920
Closes #19921

Note that ilevkivskyi has a suggestion to change overload inference
here:
#19920 (comment)

While that is right, it is a little separate, and I think it is okay to
remove the crash given that it now affects numpy and has been reported
in other contexts too. We just keep the originally inferred tuple

The `env_class` object associated with a mypyc coroutine is not
immediately cleared when the coroutine completes. This can significantly
increase memory usage, since the env class may hold references to
captured locals and values spilled across suspension points.

The objects are eventually collectible by the GC but the collection
might be delayed in cases where a nested coroutine is awaited, eg.

```python
async def allocate(size: int) -> None:
    payload = bytearray(size)

    async def nested() -> int:
        return payload[-1]

    assert await nested() == 0
```

With nesting mypyc creates env classes for both `allocate` and `nested`
that may reference each other and form a cycle.

To fix this, clear the `env_class` immediately after the coroutine
completes. This matches behavior of cpython, which clears frames of
completed coroutines immediately in the eval loop.

Closes mypyc/mypyc#1204, a `mypyc` issue.

The issue likely unblocks `isort` from being compiled with `mypyc`,
which is why I took an interest it. Full disclosure is that I used AI to
write this, it provided this test and code just from the linked issue
but it looks sensible to me. I did not use additional prompting. Like I
said, fix looks sensible: we now call `dataclass_type` for each entry in
the MRO, instead of once.

I am aware the contributing guidelines say new contributors are not
encourage to use LLMs but I hope the size of the PR and my track record
as open source maintainer gives some flexibility here. I'm very aware
that reviewing AI written PRs creates maintainer burden, but I'm
committed to getting this over the finish line. To that end I:

Tested locally with `pytest mypyc/test/test_run.py ` and that looked
okay.
Also tested that without the changes the added test would fail on
`master`.
Also tried to run CI on my local fork
(DanielNoord#1), which succeeded.

Feel free to push changes to the branch or cherry pick this into another
branch. I am not necessarily interested in getting credits for the
fix/PR, I'd just like to continue with my attempt to get `isort`
compiled and for that I need this in a `mypy` release :)

Closes mypyc/mypyc#1210

This fixes another issue I ran into while trying to use `mypyc` for
`isort`.

I am aware the contributing guidelines say new contributors are not
encouraged to use LLMs but I hope this can get the same handling as
#21785. I have tried to ensure this
patch works as much as I can by:

Testing locally with `pytest mypyc/test/test_run.py `.
Tested that the test fails on `master` without the changes.
Also tried to run CI on my local fork
(DanielNoord#2), which succeeded.

As with the previous PR, feel free to push changes to the branch or
cherry pick this into another branch. I just want to unblock `isort` :)