ParamSpec semantics are not correct for generic/overloaded functions

[warsaw]

Describe the Bug

Here's the example code:

import shutil
from contextlib import ExitStack


def foo(tmpdir):
    with ExitStack() as resources:
        resources.callback(shutil.rmtree, tmpdir, ignore_errors=True)


def bar(tmpdir):
    shutil.rmtree(tmpdir, ignore_errors=True)

and the output of pyrefly check example.py:

ERROR Missing argument `onerror` in function `contextlib._BaseExitStack.callback` [missing-argument]
 --> example.py:7:27
  |
7 |         resources.callback(shutil.rmtree, tmpdir, ignore_errors=True)
  |                           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  |
 INFO 1 error

What's odd to me is that pyrefly complains about onerror (which clearly is a default argument), but only when called through ExitStack.callback. When called explicitly, pyrefly (correctly, IMHO) accepts the call with the missing onerror argument.

% pyrefly --version
pyrefly 0.48.0

Sandbox Link

No response

(Only applicable for extension issues) IDE Information

No response

[grievejia]

Minimized repro:

from typing import Protocol, overload, Callable

class Foo(Protocol):
    @overload
    def __call__(
        self,
        ignore_errors: bool,
        onerror: int | None
    ) -> None: ...  # overload_0
    @overload
    def __call__(
        self,
        ignore_errors: bool = False,
    ) -> None: ...  # overload_1

def callback[**P, T](callback: Callable[P, T], /, *args: P.args, **kwds: P.kwargs) -> Callable[P, T]: ...

def test(rmtree: Foo) -> None:
    callback(rmtree, ignore_errors=True) # Expected no error but got: Missing argument `onerror`

The issue seems to be overload selection on paramspec forwarding. IIUC, when a better match overload_1 is presented, we might want to be able to prefer that over the non-successful match overload_0. Mypy, pyright, and ty all seem to be pretty consistently doing that.

---

AI root cause diagnosis: When checking if an overloaded callable is a subset of Callable[P, T]:

1. Pyrefly tries the first overload (overload_0 requiring onerror)
2. It succeeds in matching, and binds P to overload_0's parameters
3. Later, when checking *args: P.args, **kwargs: P.kwargs against the actual arguments (ignore_errors=True), it fails because P was bound to the first overload which requires onerror

The fundamental problem is that ParamSpec inference happens before the actual call arguments are known. For regular functions, this works fine. But for overloaded callables, the choice of which overload to use should depend on the subsequent arguments.