@lru_cache interferes with @overload'ed definitions

[bukzor]

This is a simplified version of what's seen in stdlib urllib.parse.

Code sample in pyright playground

# pyright:strict
import functools
from typing import AnyStr, NamedTuple, assert_type, overload

class SplitResult(NamedTuple):
    pass

class SplitResultBytes(NamedTuple):
    pass

@overload
def urlsplit(url: str) -> SplitResult: ...
@overload
def urlsplit(url: bytes) -> SplitResultBytes: ...
@functools.lru_cache(typed=True)
def urlsplit(url: AnyStr) -> SplitResult | SplitResultBytes:
    if isinstance(url, str):
        return SplitResult()
    else:
        return SplitResultBytes()

x = urlsplit("ohai")
assert_type(x, SplitResult)

As written, currently, this assert_type fails, but removing the lru_cache removes the error.

[erictraut]

Thanks for the bug report. This will be addressed in the next release.

[bukzor]

I dug into this and found that it seems to be a bug in typeshed, rather -- lru_cache wrapper uses ... as the paramspec which means pyright loses the overload information. pyright and mypy give exactly matching results for the below program.

What did you find?

from typing import Callable, ParamSpec, TypeVar, assert_type, overload

P = ParamSpec("P")
T = TypeVar("T")
def cached_broken(f: Callable[..., T]) -> Callable[..., T]: return f
def cached_worken(f: Callable[P, T]) -> Callable[P, T]: return f

@overload
def g(x: float) -> complex: ...
@overload
def g(x: str) -> None: ...
def g(x: float | str) -> complex | None:
    del x
    return None

cg = cached_broken(g)
assert_type(cg(1.2), complex)
assert_type(cg("1.2"), complex)

cg = cached_worken(g)
assert_type(cg(1.2), complex)
assert_type(cg("1.2"), None)

[erictraut]

This is addressed in pyright 1.1.376.