TCH002 false positive with singledispatch

[flying-sheep]

Playground link

functools.singledispatch’s register decorator can take a function with the first argument being type-annotated. That type annotation will then be used to dispatch.

The following code should therefore be correct with all rules active, however it claims rule violations on the annotated lines.

"""Test module."""
from __future__ import annotations

from functools import singledispatch
from typing import TYPE_CHECKING

from numpy import asarray  # TCH002: Move third-party import `numpy.typing.ArrayLike` into a type-checking block
from numpy.typing import ArrayLike  # TCH002: Move third-party import `scipy.sparse.spmatrix` into a type-checking block
from scipy.sparse import spmatrix

if TYPE_CHECKING:
    from numpy import ndarray


@singledispatch
def to_array_or_mat(a: ArrayLike | spmatrix) -> ndarray | spmatrix:
    """Convert arg to array or leaves it as sparse matrix."""
    msg = f"Unhandled type {type(a)}"
    raise NotImplementedError(msg)

@to_array_or_mat.register
def _(a: ArrayLike) -> ndarray:
    return asarray(a)

@to_array_or_mat.register
def _(a: spmatrix) -> spmatrix:
    return a

[flying-sheep]

@ivirshup support for pydantic-like and attrs-like patterns can already be achieved via flake8-type-checking.runtime-evaluated-base-classes and flake8-type-checking.runtime-evaluated-decorators, respectively