🧱 static integration testing - #35

So numpy isn't compatible with our HasArrayNamespace protocol, because numpy uses string literals, instead of str in the input position (which is contravariant).

What string literals does it expect? Limited to the list of compatible versions?

What string literals does it expect? Limited to the list of compatible versions?

In numpy 2.3.1 it [looks like this]:(https://github.com/numpy/numpy/blob/4d833e5df760c382f24ee3eb643dc20c3da4a5a1/numpy/__init__.pyi#L1684)

_ArrayAPIVersion: TypeAlias = L["2021.12", "2022.12", "2023.12", "2024.12"]

# --snip--

def __array_namespace__(self, /, *, api_version: _ArrayAPIVersion | None = None) -> ModuleType: ...

in 2.0.2 it looks like this:

def __array_namespace__(self, *, api_version: str | None = ...) -> Any: ...

and <2.0.0 there was no __array_namespace__

Huh, why is mypy using the stubs from numpy 2.3.1 when numpy 1.25.0 is installed?
https://github.com/data-apis/array-api-typing/actions/runs/16028175155/job/45221469410?pr=35

So mypy apparently completely ignores build isolation...

The workaround is to use uv pip install, instead of uv run --with, which layers the dependencies. It's disgustingly ugly, but at least the tests are failing for the right reasons now...

Ok, now the question remains: What to do with numpy<2? Ignore it (for now), script our way around it by using array-api-compat if numpy<2, or split it up into two separate integration tests?

My feeling is that we are going to end up using array-api-compat anyway, no? I guess not for older versions of the standard, but as the standard is developed.

My feeling is that we are going to end up using array-api-compat anyway, no? I guess not for older versions of the standard, but as the standard is developed.

for the purposes of integration testing it might be a good idea to rely on it as little as we can

My feeling is that we are going to end up using array-api-compat anyway, no? I guess not for older versions of the standard, but as the standard is developed.

Testing both is also an option 🤔

may as well ignore it for now if we can pass without it, and see what happens down the line

Nevermind, I managed to get this working on numpy<2 by using np.array_api there. It required a little bit of black voodoo magic in the gha workflow, and it's not ideal in terms of IDE support, but it's the cleanest way I could think of.

@lucascolley did you want to give this another look?

Could we open an issue :)

#40