I don't believe these named tuples are actually available in the main namespace (as __module__ now claims), no? This could potentially cause issue I imagine with anything relying on __module__ + __name__ for retrieving an object, e.g. does pickling even work now?

Good point - @rgommers Do you think that namedtuple classes of return objects for setops should be present in the main namespace? Now when I think about it they shouldn't and I would leave __module__ as numpy.lib._arraysetops.

We had similar discussion for index tricks, e.g. np.r_ which is <numpy.lib._index_tricks_impl.RClass at 0x102fd6b20>. RClass return class is in a private module only.

I don't think those should be public. But yes, pickling should work, and adding a test to confirm that would be useful. So __module__ should just not be set, then it should be automatically correct.

The return these here should be named tuples as well, just as is the case during runtime. It's a bit annoying, but that does mean the named tuples definitions will have to be repeated here in the .pyi file:

# NOTE: still needs a `from typing import Generic` statement

class UniqueAllResult(NamedTuple, Generic[_SCT]):
    values: NDArray[_SCT]
    indices: NDArray[intp]
    inverse_indices: NDArray[intp]
    counts: NDArray[intp]

class UniqueCountsResult(NamedTuple, Generic[_SCT]):
    values: NDArray[_SCT]
    counts: NDArray[intp]

class UniqueInverseResult(NamedTuple, Generic[_SCT]):
    values: NDArray[_SCT]
    inverse_indices: NDArray[intp]

Sure! I added these classes to typing.