the 0.25.x behavior was wrong
as it returns a 2d ndarray

the 0.25.x behavior was wrong
as it returns a 2d ndarray

that's whats returned by numpy on the underlying arr, see numpy/numpy#13610 (numpy/numpy#15555 (comment))

>>> arr = s.__array__()
>>> arr
array([-2.12345154,  0.50973858,  0.11203232,  0.30739767, -0.78566872])
>>>
>>> np.argwhere(arr < 0)
array([[0],
       [4]], dtype=int64)
>>>

not broken anything so marking as ready for review. do we want an atomic PR to just address the np.argwhere or do we want to look at the other __array_wrap__ on Indexes as well and include more NumPy functions.

that's whats returned by numpy on the underlying arr, see numpy/numpy#13610 (numpy/numpy#15555 (comment))

I don't think those links are relevant.

np.argwhere returns an array with different dimensions see, https://numpy.org/doc/stable/reference/generated/numpy.argwhere.html and so we cannot box the result in a Series. (unless we reshape the data and break consistency with NumPy)

There is a problem (stemming from numpy only passing context for ufuncs) if the result dimensions and shape match the input, boxing can produce confusing results

>>> df = pd.DataFrame([[0, 1], [2, 0]], index=["a", "b"], columns=["foo", "bar"])
>>> df
   foo  bar
a    0    1
b    2    0
>>>
>>> np.argwhere(df > 0)
     a  b
foo  0  1
bar  1  0
>>>

here boxing in a DataFrame with the same labels makes no sense. without the context we don't know which function is being called.

maybe we just raise if context is None instead or always return a Numpy array and push on with implementing __array_function__ where we know the calling function.

@jreback closing this. There are issues with boxing the return of numpy operations eg. #35334 (comment) I think the best course of action is leave the wrapping as is to avoid breaking api changes. Any issues, such as #35331 will need to wait for array_function to be implemented. (even though that issue was a regression)

The fundamental issues are that the semantics of an array and a DataFrame are significantly different (and hence the return of np.argwhere shouldn't be boxed) and numpy only passing context for ufuncs (hence we don't know np.argwhere is the function calling array_wrap)