MAINT, TST: failure in test_axis_nan_policy_decorated_positional_args with array API

[tylerjereddy]

I regularly run the full array API suite locally, but can be reproduced with focused call as well: python dev.py test -t scipy/stats/tests/test_axis_nan_policy.py::test_axis_nan_policy_decorated_positional_args -b all

___________________________________________________________________________________________________________ test_axis_nan_policy_decorated_positional_args ___________________________________________________________________________________________________________
scipy/stats/tests/test_axis_nan_policy.py:585: in test_axis_nan_policy_decorated_positional_args
    stats.kruskal(samples=x)
        message    = "kruskal() got an unexpected keyword argument 'samples'"
        rng        = Generator(PCG64) at 0x7FF32ED32180
        shape      = (3, 8, 9, 10)
        x          = array([[[[           nan, 2.69786714e-01, 4.09735239e-02, ...,
          7.29496561e-01, 5.43624991e-01, 9.35072424e-0...    [6.60372970e-01, 9.65276732e-02, 1.06143803e-01, ...,
          3.86779163e-01, 9.80181895e-01, 1.39465441e-01]]]])
scipy/stats/_axis_nan_policy.py:494: in axis_nan_policy_wrapper
    samples = _broadcast_arrays(samples, axis=axis)
        _no_deco   = False
        axis       = 0
        d_args     = {}
        default_axis = 0
        hypotest_fun_in = <function kruskal at 0x7ff3381de020>
        hypotest_fun_out = <function kruskal at 0x7ff3381de020>
        intersection = set()
        is_too_small = <function _axis_nan_policy_factory.<locals>.is_too_small at 0x7ff3381ddee0>
        keepdims   = False
        kwd_samp   = []
        kwd_samples = []
        kwds       = {'samples': array([[[[           nan, 2.69786714e-01, 4.09735239e-02, ...,
          7.29496561e-01, 5.43624991e-01, 9...   [6.60372970e-01, 9.65276732e-02, 1.06143803e-01, ...,
          3.86779163e-01, 9.80181895e-01, 1.39465441e-01]]]])}
        maxarg     = inf
        n_kwd_samp = 0
        n_out      = 2
        n_outputs  = 2
        n_samp     = 0
        n_samples  = None
        nan_policy = 'propagate'
        override   = {'nan_propagation': True, 'vectorization': False}
        paired     = False
        params     = ['nan_policy']
        reduced_axes = 0
        result_to_tuple = <function _axis_nan_policy_factory.<locals>.result_to_tuple at 0x7ff3381dde40>
        samples    = []
        sentinel   = None
        temp       = []
        tuple_to_result = <class 'scipy.stats._stats_py.KruskalResult'>
        used_kwd_samples = []
        vectorized = False
scipy/stats/_axis_nan_policy.py:19: in _broadcast_arrays
    xp = array_namespace(*arrays) if xp is None else xp
        arrays     = []
        axis       = 0
        xp         = None
scipy/_lib/_array_api.py:128: in array_namespace
    return array_api_compat.array_namespace(*_arrays)
        _arrays    = []
        arrays     = ()
scipy/_lib/array_api_compat/common/_helpers.py:283: in array_namespace
    raise TypeError("Unrecognized array input")
E   TypeError: Unrecognized array input
        _use_compat = True
        api_version = None
        namespaces = set()
        xs         = ()

During handling of the above exception, another exception occurred:
scipy/stats/tests/test_axis_nan_policy.py:584: in test_axis_nan_policy_decorated_positional_args
    with pytest.raises(TypeError, match=re.escape(message)):
E   AssertionError: Regex pattern did not match.
E    Regex: "kruskal\\(\\)\\ got\\ an\\ unexpected\\ keyword\\ argument\\ 'samples'"
E    Input: 'Unrecognized array input'
        message    = "kruskal() got an unexpected keyword argument 'samples'"
        rng        = Generator(PCG64) at 0x7FF32ED32180
        shape      = (3, 8, 9, 10)
        x          = array([[[[           nan, 2.69786714e-01, 4.09735239e-02, ...,
          7.29496561e-01, 5.43624991e-01, 9.35072424e-0...    [6.60372970e-01, 9.65276732e-02, 1.06143803e-01, ...,
          3.86779163e-01, 9.80181895e-01, 1.39465441e-01]]]])

@mdhaber maybe this slipped through? Or am I doing something wrong.

[mdhaber]

I have been seeing that, and I thought we fixed it already in a PR or have a PR open with a fix. IIRC it started when we changed _get_nan. It's an edge case nicety - it's checking for a particular error message but it's getting another. I'll take another look.

[mdhaber]

Ths comes up because a function that accepts a variable number of positional arguments gets zero positional arguments. Now that we're adding array_namespace very early in the execution of a lot of functions, functions that handled the no-arg case gracefully in the past now start to get the error message that array_namespace wants to raise. This has come up in at least two different ways so far, so I'll make sure a test like this runs in the array API job.