If only given 0d inputs, even if they are of type ndarray, ufuncs will decay their output to a scalar via [()] (as noted in #4563, #5819). While we can't change this behavior now without creating signficant pain downstream, we could add a way to opt out of it.

#13100 raises a case where np.fix resorts to calling np.asanyarray(np.ceil(x, out=out)) in order to ensure that the result is an array. Unfortunately, this has a number of draw-backs:

• It discards duck-types, like dask arrays
• It changes the dtype of 0d object arrays containing array-likes

Proposed implementation:

• Create a new np.leave_wrapped sentinel object that can be passed as an out argument
• Add support in ufunc.__call__, np.take, ... for passing out=np.leave_wrapped meaning "do not call PyArray_Return", causing the result to never be a scalar
• Expose PyArray_Return to python as np.core.unpack_scalar
• Implement np.fix as:

def fix(x, out=None):
    if out is None:
        do_unwrap = True
        out = np.leave_wrapped
    else:
        do_unwrap = False
    res = nx.ceil(x, out=out)
    res = nx.floor(x, out=res, where=nx.greater_equal(x, 0, out=np.leave_wrapped))

    if do_unwrap:
        res = np.unpack_scalar(res)   # PyArray_Return
    return res

def fix(x, out=None):
    res = nx.asanyarray(nx.ceil(x, out=out))
    res = nx.floor(x, out=res, where=nx.greater_equal(x, 0))

    if out is None and type(res) is nx.ndarray:
        res = res[()]
    return res

def fix(x, out=None, *, unpack_scalars=True):
    res = nx.ceil(x, out=out, unpack_scalars=False)
    res = nx.floor(x, out=res, where=nx.greater_equal(x, 0, unpack_scalars=False), unpack_scalars=False)

    if unpack_scalars and out is None:
        res = np.unpack_scalar(res)
    return res