Interpolation behaviour inconsistent with numpy?

Hey all,
When running dataset.interp(time=dataset.time) fills with np.nan if one of the neighbor is a np.nan even when interpolation is not actually needed.

Here is the sample code to reproduce the issue :

def test_crop_times_nan() :
        ds = xr.Dataset(
            data_vars = {
                "some_variable" : (['x', 'time'], np.array([[np.nan, 0, 1]]))
            },
            coords = {
                "time" : np.array([0,1,2])
            }
        )
        result = ds.interp(time=ds.time)

        # result["some_variable"].value == [nan, nan, 1.0]
        # whereas [nan, 0, 1.0] is EXPECTED
        xr.testing.assert_allclose(ds, result)

Please note that numpy does not have the same behavior :

>>> import numpy as np
>>> np.interp([0,1,2], xp=[0,1,2], fp=[np.nan,0,1])
array([nan,  0.,  1.])

Is that an intended behaviour for xarray?
If so, does this mean that I first have to check if an interpolation is needed instead of doing it no matter what (and use reindex instead of interp if it is not needed) ?
(this will be kind of tricky if interpolation is needed for certain values and some not...)

Thanks for your help ;)

Environment:

Start by running the provided Dataset.interp(time=dataset.time) reproducer and compare it with np.interp and xr.testing.assert_allclose. Trace the Dataset.interp path for coordinates that already match, then add a regression test showing that existing NaN values are preserved when no interpolation is needed; done means the result is [nan, 0, 1].