Problem using assign_attrs() in map()

What is your issue?

In MetPy we have an attribute managing helper that calls DataArray.assign_attrs() within a callable passed to Dataset.map(). With recent versions of xarray, it is impossible to have the attributes assigned using assign_attrs() to be present in the result. See the following example:

import numpy as np
import xarray as xr

ds = xr.DataArray(np.zeros((3)),
    coords={'b': xr.DataArray(np.arange(3), dims='b')},
    dims=['b'], attrs={'units': 'kelvin'}, name='test').to_dataset()

def add_attr(da):
    d = da.assign_attrs(my_attr='foobar')
    return d

ds2 = ds.map(add_attr, keep_attrs=True)

print(ds2['test'].attrs)

This gives {'units': 'kelvin'}. If instead, keep_attrs=False is used, then we get {}.

This boils down to the fact that there are only two options for a DataArray processed within map():

1. Get a direct copy (really, reference) to the original DataArray's .attrs
2. Have .attrs set to {}

There's no option to tell map() to just leave the created DataArray alone. The only work-around I've found is to modify the original DataArray directly instead of using assign_attrs(), which is suboptimal. Things worked fine as recently as xarray 2025.6.1. I know the keep_attrs default changed to True this fall, but I'm surprised setting to False doesn't restore the previous behavior.

Start by running the provided Python reproduction with Dataset.map(), DataArray.assign_attrs(), and both keep_attrs settings. Trace how map handles the returned DataArray attributes and compare the result with the expected preservation of my_attr alongside the original attrs. Done means the example retains the attributes assigned inside the callable without breaking existing keep_attrs behavior.