Merge fails when sparse Dataset has overlapping dimension values

Sparse numpy arrays used in a merge operation seem to fail under certain coordinate settings. for example, this works perfectly:

import xarray as xr
import numpy as np

data_array1 = xr.DataArray(data,name='default',
                           dims=['source','receiver','time'],
                          coords={'source':['X.1'],
                                  'receiver':['X.2'],
                                  'time':time}).to_dataset()
data_array2 = xr.DataArray(data,name='default',
                           dims=['source','receiver','time'],
                          coords={'source':['X.2'],
                                  'receiver':['X.1'],
                                  'time':time}).to_dataset()

dataset1 = xr.merge([data_array1,data_array2])

But this raises an IndexError: Only indices with at most one iterable index are supported. from the sparse package:

import xarray as xr
import numpy as np
import sparse

data = sparse.COO.from_numpy(np.random.uniform(-1,1,(1,1,100)))
time = np.linspace(0,1,num=100)

data_array1 = xr.DataArray(data,name='default',
                           dims=['source','receiver','time'],
                          coords={'source':['X.1'],
                                  'receiver':['X.2'],
                                  'time':time}).to_dataset()
data_array2 = xr.DataArray(data,name='default',
                           dims=['source','receiver','time'],
                          coords={'source':['X.2'],
                                  'receiver':['X.1'],
                                  'time':time}).to_dataset()

dataset1 = xr.merge([data_array1,data_array2])

I have noticed this occurs when the merger would seem to add dimensions filled with nan values.

Start by running the sparse.COO reproducer in the issue and compare it with the dense NumPy case. Trace the xarray merge path where overlapping coordinates introduce dimensions filled with NaN values, then verify that merging the sparse datasets completes without the sparse indexing error and preserves the expected dimensions.