My initial thoughts was to opt for D because it is easier to maintain (less special cases, less complexity, internally we could get rid of assert_unique_multiindex_level_names after dropping multi-index virtual variables) and also because an error is raised when there's actually an issue, e.g., we could do stack and never plan to do unstack later so in that case multi-index level vs. dimension name conflict is not an issue?

I was thinking about this kind of behavior (even though reverting previous changes is not ideal):

```python ds = xr.Dataset(coords={'dim0': ['a', 'b'], 'dim1': [0, 1]}) ds = ds.stack(dim_stacked=['dim0', 'dim1'])

This works: dim0 is not a dimension in b

ds['c'] = (('dim0',), [10, 11, 12, 13, 14, 15])

raise a nice error message here: conflicting sizes for dimension 'dim0'

ds.unstack(dim_stacked)

raise a nice error message here: conflicting sizes for dimension 'dim0'

ds.sel(dim1=0) ```

However, this may be confusing in the case of integer-based vs. label based indexing:

```python

label-based selection along dim_stacked dimension (length=4)

ds.sel(dim0='a')

integer-based selection along dim0 dimension (length=6)

ds.isel(dim0=0) ```

So a more general rule like option C is not that silly after all?

If a dimension name matches the name of a coordinate:

• Ok if this coordinate is a "dimension" coordinate (coordinate name == dimension name)
• (maybe ok if this this coordinate is not indexed?)
• Error otherwise