Every ggseg_atlas keys its core on three columns, and they are not
interchangeable.
The two conventions for region
A parcellation that names its structures anatomically gives both
hemispheres the same name, so region aggregates them and hemi
disambiguates: grouping by region gives one row per structure.
A parcellation that numbers its parcels does not. region_0001 and
region_0002 are two different parcels, not one structure seen twice,
and no hemisphere-free name exists to recover – the source never gave
one. There region is 1:1 with label and grouping by it aggregates
nothing.
Both are correct for their source. What was missing was a way to ask
which you have, short of testing whether any name is shared across
hemispheres. atlas_region_semantics() is that question.
