This should be a considerable performance improvement, as we
don't need to query swh-graph for every head (which includes
a lot of duplicates).