sub_relations
we don't track them when canonicalizing or when freshening, resulting in instable caching in the old solver, and issues when instantiating query responses in the new one.