Good point. It is misleading if different flakes end up producing the same derivation, and we don't want to resign our build trace entry to account for that (which would amplify reads into writes). Separate indirection for this eval->store layer accounting sounds good.