It could also be that we propagate this "resolve by subgraph" logic to its logical extreme and figure out some way to cleverly run ivy very fast and cache the results well