We used to store 'discovered' entities at the global scope, which makes it easy to determine which clusters are accessed by an entity.

However, when considering pruning, we hit a scalability concern: for every entity, we have to verify if there is no flow using it. This search spans all the flows, which is a large set.

The proposal is to insert 'discovered' entities in every cluster they are detected. This duplicates the information, but splits the pruning cost.