hm, i'm not sure i understand why we wouldn't hit the cache in practice. it was necessary to make the performance of merging cargo snapshots together at all reasonable, so it's definitely an improvement that definitely blocks the rust backend. the cache is doing the expected thing here. as mentioned in the description, there was also a separate commit needed to add

merkle_<http://trie.rs|trie.rs>

on top of the caching, in order to make merging those snapshots reasonable at all. both the caching and the speedup are required.