Looked like a race condition if two blocks are reading from the same buffer; one block calls prune before the other one can, and then it no longer gets its tags. This puts back in the original protections we used when pruning with vectors, only we don't have to restart the iterator at the beginning and can exit sooner because of the sort order. Likely slower than the multimap erase, so we will have to reinvestigate what kind of performance hit we take and if we can do better.