If you push a hash table which handles collisions by linear probing that hard, and any non-randomness appears, you can get long runs and are in trouble. I knew about Robin Hood hashing, but didn't know it was being used to justify 90% table density. Usually, above 50% full and it's time to double the size of the table.

What does the Rust implementation do for deletions? The classic solution is to mark the deleted cell as available for reuse, but non-empty for the linear part of the search during lookup. If you do that, at some point you have to compact the table or recopy it or something. This is an argument for a hash table with linear lists for each hash slot. More memory allocations and cache misses, but better behavior as the table fills up.

(Remember the YC discussion in which some job interview question insisted that hash tables operated in constant (O(1)) time? Not quite. There's collision overhead and growth overhead.)