The thing is that PR #97316 gives an explanation about which errors may happen but only in regard to modified keys, and not in regard to other errorneous implementations of Eq and Hash (in particular the violation of k1 == k2 -> hash(k1) == hash(k2)).
I think that's what @taperedDominoes pointed out, if I understand right.
However, if Eq and Hash are implemented in malicious ways, then of course HashSet will also be malicious. So I think the docs are good as they are now.