> Maybe a new comparator is called for (like PY_EQ_FOR_HASH_COLLECTION), which would yield float("nan") equivalent to float("nan") and which should be used in hash-set/dict?

It is difficult to do from technical point of view because all rich comparison implementations support currently only 6 operations. If you call tp_richcomp with something else it will crash. So we need to add new slot tp_richcomp_ex and complex logic to call either tp_richcomp_ex or tp_richcomp and correctly handle inheritance. It is too high bar.

> Does GH-26679 need to be backported to the 3.10 branch?

I thought that the original change was 3.11 only, but seems it was merged before the split of the 3.10 branch. So PR 26679 needs to be backported. I would add a NEWS entry if I knew this.