Related to #6020 but IMO should be considered separately. Mostly a mirroring of LLVM-175086.

In flat_map and flat_set, sometimes we call unique with a synthesized equivalence relation to deduplicate elements after sorting. And it's probably redundant to call the comparator twice when deduplicating the sorted range.

When operator== is reliable, we can just use it for unique (addressed in #6024).
When it's not reliable, I think it will be better to

• introduce a modified version of unique which doesn't require equivalence relation, and
• pass something similar to key_compare to that algorithm to deduplicate elements.

Moreover, it turns out that key-equiv in the standard wording is possibly pointless. It's unclear to me whether the current wording requires particular side effects from synthesized equivalence comparisons. If so, we should also file an LWG issue to get rid of such requirements.