I just tried to replace rank32 with a scalar binary search within a node (although my binary search is a bit unusual): it is worse but still ~5x faster than std::set (and still beats Abseil by ~1.5x).

I don't know for sure since I have not compared against them, but I think these particular structures won't be faster even for large arrays. They both perform roughly $$$\log \log 2^{32} = 5$$$ iterations — same as B-trees for all reasonable dataset sizes, — but they are probably slower, at least in their basic implementations:

• The vEB tree uses recursion and branching, which is expensive.
• The x-fast trie uses hash tables, which is even more expensive.

On top of that, they are not storing the data that efficiently, which is bad for the cache performance.

That said, I think that some universe-reducing approaches may be faster when properly optimized — at least for large enough dataset sizes. Implementing an associative array this way is one of the things I'm going to try next.