This is obviously an improvement, but I think we should optimize the whole algorithm as well. _unicode_lookups appears to be a very simple translation of the UCD case folding data, and we could probably get better performance (and better code size) by preprocessing that data to something more compact.

A starting point would be figuring out the offsets for each character, then merging ranges with the same offsets and storing just the range start and length. Obviously this is more "work" to look up but it's a lot less thrashing. If you start the search by looking at the first range then you end up doing basically the same thing as this optimization except you might not get the bitwise version (which probably won't matter, arithmetic has the same TP as bitwise ops and there aren't any dependency chains on the output).

You'd probably want to first check at the beginning of your map (in case of ascii) then start searching from wherever the last processed character was.