I actually mentioned this in the article! That's a nice catch on the L2 cache reasoning and modulo idea, that saves on spenny heap allocations everywhere. I di

I actually mentioned this in the article! That's a nice catch on the L2 cache reasoning and modulo idea, that saves on spenny heap allocations everywhere. I did the first sweep of optimisations to kill the super-slow regex and turn the slowest operations into a set matching operation. I considered pre-processing all the potential interesting codes, reducing everything to a simple dict/set matching, but by that point the actual operation to generate OTPs was orders of magnitude slower than everything else, so there would be negligible user-facing benefit to doing so.