This is nice writeup. Thanks. Another commenter said will've taken them 2h just to sketch out ideas; sans LLMs will've taken me more than 2h just to collect all this info let alone start optimizing it.
I think I'd be able to make some progress optimizing this program in two hours but probably not much. I'm not a performance engineer but have designed exotic emulated CPU architectures before, so that helps a lot.
I gleaned about half of this comment in a few minutes of just skimming the code and reading the comments on the functions and classes. There's only 500 lines of code really (the rest is the benchmark framework).
On the whole I don't think I'd perform all that well on this task given a short time limit but it seems to me to be an extremely well designed task given the stated context. The reference kernel easily fits on a single screen and even the intrinsic version almost does. I think this task would do a good job filtering the people they don't want working for them (and it seems quite likely that I'm borderline or maybe worse by their metric).