To be honest, I'm really not convinced this is worth it the effort. VLAs work, they have worked for decades, and they're not suddenly going to stop working - and if they do, then we can get rid of them.

I don't really have anything against VLAs, but being able to fall back on heap allocations when the stack is filling up seems good.

VLA dynamic sized small vec (do some sort of stack check to decide VLA or regular heap vector) works, but unconditional VLA doesn't satisfy the above.

Indeed I'm not doing this for the VLAs but for the stack overflows. The PR only had that as a title for continuity with @thufschmitt's work.

VLA dynamic sized small vec

This would seem optimal but leads to somewhat more complicated code because the VLAs can't be put into a type. Also this suggests a small performance penalty with VLAs:

It's also incidentally consistently 1% faster on the benchmarks.

In fact I'm quite ok with VLAs, as long as they're not chocolate flavored.
It's just that we have a big reason (stack overflow) and two minor reasons to avoid them.