Or oops, there was a methodology error in there.
The first one is 2 minutes and ~4GiB of ram, not 2s. I was accidentally leaving part of it un-evaluated.
Still, a darn sight better than the foldl bit.
Ultimately I ended up with this non-foldl max https://github.com/euank/advent-of-nickel-2022/blob/689b48a0b/lib.ncl#L33-L38 ... that ended up being plenty fast.