This is cool and clever. But if it doesn't sort in O(1) memory, it's not quicksort. With the lambda calculus lacking a concept of place, it's always going to be tough ensure sorting in place...or randomized pivots as in Hoare's algorithm.
You don't have to do this if you're already using a hybrid sort to prevent Quicksort's worst case, like introsort or (shameless self promotion) pdqsort (https://github.com/orlp/pdqsort).
I haven't run the math, but I don't believe introsort fixes this. It's a constant-factor optimization.