Generational GC might make heap-allocating your activation records a viable option, even if you're not satisfied with CPython levels of performance. I mean in a sense that's what Chicken does, right?

That is the argument made by Andrew Appel’s paper “garbage collection can be faster than stack allocation” https://www.cs.princeton.edu/~appel/papers/45.pdf and it’s the compilation strategy used by SML/NJ as set out in Appel’s book “compiling with continuations”

Well, I know, but I think he turned out to be wrong in that case. Fast enough, maybe.

It's pretty hard to beat the pushj/popj approach, except that you have to lock out the GC before returning (as it may be scanning the activation records to find roots).

Yeah, whatever Appel thinks, the stack is always faster. But sometimes the heap penalty is affordable. Before generational GC it was ridiculous.