I find it amusing how often trying to do complex tasks with simple constraints, always ends up back at stack based languages. Probably why the stack based turing machine is academically important.
I suppose that prior art influences things. In my case, I went "a separate text editor won't fit, I need a REPL", and that basically filtered to just Forth and Lisp. It would be interesting to see what subset of Lisp would fit in a bootsector, but I wouldn't get my hopes up.
There's sectorlisp that appears to fit into two sectors but is yet shy of their single-sector goal: https://github.com/jart/sectorlisp