It seems to me (granted I haven't gotten it to work reliably at all yet) that the REPL is somewhat fragile, and I wonder if it might be easier as well as smaller to use a remote (tethered) compiler, as some Forth systems do.

Re the repl, I may have figured out what went wrong with user-defined functions. Instead of saying (define (sq x) (* x x)) I apparently have to say (define sq (lambda (x) (* x x))). Maybe that is documented in the paper, which I still have to read.