I've been working a bit with Baker's COMFY-6502 code; a few notes of what I have learned so far.
First, a couple of tiny bugs in the genbrc; the code miscounts the size of the branch instructions, meaning that (- l 2) should be a (+ l 2) and so on. I found it handy to enumerate each clause of the cond. Each clause handles a particular case, such as when the "lose" continuation can be reached by a short branch instruction, while the "win" continuation is far enough away to require an absolute jump.
Second, genbrc, genbr, and compile all provide the address of the resulting "continuation" as the return value. This is perhaps clear when one traces out all the recursion, but it isn't explicitly mentioned. One interesting case is genbrc when the "win" and "lose" branch destinations are the same: one could simply emit an unconditional branch to that destination, and return the address of that branch but actually returning the destination works just as well. (In the emit routine, if the continuation does not happen to be the next instruction, the unconditional branch is, after all, emitted, moving the continuation to the front of the instruction stream, ready for the non-branching instruction to be emitted just in front of it.)
As for "upgrades" to the package, I have been focused up to this point on moving the knowledge of 6502 opcodes and addressing modes from magic decimal numbers to symbolic processing. Instead of simply emitting a decimal opcode, I have been changing the code to emit symbolic opcodes, such as (ADC ABSOLUTE), with routines to reduce these symbolic forms to the appropriate opcode, including checking for invalid opcodes. Baker's original code will happily emit opcodes with addressing modes not supported by the chip. Currently, the compiler emits these symbolic codes into the code vector, awaiting a processing step to convert them to the decimal equivalent. Ideally, one would detect invalid addressing modes in the compilation stage, not the post-processing stage.
Some other changes I am contemplating are allowing for symbolic jump destinations, so that object code can be relocated and external labels could be used to relax the restrictions of the current scheme, which requires manually sequencing compilation and storing of addresses.