Don't feel too much like debugging anymore. It doesn't make sense to assume a logical error in the compiler anymore (the combinators compiled look entirely standard). So, I must have missed something in the code. Recompiling again with an extra check.
There seems a lot of room for optimization of the code, given the few examples I worked through. I eta-expand a lot of expressions which can be dumbed down. Should look if I still run the partial evaluator, might have switched it off, should see how often and where it is run. Cool, I hope I am gonna take about 50% off of code size with some optimizations.
(Looks like I am serializing a piece of text where it expects an AST. It's the permissiveness of stage 1. So, either fix that once or just accept it and debug stuff by hand. Found the bugs, need to print, again.)
(I removed some bugs, but the AST is very probably garbled. I.e., it finds texts where it expects AST nodes and cannot serialize those. I never notice, since it occurs only at positions -like a type node- where I expect texts anyway. Not sure.)