You're right it seems - as written, the parsing overhead from the requirement to make the entire system run the GPU makes it impractical. A hybrid system could outperform CPU or GPU alone.
What I really want is a scheme->GLSL shader compiler - much narrower and more tractable.
In case you haven't seen the applications of nanopass compilers to do similar things, specifically harlan, here you go: https://github.com/eholk/harlan
What would it give you to compile scheme to glsl? Glsl is already a pretty direct yet high level representation of shaders. Why put another layer on top of it?
What I really want is a scheme->GLSL shader compiler - much narrower and more tractable.