Why is it that 35 years after the Futamura projections were first described, when we want to produce a new compiler, we don't write an interpreter and pass it through a high quality, general partial evaluator?

Whenever the Futamura projections are described, it's always at a conceptual level, in general terms. Reading a blog post like this one, I'm always left to wonder (a) why does this not end with a link to a github implementation of a general partial evaluator and (b) why is it the case that none of the compilers I use were produced with these techniques?

There must be some set of tradeoffs preventing this approach from living up to the "compilers for free" promise, but I haven't found anyone with a real explanation. Either there must be hard challenges which make implementing a general partial evaluator very difficult, or the results that we'd get from using such a partial evaluator are poorer than what we get through other means right?

Perhaps the greatest success in this direction is PyPy. At some level the RPython toolchain is general enough that people have used it to produce JITted versions of ruby, php and a few other languages. But these are a handful of alternative implementations to existing languages, where the predominant implementation doesn't use this toolset. And PyPy's own docs note that the JIT compiler needs some amount of hinting from the interpreter; you can't just throw any interpreter (written in RPython) at it.