The approach that you suggest (including the compiler with the binary) is exactly the "bundlemonkey" approach that jonadab ruled out. And my claim is that if you rule that out, then you run into problems with eval.

I had assumed that jonadab meant that the building-the-interpereter technique wouldn't do the trick because it still wouldn't be a "real" program and you would still suffer from the speed issues that interpreted code would suffer in comparison to optimised native code.

Personally I don't think it would solve the problem in any case. A language bigot of whatever colour will remain a language bigot. Minor things like evidence never seem to change anything.