[Retra]

We don't actually need to butt into the Halting Problem or Rices Theorem, though.

Ideally, we would have proven-to-terminate non-turing complete DSLs that can be composed together to form more powerful DSLs. The problem is that we don't really know how to compose programs well, and when you demand the ability to peer into arbitrary executing code and modify it on the fly, you don't make it easier to learn how to do this.

If you were forced to use an architecture that separated things properly, you'd be forced to deal with those composition problems, rather than 'hacking' programs using a level of power that is unnecessary for the problem you hope to solve. (And you are hoping to solve a problem, not just run a program that fails to terminate.)

(I'm not arguing that this is the 'right way' to do things. It's just a way that should be considered properly.)