The JVM byte code situation isn’t a great example because that was a series of deliberate design choices for lots of complex reasons. And, the JVM absolutely can guarantee memory safety at the bytecode level. It’s just working with a slightly more dynamic type system than Java source.

What would happen if you tried to do PCC for InvisiCaps and FUGC is that it would ultimately constrain what optimizations are possible, because the optimizer would only be able to pick from the set of tricks that it could express a proof for within whatever proof system we picked

Totally not the end of the world.

Do I think this an interesting thing to actually do? I’m not sure. It’s certainly not the most interesting thing to do with Fil-C right now

Yes, I agree with you that a correct JVM can guarantee memory safety at the bytecode level, but what I meant to express was that many JVMs have had bugs that caused them to fail to do so, for essentially social reasons which I expect to cause problems with other PCC systems as well.

Maybe you're right, and those problems are not inevitable; for example, if you could reduce the proofs to a tiny MetaMath-like kernel that wouldn't need constant "maintenance". As you say, that could move the compiler's optimizer out of the TCB — at least for the security properties Fil-C is enforcing, though the optimizer could still cause code to compute the wrong answers.

That seems like something people would want if they were already living in a world where the state of computer security was much better than it is now.