How hard could it be to write the thunking layer?
Keep in mind the current mindset is that everything has to be portable across all of the (hypothetical) backends that Rakudo (hypothetically) supports, so writing a thin embedding layer in C won't do. You have to write it in NQP and then hope that the backend allows you to invoke C functions in an efficient way. That's C for Parrot, JNI for the JVM, something like P/Invoke for the CLR, wishful thinking for v8, and magic unicorn fairy dust for MoarVM.
It's gotta be easier than recreating P5 + XS + everything else.
If P6 is optimized for -Ofun, then plumbing and maintenance and packaging and all of the other pieces of finish work that aren't glamorous must not be fun.