In support of the requirement for high-fidelity (atom-for-atom) modeling is the notion that an evolved computer would converge toward behaviors that supervene on specifics of the host environment. If porting a binary to another CPU architecture is tough, how easy will it be to port a mind to a simulated simple physics? How many edge cases will it have to get right to even run at all? If brains are hacks designed over millions of generations to surf overlapping fitness functions, it makes sense they'd find implementation (real physics) dependent optimizations that compound in ways which fall apart in toy physics. That's not to say we can't add cool peripherals.

ps even with atom-for-atom modeling, how do you know the behavior doesn't depend on relations which are not computable? If physics ranges over the reals, some of those edge cases might be hard to find with a simulator.