Let's take a step back here.
I agree (if that's the point) that you cannot implement iterators (or anything else) using direct pointers if you allow reallocation. That's pretty clear, undoubtedly.
My question is specifically what I am paying the overhead of .borrow_mut() for in a Ref<RefCell<>>. (*) If all accesses, mutable or not, go through the Ref (so no direct pointers are kept anywhere), would we agree that there's no potential for use-after-free of the heap allocated object to which the Rc refers?
That then leaves the issue of the compiler assuming that no mutable references exist (and thus being able to perform the same optimizations a C compiler could when a pointer is declared with the restrict keyword). This issue could be side-stepped with a marker trait or something similar like it's already done for some types.
My question sort of is, what else? What are other examples where things would completely break under the alternative scenario that, say, multiple calls to borrow_mut() were allowed to succeed or perhaps equivalently, if Rc<T> allowed mutable access to T (and accesses weren't optimized under the noalias assumption).
(*) I'm especially interested in that since so far no one has been able to present a solution to the alternative question I posed of how to avoid interior mutability that doesn't involve double indirection or hashing or indexing via integers, or avoiding direct references altogether.