In general: looks good.

Ad 1:
Since setfault is in locore.S and you're not merging that code, I would defer this question until later. That said: I don't know whether either of the Book-E or AIM functions is optimized and precisely save the minimal set of registers or whether they're still relatively quick-n-dirty so so that it works.
All they need to do is signal to the trap handler, when it handles a page fault, that user space is accessed and that control should return to the function doing the access with enough state saved that this can be done without violating the ABI. Thus the question mostly pivots around what registers need to be preserved across calls (and consequently what registers can be considered scratch) and what the user-space accessor function saves extra to deal with the trap.
Put differently: they are what the comment says; very similar to setjmp. See what setjmp saves and why and you should have a good starting point. Note that setfault is rather special-purpose, so it's possible it can save less.
Most/some architectures do not need to save anything. They just need to set a flag and an instruction address to return to from interrupt. You may want to ask yourself why PowerPC uses a setjmp approach that requires saving registers in all cases instead of a lazy let-it-trap approach that only requires saving registers when a fault happens... There's room for a noticeable performance improvement by letting the standard trap code save and restore registers as needed...

Ad 2:
Since Book-E uses arbitrary codes, it makes sense to match these codes to the vector addresses of AIM. Exceptions specific to Book-E can use codes that is not likely to ever be used: something larger than EXC_LAST for AIM.