What fee it calculates depends on what transaction outputs it decides to use to fund the new transaction, and (I could be wrong with this point) I don't believe it does so deterministically. So there certainly isn't a way to do it without all the information in wallet.dat, and that might not even be enough.
This is correct.
I believe the solution is providing an RPC call that creates a preliminary transaction (which isn't broadcast yet), but can be inspected. At this point, it can either be committed (broadcast) or reverted through another RPC call.
Yup, but it needs to lock all transactions involved in the potential spend until it gets approval / rejection / timeout. For almost all systems that I can think of, this is worse than just paying the fee and noting the amount later.
