Real life savestates are problematic because of our brains. Should the savestate include the current brain state of the user?
If the answer is yes, then people would just repeat the same action over and over again after loading a savestate. That's because they could have no memories of the failed attempts.
What if the answer is no, so that we could load a savestate without losing our "future memories"? The problem is, it would cause the loaded savestate to be inaccurate copy of a world state. This might not necessarily lead to desync, but it would certainly be cheating. Playing the movie back would look strange from the
electroencephalography point of view. Someone dodging raindrops means she's got brains that change their state spontaneously, faster than physically possible. Also, there would exist no causal history that lead to the obscure dance preformance, because it is the failed attempts in the alternative future scenarios that caused the succesfull dodging.
I'm not saying it's impossible. It's just that the tool-user can't be part of the world that is recorded. You'd have to let someone else control you. Also, some form of
determinism might be required to avoid desyncing movies.
By the way, I believe the difficulty of dodging raindrops would equal the difficulty of completing SMB in 4:59. (But that depends on how hard it is raining and how fat you are.)
