I read an interview with a CA bod somewhere that stated that they didnt do collision detections on models in order to determine hits and misses, but its pretty clear that the rates of attacks are determined by the model animations.

I'm guessing the code structure looks something like this:

Code:
loop
    perform calculations to determine attack/defend hit/miss result (effectively instantaneous)
    play appropriate result animation (takes some time)
until (its time to do something else)