I would model this as the position being the integral of the ball's velocity, and switch the velocity, this way, you don't need to do complicated things with trying to grab the position from the last section of the behavior. I tried that too, but the result was the same. I don't see how that would help anyway, since integral is just a stepper over an accumE. I even tried a function-valued behaviour that had to be supplied time explicitly, which did move at least, but still <<loop>>-ed, so it never switched.
Re delayed switching, Reactive already has a kind of delayed switching, that happens automatically when necessary. If you attempt to take the value of a Reactive value at the exact moment of a switch, you will receive the value from the previous step. Oh yes, now I went back to the 'Simply efficient...' paper, and it says so indeed. I just misunderstood that part (off by one thinking, as always ;). However, the problem still remains, because it's not only the new value that depends on the current one but the time of switching as well. This knot is apparently too tight for the time being.
Gergely -- http://www.fastmail.fm - The professional email service