Per my understanding, what's so exciting about this is precisely that you're wrong - it does not need to be known at compile time, but can prove symbolically that the resulting length will be X + Y. And of course, given that, it's not statically creating a distinct instantiation of the Vect type for every N at compile time!
I'm not sure the exact approach, but there's definitely some heavy lifting. My understanding is that most (all?) dependently typed languages are essentially theorem provers at heart.
