And really? Just any random sequence gets you a new fold. I mean, it won't be very useful if you pick a random one, but it'll work and be a new one.

I think this is just an artifact of natural selection basing new proteins on existing ones, not an actual useful ("rational" if you can call natural selection rational) selection limit. I don't think that if you designed proteins from first principles you'd see this limitation in your results.