The best hypothesis on the scaling explanation is that at the point where your tweet is injected into the message bus, they have your full friend-of-followers network available. With that information, they can reduce the number of followers that they send a message to. Otherwise, they blast it to everyone, and then figure out on the receiving end if it's an @reply that that person wants to see.

What doesn't make sense about this is that if they were to store the @reply preference for your followers along with the friend-of-followers graph, they could maintain the original behavior, and avoid unneeded traffic on the message bus.