An active learning approach would be cool, where the formatter process a large corpus of code and asks a human operator for advice when it is especially uncertain about the best place to split.

Even if training the splitting policy from scratch is too hard, learning would be a good way to establish cost weights on different kinds of rule violations.