I wonder how hard it would be, given the detection algorithm, to fix this automatically. Seems like simulated annealing would be a pretty good fit, where you perturb the word spacing on each iteration and the energy function is the total river length. The fitness landscape is rich enough with good solutions that you should rarely need more than a few iterations to arrive at one.

Still, that's a pretty expensive fitness test, and I wonder if there's a more elegant and efficient approach than Monte Carlo methods.