Knuth-Plass is fundamentally a dynamic programming algorithm that depends on being able to say "If in the future I determine that there is a line break at point X, I know for certain that the way I should have broken the paragraph up until X is B", where B depends only on information before X.

If you're trying to avoid rivers, then based on information after point X you might find that your sequence of breaks B wasn't actually the best after all, because it ended up interacting badly with later spaces and producing rivers. So you can no longer prove that you've achieved a global optimum of the penalty function.

Of course you may be willing to take this hit and settle for non-optimum line-breaking (and you still might do better than an algorithm that doesn't take rivers into account at all).