Fix learning for decision that doesn't create decisions

During the rewrite and fix of the learning of the last leaves, an additional constraint was added for simplicity that if a decision wants to do no choices yet wants to propagate, it needs to return a unique choice into which to do the propagation. This additional constraint has not been not well communicated to other developers and it is very fragile. So this branch fix the learning in this case.

This branch also ensure that errors in the schedulers leads to unknown instead of sat back making the former the default.