The crux is this assumption by Paxos:
Liveness(C;L): If value C has been proposed, then eventually learner L will learn some value (if sufficient processors remain non-faulty).
The bad case for the two generals problem are when messengers are continuously intercepted. The "if sufficient processors remain non-faulty" part eliminates that possibility.
In other words, if messages are continuously dropped then Paxos is not required to (and won't) finish.