I should add this question to one of my exams and see if my students get it right. I'll be a little grumpy if they don't, but it's a great question. :)
I suppose you can mathematically construct some kind of non-total message loss such that Paxos would never make progress. But that kind of message loss won't persist forever in a real system... it would basically be message loss with knowledge of the algorithm?
A: Prepare(1)
B: Prepare(2)
A: Accept(1, Va) # fails.
A: Prepare(3)
B: Accept(2, Vb) # fails.
Etc... In practice randomized backoffs when this happens causes it to converge extremely fast, but the exact right situation COULD cause it to delay forever. Don't lose sleep over it though.
It is sort of Byzantine in that the "system" is inserting very specific message losses with knowledge of the algorithm.
Good question. I think the answer is that you don't build large systems using consensus; you bootstrap large systems with very small systems using consensus. The very small systems are reasonably assumed to have no Byzantine faults (i.e. just like you more or less rely on a single database server not to have faults).
All the systems I know of that use consensus are meant to be small, e.g. run on 5 machines or so (and of course the membership is fixed). Google's Chubby, Yahoo's ZooKeeper, and similar systems like doozer and etcd all work like this.
Consensus doesn't "scale" anyway (the latency isn't bearable). If you only have 5 machines, the likelihood of Byzantine faults over a reasonably long time period is low. The main problem you will see is your own software bugs (i.e. not bugs due to faulty CPU, memory, disks, switches, etc.).