The key point is that after 2 hours, all of the generals can independently assess, by examining the previously mined blocks in the chain they are working on, how much CPU was spent working on the solution, and can "see" how many nodes are in the network, and hence can see if all the nodes have worked on this solution (if yes, they all know of the arranged time of attack).
This doesn't fully solve the problem (one general could be rouge, or one might be killed just before the attack, ...) but it at least raises the chances :)
That's the thing about BG; you can keep increasing the chances, but if your requirement is that you must be certain that all the other generals will attack, and at the same time, then we know of no solution. We also have more than one good proof that this is impossible.
I think what happens here is that any node that doesn't get the message due to lost message, does not advance to the next block, and hence is working on a shorter chain than the nodes that got the message. Eventually it will receive another message with a longer chain and be forced to abandon its own work and adopt the new longer chain. It's highly improbable that nodes working on shorter chains could ever catch up to those working on the longest version.
There may be some threshold of lost messages where the system begins to break down, but assuming no intelligent attacker is behind it and the losses are random, they affect the whole system, not just the prime chain. The shorter chains would be roughly equally affected.
>How do the other generals know that he hasn't got the final block that is long enough?
He keeps working on whatever he has until he receives a longer blockchain from neighbors, which forces him to abandon his shorter one and start working on the longer one right away.
>Or is it just the case that a majority of generals know when to attack? I thought it had to be all of them, but maybe that's the two-general problem and Byzantine-general's is an easier problem.
Eventually it is all of them. The big difference between bitcoin and the theoretical BGP or TGP is that the theoretical problems have an end state where it is simply good enough for a majority to decide on an attack time, they attack, win, share the spoils, and the process is over. Simple majority is all that was needed.
Bitcoin takes that a step further and never stops. As soon as it reaches a simple majority decision, all nodes abandon their own work and adapt that decision - the longest blockchain - as soon as they receive it, and the process starts over from that new point. Each solution creates a longer blockchain, which forces unaninimity.
Say that after some time a block that is long enough is made by some general A and is distributed to generals B and C. General D doesn't get it though - his messenger is killed in transit. He helped make an earlier block, but has never seen the fully completed block.
How do the other generals know that he hasn't got the final block that is long enough?
Or is it just the case that a majority of generals know when to attack? I thought it had to be all of them, but maybe that's the two-general problem and Byzantine-general's is an easier problem. I was pretty sure there were multiple good proofs of the impossibility of a solution.