Because the network follows a longest chain is valid rule? If it doesn't then you are relying on a node knowing that an alternate chain came "later" and not all nodes will no that. As I already pointed out up thread imagine you are a new node, you connect to the network and receive two competing chains A & B. A is longer. Which chain do you use? If you use A and other nodes use B that is a problem (isolation attack and network fork due to non deterministic chain selection). If they are choosing B over A because they "saw it first" there is no way for you to confirm that or even know that.
Still it doesn't need to be 10,000 blocks. A 51% attack can be accomplished with a reorg of any length.
So, you are talking about two different things:
1) new nodes
2) existing nodes
Both handle things differently.
If I have a rig (or more accurately a massive mining farm that is a majority of the hashing power) I have incurred a cost and I am taking a risk by executing an attack. The difference with PoS is that an attack can be executed without cost or risk. That would only be true for PoW if I could build a farm, then sell it, and then somehow execute an attack after the sale with the farm I don't have. It was tongue in cheek to show that since PoW can't be exploited by history, I can't perform an attack with no cost of risk
No, you do not. Because you are anonymous and you can do it over and over again because nobody can punish you for publishing these blocks.
The network can't punish the bad guy.
It can because the consensus power is known to the network entirely. Existing nodes know who is going to generate the next block and therefore will not accept any derivations or re-orgs especially not those coming from ages ago.
The whole point is that PoS, the bad guy can attack without cost or risk. How exactly does a PoS punish an anonymous entity who no longer has anything at risk and can attack you with no cost.
No cost/risk = depends on how the network punishes him. Distributing his coins, removing his coins, whatever. Because the consensus power lies within the network, the network can decide what to do if a bad guy tries to bring it down.
If that were true that checkpoints wouldn't be needed. There is no PoW "problem". There is a limitation that both PoS and PoW share and that is the security model only works if the attacker has less than half of the resource. An attacker can buy computing power and an attacker can buy a stake neither are closed systems.
It is
THE PoW problem. The network
CANNOT simply punish bad guys. How could it? The consensus power lies outside of the control of the network.