We always heard that to be the case, but as long as the double spend only had a few victims, I am not so certain.
If it were fiat , it would be akin to someone using counterfeit money to buy good or services.
In real life , No one reimburses the person that sold his car for counterfeit money ,
they just try and arrest the guy who did the counterfeiting and only give back the car , if they catch the counterfeiter, and track down the car.
Well, if I have spent lots of time and money to attack a PoS currency, then I would try to scam all existing exchanges for the maximum amount. That would cause heavy disruptions and in most cases, delistings.
Replacing the doublespend amount with a hard fork implies a centralized authority making that decision,
a truly decentralized resource such as gold , no one makes the pretense that stolen gold will be replaced unless the person that stole it is apprehended.
It is funny, we want crypto to be decentralized, but we also want centralized protections.
A hard fork is not centralized per se, it can be the voluntary decision of the majority of the economic actors invested in the currency, in the case of a PoS coin, to run another client. While in past hard forks (ETH/ETC as the prime example) a kind of "leadership" led to the hard fork, that's not mandatory at all. It's enough if all merchants and exchanges simply use the new client.
Anyway, that would be a point against PoS, not for PoS.
There is only 1 way to Censor a Proof of Stake coin Transactions for an extended period,
but it is not 51% attack , it is 100% control of the full nodes, (Which is almost impossible)
I'm sure that with a very high supermajority (95% e.g.) it would be possible to censor transactions even with longer dormant periods - the longer the "dormant period" is, the higher has to be the supermajority. The attacker would have to ensure that he always has enough active (non-dormant) stake to orphan blocks found by the honest minters. But that's mainly theoretical, in what you're right is that 67% is only enough if the dormant period is pretty short, and such an attack should be prohibitively expensive.
A problem that could arise, however, is that the attacker could increase his stake when he succeeds double-spending. He double-spends and with the coins he sold (scamming the exchange) simply re-buys coins "honestly" at other exchanges. He would need plenty of sockpuppets, but there may be still plenty of relatively anonymous ways to buy the coins (and if he's working for a government, he can use "real fake identities"). So once he has 51%, he probably is likely to increase his stake until he gets a supermajority - or the honest minters hard fork away in an ETH/ETC manner.
Regarding DPoS and similar approaches, we mainly agree - with the exception of Cardano which really looks interesting.