If his attempt fails, his stake won't be blocked, because his chain wouldn't be selected at all. It's as if the attack didn't happen. So he doesn't have to wait.

Where you're right is that with a majority as low as 51% it is probably difficult to control a chain permanently. But that only applies if the other 49% all mint actively. With two thirds of the active stake it should be possible to control the chain permanently and censor transactions, regardless of dormant periods.

Now even with "only" 51% the attack can do a lot of harm. The attacker can try to attack/double-spend again and again. No exchange would be safe, and so the coin would be probably delisted from all exchanges until the 51% scenario ceases - or exchanges would have to set, as you wrote, the confirmation threshold to 100% of the reorg limit, which are typically days. If the attacker doesn't sell his coins because his intention is to destroy it (e.g. because he short-sold coins before) then the only way to stop that scenario (that makes the coin de facto unusable) is a complicated hard fork "tainting" all UTXOs that have been part of the attack and block all tainted UTXOs.

At a first glance this approach looks good - but why is no PoS coin doing that? I think that it's possible this approach could add attack vectors for limited short-range attacks using network disruptions to confuse badly-connected nodes.

A miner with 51% of the hashrate would not get all blocks, so he also cannot censor transactions.