>I seem to recall the proposed staking minimum being around $200,000...
No, it's 32 eth, which is a bit over $4k.
>https://download.wpsoftware.net/bitcoin/pos.pdf
Stake grinding is an obsolete attack (solved by randao, in the future strengthened with VDF asics).
The second argument that weak subjectivity is somehow unsafe is at odds with reality: it assumes some far away hermit that runs an old node after 10 years of hibernation, with no ability to communicate with others otherwise. In reality, crypto is a technology for resource allocation among humans also participating in that specific system, which means the only constraint is to make the bonding period sufficiently long that manual decisions are feasible and not overly costly. A system that requires a node to run for few minutes every few months to follow the same chain fulfills those conditions.
>Moreover, as was recently demonstrated with "steem" PoS
No, steem has DPoS, which is very different in practice. It has inherent centralization because there are only 21 witnesses, as opposed to potentially millions. It has stake delegation baked in the protocol which ensures all witnesses are public figures that know each other, which makes a cartel the expected outcome. Nodes are by design heavy which makes outside verification very hard.
Eth2 has to support up to millions of nodes at once. It has pro-decentralization penalties - penalties grow if others are misbehaving at the same time - which means if most of the network is on aws and it goes down, they start to lose their stake very fast, as opposed to random home node going offline in an uncorrelated manner for (most likely) no penalty at all. Same goes for slashing incidents due to contradictory voting.
The system is verifiable externally and can be randomly sampled, because it's stateless and state root is part of the consensus. This also means a block that tries to do something against the rules automatically functions as a fraud proof given only its parent's block header. It's not possible to design a system that's more easily verifiable: all it takes is one person somewhere to observe incorrect behavior to alert the others.
There are going to be centralized staking services, but they are inherently going to charge some fees, and given how light one staking node is going to be and the correlation penalties, most likely they aren't going to be a significant portion of the network.
>never-arriving
It turns out it's not easy to design a system with all these characteristics. PoW is an easy and a temporary hack solution, but that's all it is. Mining (at least sha256) is now fully centralized in China.
If a PoW network ever became really important - not as a speculative toy mainly for rich Westerners, but as something used by countries like Iran to evade sanctions on a massive scale - mining would became fully regulated with enforced kyc on every transaction. It's trivial to do, there's no way to hide those mining farms.
PoS can fully function on tor or other anonymizing network.