That assumes there's some way to grind through lots of computations to manipulate proof of stake, which isn't the case with the newest designs.

Ethereum's proof of stake FAQ has responses to that article: https://github.com/ethereum/wiki/wiki/Proof-of-Stake-FAQ

Not really, it just assumes on the economic principle that your costs and your revenue will converge. In a free market, if you are spending 50$ for a chance of winning 100$, everyone else is incentivized to spend 51$ and more to have a higher probability than you. Eventually the people spending 99$ win every block.

If the costs don’t manifest themselves in the protocol, they will be external costs. Hence, “obscured proof of work”.

In other words, the problem is not that Bitcoin costs us X per unit produced, it is that we value Bitcoin at 8000$ per unit.

That only works if there actually is a way to spend $51 and get a higher probability, and that's not necessarily the case. E.g. the Dfinity protocol uses threshold signatures so with any m-of-n stakers submitting, the same random number will be produced.

Besides, in Casper at least there are internal costs: the locking and potential loss of stake.