Right. Exactly setting the timestamps isn’t something a computationally-bounded miner can do because it is still a Poisson process.
If you try to adapt Sjors’ scheme to a real algorithm for deciding timestamps, I think you exactly get the scheme I posted above. And we simulated that: it has no compounding effect because the attackers’ block production rate is affected by the difficulty adjustment they caused, providing negative feedback. But, again, even if there was a compounding effect it would not be a concern, as mentioned above, because it’s at worst equivalent to attackers just choosing to reduce their hashrate.