A single actor controlling a majority of hash power isn't a problem. Double-spends are a problem. If anyone double-spends, it will be noticed because there will be a permanent record of it in the form of orphaned blocks, and bitcoin value will decrease because Bitcoin transactions will have become less trustworthy.

Thus, even in the case where a single actor has 51% hash power, that actor has a strong financial incentive not to conduct a 51% attack.

The Sybil-style independent mining pools you describe aren't really hiding what people would be looking for, which is a coordinated preference for the fork that causes the double-spend. (Plus miners would have to accept the relayed second transaction into the mempool to begin with, which today nobody does. RBF might have changed this, but even then you should still be able to reason about whether anyone is even contemplating allowing a double-spend.)

I'm not sure if I'm talking past you -- am I correct that we disagreed on what a 51% attack looks like?

Malicious pools can prioritize their transactions and manipulate activity in the network to grant them as many block rewards as they want.