I think you’re right that a UDT agent who chooses policies could have preferences over those policies that satisfy independence, with the apparent node-level independence violations being just an artifact of projecting policy-level optimization down to individual nodes.
The core argument is against the standard application of independence as it is done in practice: to individual lotteries and individual choices, with branch-by-branch evaluation, which is how it’s used in the economics literature and in most discussions of rational agency. Both UDT and EE move away from this level.
Or, in other words: for the things over which people actually attempt to maximize expected utility or model it, expected utility maximization is a wrong strategy, because independence doesn’t hold for these things.
Also, for EE specifically, the ergodic mapping makes the effective utility function dependent on the dynamics of the process, which means that choices across different dynamic environments (where the dynamics themselves are part of what you’re choosing) may violate independence even at the trajectory or policy level. Whether this is a real violation or can be further absorbed by moving to yet another meta-level is, I think, an open question.
Overall, it looks like there always may be some “meta” level on which independence holds, but is it useful/meaningfull in practice? You may be able to rescue independence by climbing to higher and higher levels of abstraction, but at the level where EUT is really deployed, where economists fit CRRA parameters to observed choices, where financial advisors recommend portfolio allocations, where rationalists tell each other to shut up and multiply, independence fails and EU maximization gives the wrong answers.
Ihor Kendiukhov comments on On The Independence Axiom