I think the expectation is that, if all humans had the same knowledge and were better at thinking (and were more the people we'd like to be, etc.), then there would be a much higher degree of coherence than we might expect, but not necessarily that everyone would ultimately have the same utility function.

Or is it equally plausible to assume that some important values cannot be extrapolated coherently, and that a Seed-AI would therefore provide several results clustered around some groups of people?

There is only one world to build something from. "Several results" is never a solution to the problem of what to actually do.

I think the idea is that CEV lets us "grow up more together" and figure that out later.

I have only recently started looking into CEV so I'm not sure whether I a) think it's a workable theory and b)think it's a good solution, but I like the way it puts off important questions.

It's impossible to predict what we will want if age, disease, violence, and poverty become irrelevant (or at least optional).