Why? As long as the translator between two different base languages is much smaller than Finnegans Wake, I don't see the problem.

Because if we are deciding whether there are gains to be made by including a simulation of the universe in the compression algorithm. In that case the comparisons to be made are between the representation of a universe simulation, what efficiency this can gain compared to the base language and the translation cost between languages. Since we can expect any benefit to using a universe sim to be rather minimal this matters a lot!

I was actually only allowing the possibility that simulating the universe could be an "ultimate high concept trick" if you were going to be compressing things a whole heap more arbitrary than Finnegans wake (or any human produced data). If you are just talking about compressing human works and lets say using any old language like "ruby" then including a simulation of the universe as part of the message is an ultimately terrible concept. Even narrowing things down from "something represented in the universe" to "something conveniently expressed in ruby" provides an enormous amount of information already.