As the docs say, using refcat can lead to all sorts of problems and is not very likely to be what you want - you can’t model all mu as absolute values, because your data only gives you information about the relative proportions of each outcome (that’s why one category is treated as “implied”).
But you want to compare, right?
I usually find it easier to use model predictions to determine differences. For example you can “ask the model” something like “If I repeated the experiment exactly, what would be the mean relative proportion in each category” - for this you would use posterior_epred. I wrote a bit on how this is done elsewhere (in a bit different context), feel free to ask for clarifications
Hope that clarifies more than it confuses :-)