@svilup do you have some ideas on this problem? I think @akotlar would like to get an ordered simplex prior but also be open to other approaches to fight against the label switching issue.
The topic Ordinal regression with a principled prior seems to be relevant.