How about side-stepping the problem of big lists of scopes by training a zlib or zstd dictionary with the list of scopes, and then compressing the scopes in the

How about side-stepping the problem of big lists of scopes by training a zlib or zstd dictionary with the list of scopes, and then compressing the scopes in the jwt token using this dictionary? Obvious benefit is that you can still represent a scope that is not included in the ones used to train the dictionary (vs. the proposed approach that breaks down in this case)