Protobufs are too high level and too opinionated about the structure of the data. It may or may not be "the right tool for the job" depending on the problem being solved.

The goal is to have a schema language that is as universal as byte streams are.

This necessarily means that the language should be open to being extended with new custom types. (Implementations that don't understand a type can always fall back to treating a type as an unsized byte stream.)