I don't know anything about this spec, but typically the reason is that you want to be able to generate and verify signatures in a place and at a time where the transport isn't known. Therefore there are no "bytes" to sign. In essence the idea is to define an abstract transport (e.g. encode as JSON) and sign that. Then subsequently it doesn't matter how the bits are sent from A -> B -> C, you can always verify the signature by recreating that abstract encoding.

Obviously this is less efficient than signing the transport payload, but that doesn't help if you just don't have access to the transport payload, or when there are N different kinds of encoding used in different places in the system.