There is a rule in software development, never optimize early.

So I'm going to just use Gotiny - the fastest binary/wire codec I know of, nearly 10x faster than Protobuf. and if there seems to be a need to minimise preprocessing I'll think of other ways to do this.

One thing that immediately comes to mind is the fact that encoded blobs of data themselves can be layered inside other encoded segments, making a progressive decode possible without wasting time on a custom codec.