Mostly because it comes out of the RDF/Linked Data people, who have always had strong ties to the web and thus HTTP.

I’m not sure there’s a real reason. Wikipedia says it’s based on pump.io, so maybe it grew out of a prototype?

It seems like a good place to start as any. Putting it on top of TCP seems like an optimisation, that can always come later if necessary.

e.g. the spec also utilizes headers