HTTP caching doesn't work well with compound documents that represent a graph of objects. The kind of caching described in JSON API allows an application to group together requests for documents while avoiding making requests for documents it already has. The only way HTTP caching works is if every request for a document is 1:1 with an HTTP request, which doesn't map onto my experience (and our experience with Ember Data) at all.
> I see you have an application specific need for a certain thing but I don't think it generalises enough.
This functionality was extracted out of a general purpose framework used by a number of applications that made heavy use of it. The ability to send a normalized graph of objects in a single payload, and then make a small number of additional requests, only for the documents that the client doesn't already have, is a huge win for clients working with a non-trivial number of related objects.
The rest of your concerns are very valid. FWIW, I envision future extensions being added to the `meta` section, which is already reserved. ID, URL, and rel are the building blocks of the graph, while other kinds of extensions (like pagination, etc.) are optional metadata. But you raise a good point and I will give it some thought :)