It's more of a social agreement. If you get a JSON object from a place you expect a JSON Feed and it has a title and items, then it'll probably work, even if it omits other things.
If it's JSON your program expects then I should be able to throw any valid JSON at your program and it should work. Granted, it probably won't be a very interesting program precisely because JSON is just generic data without any meaningful semantics.
This spec is entirely about attaching semantics to JSON documents, but all that gets lost when you forget to let people know the document carries semantics and just call it generic JSON. Maybe that doesn't matter to a JSON-feed specific app that thinks any JSON is JSON-feed (an equally egregious error) but if there's an expectation that I should be able to point my catch-all program (i.e. web browser) at a URL and it should magically (more like heuristically I guess, potato/tomato) determine that the document retrieved isn't in fact just any JSON then things are about to get real muddy. Web browsers aren't particularly social, so I suspect a social agreement probably won't work that well.
Media types aren't just something that someone thought was a nifty idea back in the dizzy, they are pretty important to how the web functions.
That's not a valid argument, because JSON is just a serialization format for an arbitrary data structure. You can't throw any arbitrary data structure at any program that accepts data and expect it to be able to accept it. Every program that accepts input requires that input to be in a specific format, which is nearly always more specific than the general syntax of the format. And aside from programs that make strict use of XML schemas, they pretty much all use the handshake-and-wink method for enforcing the contract. (Or to put it another way: documentation and input validation.)
My take on the author's approach is that the content-type is specifying the syntax of the expected input, and the documentation specifies the semantics and details of the data structure. In that respect, the program works like most other programs out there.
If I'm pulling JSON from any API, I expect it to match a certain schema. If I expected { "time": 10121} from a web API they send me "4", then sure, that's valid JSON, but it doesn't match the schema they promised me in the API.
Something that's JSON should be marked JSON, even if we're expecting it to follow a schema.
