I agree that if anything between you and the server is going to cache requests then you need well-specified behavior around that. But that particular problem can be solved by a specification much, much simpler than all of HTTP.
I also see that it's useful to know that certain requests will not be state-changing for purposes of pre-fetch. But while I agree this is important for the browser, I question how important it is for APIs. API calls are usually constructed programmatically -- V8 isn't going to start just randomly guessing what XMLHTTPRequests you might come up with next and start fetching them. This is where I think that browser-based paradigms with a human user diverge from programmatic ones.