If the content-type for Atom doesn't sufficiently distinguish between the types of feed you want to serve for HTTP content negotiation to work, it seems to me you should be looking at ways to extend the content-type (like the type parameter that was proposed to distinguish between Atom entry documents and Atom feed documents). Inventing a new form of content negotiation that requires parsing links from the top of an Atom feed is just twisted.
+1. Media parameters are not sufficiently well defined yet, but I believe that this is the right solution to this problem.