The real problem here AIUI - at least in the context of HTML 5's inferred rel="feed" bit - is not just entry documents, it's any Atom document which wouldn't normally be considered a "feed" by a typical user; something that most people would be interested in subscribing to. An example I gave on the whatwg list was an MHTML-like (MIME multipart) package, but there are many other possible examples of course; not all RFC 4287 feed documents are "feeds" in this sense.
If HTML 5 (and current practice) doesn't change, but we defer to them for the specification of autodiscovery, then a new media type would be one way forward. But it should be reusable for all non-"feed" (i.e. from a user POV, as above) Atom documents, not just entry documents; perhaps application/atom-no-feed+xml. It's an ugly hack, but it's better than the alternative of many more specific Atom-related media types, which atomentry+xml might set a precedent for. Another way forward, because the rel="feed" inference is triggered not just by the media type but by the "alternate" relationship keyword, is to create a non-feed alternate relationship ("alternate-non-feed"? ick). I prefer the new relationship to a new media type because it's less disruptive; it doesn't require futzing around with existing specs and implementations. Mark.