Well, the circle-ellipse problem complicates that example a little bit: http://en.wikipedia.org/wiki/Circle-ellipse_problem.
What does it teach us? Well, one lesson is that you can't design good inheritance hierarchies in your program simply by aping ontological relationships in a Platonic Ideal World, which is the main point of my article. Another lesson is that sometimes you'll discover that your program is inadvertently violating LSP because of some nonobvious interaction, and that this is a bug, and you need to fix it. A third lesson is that mutability is tricky, in particular with respect to the Liskov subtyping relation; this same problem comes up in non-OO contexts as well, if you have mutability and any kind of subtyping relation. (OCaml polymorphic variants, for example, create a subtyping relation all by themselves, even without OCaml's OO features, and that's enough to create the problem.)
Fundamentally, though, I do agree with your point. You don't use completely realistic examples to teach concepts for the fist time. All of the incidental material that is necessary to grok the example clouds the new concepts. Sometimes this means the examples will be pedagogical toys, and that's okay. The benefit of the already known taxonomies is that the student has less to learn. The new concept is more obvious. If the worry is that the example is too much like a toy, then address that in the second example. In general, I think people put too much emphasis on first exposure to concepts. Education is iteration. If it doesn't take the first time, there are more tries coming up.