In my experience Backbone Views become the "top level" of a client-side app, but that is appropriate because they encapsulate controller logic.
I guess I just dislike the convention of interacting directly with collections from places where it shouldn't be done, e.g a todoView removing its own Todo from the collection (hence the risk of duplicating functionality throughout your code), instead of asking a proxying controller to do it. I think this is the reason as to why I've seen a lot of Backbone code where a reference in to the view is stored in the model itself (e.g http://documentcloud.github.com/backbone/docs/todos.html), which disgusts me since it completely misses the point of MVC.
Also, events passing feels more complicated than it would need to be (even if you can always include Backbone.Event).
Controllers are supposed to coordinate the interactions between views and models by propagating changes and events between the two.
For example, if you had a todosController responsible for controlling a collection of views, there isn't really a clean way (in my opinion) to utilize it from two independent Backbone views without passing the controller through to the view (which makes the view depend on the controller). By handling events in the controller instead of in the view, you remove this dependency (which is also similar to my perception of Cocoa's delegates/SproutCore's bindings, but I may be wrong).
Both Spine's and Backbone's view layer consists of rendering templates to the DOM. To me, a "proper" view layer would include things like positioning and both frameworks seem to utilize CSS to do this, for the most part.
I think of Backbone's controllers as a poor man's State chart, without any enter or exit callbacks (e.g going from "#/" to "#/todos" should trigger a callback that states that I will exit the state "#/", so I can do necessary cleanup).