One way to efficiently do a spatial query on such a sorted list is the BIGMIN/LITMAX algorithm I mentioned earlier. I believe the originating paper is: http://hermanntropf.de/media/multidimensionalrangequery.pdf (see section 4 — "Range Search").
In my particular case, I also only cared about the "black" nodes in a tree (that is, ones with some data inside), so the linear quadtree is also sparse, speeding it up more. I think since S2 is modeling the entire sphere's surface, it is fully dense. A sparse version might be "only include the cells that contain land," if that were all that was needed for the application.
Though in the sparse version, you don't have fixed-fanout (or at least, it's not embodied in the array), so can't use the trick you mention. Pros and cons (:
One key advantage of Morton codes is that, given a cell ID, you can split it into the "x bits" and "y bits", to get a genuine (x,y) coordinate in the mapped space. Then you can, for instance, easily navigate to your (spatial) neighbor by incrementing X or Y. If I understand your last sentence correctly, this is not something you can do so easily with a plain array-ifying trick.