The main problem I had with trying to implement this is, that’s not quite the case. Vertices outside of the cage get bound as well. This is probably related to bind precision, and maybe precision 9 binds would work okay, but precision 9 binds take my computer days to bind. (The best example of this problem is using a low precision bind to do a full-mesh deform, as with Blenrig, on a T-pose model: look what happens to the groin as the legs spread.)
But I’m writing again, after a long time, because I’ve come up with a solution to this problem.
What I’ve begun doing with my mesh deforms (usually even my non-dynamic mesh deforms, because it’s a convenient solution) is to create two shells. In addition to the basic mesh, I’m also now creating an outer mesh. Normals for both of these shells point outwards from the center of the shells-- it’s not a hollow mesh, it’s not for drawing, it’s for calculating. Then, I never, ever manipulate the outer shell from its bind position.
What this ends up doing is creating a smooth envelope around the mesh deform where any affected vertices interpolate from fully deformed (at the border of the inner mesh) to not-at-all-deformed (at the boundary of the outer mesh.)
For a non-dynamic bind, this isn’t strictly necessary, because you can use vertex groups to smoothly limit the effect of the mesh deform instead. But for a dynamic bind, you can’t use static vertex groups, because you don’t know where the deform is going to be active; and there aren’t good tools to create dynamic vertex groups to smoothly limit it. (A “volume” mode for vertex weight proximity modifier would let you create dynamic groups, just in case anyone is listening.)
Like I said, I’m even doing this for non-dynamic mesh deforms, because it’s easier than weight painting, and it doesn’t seem to create any slowdown beyond, possibly, a slightly slower bind.
But for dynamic mesh deforms, it’s totally essential, and these are now working well for me. You may still want to do some weight painting-- it’s easy for an arm to enter both the outer and inner area of a scapula deformer, for example. But the weight painting is often unnecessary, and can be very, very crude.
This innovation is enough for me to be using dynamic mesh deformers now. The main reason I prefer them vastly over the cast modifier solution is that I don’t have to model in anticipation of a cast modifier. Instead, I can model exactly how I want my model to look, and only later worry about how my mesh deformers should expand the mesh, and where. After all, if I don’t do anything to an armature modifying the mesh deformer, the mesh deformer won’t do anything. (And usually, what I’m doing is some combination of XZ scaling from volume conservation and Y axis translation.)