But what would that do to an Octree that is used for acceleration / compression? The ID stuff I have read indicates that their approach would be static as they are raycasting to optimize for speed and size.
The whole terrain is static and pre-computed into an oct-tree, but you can have minor changes. Basically, if your change is in an already splitted node, you just change the content. Otherwise, you split the node into smaller nodes to handle new data.
As long as you don't change the voxels arbitrily, it shouldn't inflat the data size too much. Consider this: supposed that you have a big ball made of voxels. When you use oct-tree to reppresent the ball, most of the smaller nodes are actually on the surface of the ball (assuming that the ball is completely solid inside). Now, when you want to do some dynamic deformation with the ball, it's very unlikely that you'll want to dig a hole inside the ball where no one could ever see. What you'll want to do is to make some changes on the surface. Since the surface is mostly represented with smaller nodes, you probably don't even have to split these nodes very frequently.
If, by some chance, you need to split a very large node (such as, some one decided to bombard the ball repatedly so it reaches the inner core), splitting the node is also only a local computation (i.e. it only affect the node you want to split, other nodes are unaffected), so it should be fairly quick.
So by making terrain and some mostly static structures (such as houses) with voxels, it's actually more dynamic. For example, you may be able to blow a big hole on the wall of a house, then jump into the house through the hole. In most games where houses are represented with triangles, the wall is most likely just a few triangles, so it's very difficult to actually dig a hole inside these few triangles.
Some elements are probably more difficult though. For example, if you want your voxels to have physical properties (such as, if you blow up everything under the roof, the roof should come down), then it can be more complex. But by treating every movable objects as different objects, it shouldn't be more complex than current triangle based objects, IMHO.