Shadow mapping (shadow buffers) is generally much more desirable than shadow volumes (stencil shadows). This is because in general shadow maps are much more flexible in that they dont care what kind of geometry you render. For instance, trees can cast correct shadows from foliage and things like chain link fences, and meshes dont have to be closed.

I dont know how Conitec is doing the shadow volume extrusion, but my bet is they are NOT using the GPU, which is basically a bad idea if you want multiple lights and so on. In my experience Conitec generally opts to use the lowest-tech method for compatability.

Either way, shadow volumes are really not likely to be useful going forward in the general case ( some argue that DX10 will make their use more efficient, but still they cant handle many kinds of geometry).

The soft shadow aspect is not really a consideration, as both shadow maps and shadow volumes can be processed to have softness. Creating a true penumbra is much more costly and not possible without a quite a bit of fiddling, or multiple shadow samples. The Sphere shadows are done using a multiple sample (PCF), but the shadow maps were by neccessity lower res..

However, for a good example of excellent shadows, look at the new patch of Stalker, which has essentially perfect soft-shadow maps for every light and object in high-res (no true penumbra but everything else is perfect). It is essentially high-res lightmaping done in real time (without global radiosity however..). The era of static lightmaps is certainly long gone in commercial games.