Well you've mostly answered my question I guess. I was afraid that the coupling between the stencil buffer and depth buffer would
prevent any useful data manipulation methods.
My assumed implementation would be:
1) Determine volume bounds, clip, etc
2) render depth pass \ depth fail (etc) pixels using a sparse sampler to the stencil buffer
where the umbra has 100% probability whilst the penumbra has a probabilistic falloff
3) Do something akin to blurring in the buffer (cheap 2D lerp, etc) in the penumbra regions
Am I missing some fundamental limitation here?