precise collision detection with DM and HOS is quite possible and generally would not require a lot of extra work.
You see, any intelligent collision detection system is hierarchical, the best ones go down to per-triangle collision. So the detection goes from general area checks down to general bounding volume checks ( in order of computational cost, i.e. AABB's first, then OOBB's and hierarchical OOBB's ) and only when all of those pass, exat per-triangle collision is checked between two colliding volumes ( both containing perhaps 20-50 poly maxx ).
With DM and HOS one extra step is required. One needs to extrude bounding volume around triangles ( based on precalculated max displacement value in displacement map ) or create bounding volume around HOS control points.
This way one needs to tessellate only the polys that really are colliding on the scene, and this info can be temporally cached in most of the cases.
Also, one could figure out all potentially colliding polys/surfaces first, amd then send them to GPU for tessellation as one batch, and read vertices back.
So some extra work, but probably scant little performance hit. Also, if displacement values are quite small in relation to polygon area ( like max displacement is less than 5% of longest edge of polygon ) and really only used for cosmetic purposes, collision can be safely approximated per polygon ( its not exact anyways in any system, all physics/collision systems have some tolerance in how deeply two bodies can be penetrating )
For example, on scene that ATI demoed at r9700's launch, where F50 tires were displacement mapped so you could see individual grooves in there, one could stil safely do collision detection with bounding cylinder.
Also, shadow volumes arent so impossible to deal with in DMed scene. Only thing that gets more complicated is silhouette calc, so you probably need to batch up polys that potentially contribute to silhouette, send them to tessellator, and then redo silhouette cal on resultant vertices.
All that isnt so simple as it seems on paper and will have some pathological cases where you have to implement stupid workarounds or just have to do with artifacts, but hey, thats what developers do, right ? I mean, lightmaps had its downsides as well but Quake rocked nevertheless ?
, but on the official paper Nvidia says the GF FX supports it, more precisely:
