We would all love WPO accurate collisions, but…
Conceptually, the fundamental issue here is that collision is handled on the cpu completely decoupled from rendering which is on the gpu. The collision shape use for line traces can be a simple capsule or the actual mesh. Within the scene, there is an optimized Bounding Volume Hierarchy (BVH) that the static shape is inserted into and queried against. This query (linetrace), is able quickly determines per triangle intersections because of being stored in the bvh.
On the other hand, WPO is processed on the GPU for each vertex. The vertex positions are modified (temporarily for this frame) before getting rasterized.
So the WPO deformation briefly exists only on the GPU while rendering this patch on this frame, with no mechanism to query.
If you were to apply the WPO on the cpu side, it would take a huge amount of time and memory. There are static caching techniques like Alembic which theoretically could be used to query collision for vertex deformation. But these are already typically much slower and more memory intensive than generalized WPO.
In the future, we may see the collision and physics moved to the GPU. In that scenario it might not be quite as difficult to query a gpu side cache of the wpo… will have to wait and see 