This is because you are a kilometer from the object (terrain) origin, viewing details of under a millimeter; total 32-bit floating point accuracy is only around 1/16th of a millimeter at that distance, not enough to avoid texture swimming.

Fixing this would require changing the patch generation / instancing and transform code and vertex shaders to combine the patch transform with the camera position at a higher precision. Care would have to be taken to ensure cracks don't appear between patches, as each patch would effectively have a different transform.

The new environment system uses a monolithic camera-centered mesh for terrain rendering, that is much more amenable to fixing this kind of thing; as we can build one custom high precision transform for the entire mesh, and never have to worry about cracks.

We would much rather push that system forward than try to patch the existing one.