So using a max scalar to fill the gap won't work? At least, for filling up areas below uplifted areas. Vice versa would require a min scalar.
I wonder if it would work if you'd sample the volume over X/Z for unwarped and warped, feed those into a conditional so that when the change in Y is positive, then choose the max of unwarped + warp, and when the change is negative then choose warped only.
Something along those lines.