Terrain hit test not accurate

Model always floats above the ground

Log describes the average loss on y-axis to be around 0.2~0.5, it seems that the current search method never decrease the difference between `mid.y` and `GenHeight(mid.x, mid.z)` after a certain value, why?