Thats because Epic definitely cares 
I’d just say change engine, but at this point I know i must sound like a broken disk 
If you have the know-how you can probably pull source from the last patch and find it fixed after building (assuming someone got to it). And if not, with the source you can track down what the shift is due to.
Generally speaking WPO returns the object’s pixel location.
Landscape has bothing to do with it.
The object does.
With an object placed At 0,0,0 - if you hook a debug node for float 3 into the texture output filtered by the pivot, you should get a print out of the exact object location (so 0,0,0) into the UV.
You may need to scale the uv up or down to get it readable, depends on the model in use a bit.
Anyway, its a quick and dirty way to verify.
If at 0,0,0 an unparented actor returns anything other than 0,0,0… well then the value provided is incorrect (or the hook isnt refined by the pivot, since WPO returns the pixel on scrren value and a UV covers more than 1 pixel?).
Regardless, a 1m cube should give you a z value between 100 and 0 even if you just stretched the debug texture randomly to display it.
Of note, the scale value of WPO is always Uu.
This should match the direct location output in engine - 1m being 100uu.
An easier way to test would be to divide the WPO by 100, and output the result of the B channel to the output.
You should see a UV that goes from 0 to 100 - so covers a 1m cube, the start of which is a value of 0 at 0, and 1 at 100.
A 2x2 cube would have a tiling of 2 white to back rows.
Because it tiles, you don’t know for sure the value is correct - you only know that the division is creating somewhat accurate Tiled results.
But its something…
But you can easily do an IF statement and when WPO B is > 0 lerp the value btween 2 different colors.
When its bot, output as is.
Then theoretically, any part of the mesh below Z 0 will look grayscale.
Any part above it will look as whatevrr colors you lerped.
Needless to say that if an object with a face that goes across the 0,0,0 point and extends over 2m doesnt show you both variations - the output of WPO is incorrect.
