We have re-run a test on our CG246, and the results are exactly as we expect.
We ran a test with both Extended Desktop (left two graphs) and with the internal patch generator (right two graphs).
As can be seen, with 'Null Cube' set (the left graph in both cases), and with a Bypass LUT uploaded (the right graph in both cases), both are identical (give or take slight variations due to drift, etc).
I think this proves what we have been saying.
Something looks to be wrong with your workflow - but we have no idea what that would be.