Adding on to Paul's sage advice, the output from 3dWarp is different between the two datasets because the input is different - the two oblique angles are different. 3dWarp computes a bounding box for the output, so the boxes are slightly different sizes depending on the input grid and the oblique angles of the acquisition and transformation. You could supply a -gridset option that would serve as the master and make the dataset consistent, but it would cut off a row of voxels in this case.

Instead, follow Paul's advice, and use afni_proc.py. That will transform datasets similarly to standard space. Or do everything in native space ignoring obliquity, as in Paul's second example.