@laz Wow, great find! If you are only editing the OrientationMgr class this will not be propagated to other areas of the code. Something else is likely going on with psimrcc-ccsd_t-2.
In that past I have seen this as an issue where noise around machine epsilon can cause the diagonalization to find somewhat surprising solutions. The closed form for 3x3 matrix diagonalization is not exactly complex (and easily copy/pastable), I wonder if it would be worth implementing here. See here. Although, do note that the eigenvectors can be obtained more easily via Cayley rotations using known eigenvalues.
I wonder if thats a bit of a hammer when the simpler solution should work just fine. @andysim
@loriab Can you poke Dom/Matt, this sounds like a problem for our new crack team.