In a recent commit the factor (-1)^l that was present in atomic_wfc_nc_updown
(spin-orbiit/noncollinear case only) has gone missing in the merged code.
It was not present in atomic_wfc and nobody had noticed its absence.
It would have gone unnoticed forever, if it weren't for DFT+U for spin-orbit /
noncollinear cases contained in the examples for HP: for mysterious reasons,
stresses are different enough to trigger a discrepancy warning.

I do not think that a factor mutiplying atomic wavefunctions should make any
difference in final results, so I see two possibilities here:
- a serious numerical instability, or
- some hidden implicit assumption on the form of the atomic wavefunctions
  in the stress code