Hi Yuri,
1- I think its somewhat of a consensus that at higher res. it should not be used to avoid missing out "real" detailed discrepancy. Is this true?
yes and no. Relevant methods are in active development and in general things change quickly. To get some very superficial impression, see pages 56-59 here: http://www.phenix-online.org/presentations/latest/pavel_refinement_general.p...
2- Should one try to use it in lower res data (3 A and worse)?
Definitely yes. In phenix.refine there are two types of NCS restraints, applied in Cartesian or torsion angles spaces (global/local). Torsion NCS supposed to be the best but the relevant code in phenix.refine is too fresh and requires more testing. You should try both to see which one works best for you.
3- Is the presence of NCS macromolecule specific, or crystal specific?
It's rather packing.. Say, if more than one copy packs in the asymmetric unit then you have NCS, roughly speaking..
In other words, if for the same enzyme one crystal goes to 3.1 and another 1.5 A could one have significant NCS but not the other?
I'm not sure I understood this... May some one else did? Pavel.