The tests that output twinning fractions are *not* diagnostic for twinning; they merely estimate what the twin fraction would be if the data were in fact twinned, which can only be decided on the basis of abnormal intensity statistics. (Any version of Xtriage since July should state this more clearly, since we've seen so many users make this mistake in the past.) I'm not sure what you mean by the L-test plot being "sigmoidal" -
usually this is a diagnostic feature of the NZ plot. If you could post
images of these plots (Xtriage will let you save them, probably ccp4i
loggraph will too), that might help.
Given the fact that it scales in P622 despite the ASU being too small, it may be the case that it really is twinned (I'm not exactly sure how to interpret the Refmac results), but you need to absolutely rule out other possibilities before resorting to twinned refinement. It is certainly possible to solve such a structure but very tricky to refine without fooling yourself. At 4Å resolution you will already have a difficult time coping with model bias in the 2Fo-Fc map, and twin refinement will make this even worse (whether or not you actually have twinning).
-Nat