To me, this problem has long been solved as demonstrated by the following two well-cited publications:

- The Tsukuba report, a.k.a., "A Standard Reference Frame for the Description of Nucleic Acid Base-pair Geometry". When Dr. Olson and I were drafting this report, I felt clearly the need to caution the community of the intrinsic correlations between base-pair parameters and the associated step parameters (Figure 3 there) to avoid possible mis-interpretations in structural analysis. This is specially the case for the effect of shear on twist, since the G–U wobble base-pair is common in RNA and it has a ~2.0 A shear.
- The 3DNA 2003 NAR paper. There is a subsection on the "Treatment of non-Watson–Crick base pairing motifs", and Figure 3 addressed specially on the issue:
"Large Shear of the G–U wobble base pair influences the calculated but not the ‘observed’ Twist. The 3DNA numerical values of Twist, 20° (top) and 43° (bottom), differ from the visualization of nearly equivalent Twist suggested by the angle between successive C1'···C1' vectors (finely dotted lines)."

Coming back to the shear on twist angle issue, the figure at the left shows a G–U wobble pair example (top), and a simple rationale: the base-pair is approximately of 10Å-by-5Å (as defined in SCHNArP/3DNA), so a 2Å shift will lead to an angle:

atan2(2, 10) * 180 / pi = 11.3 degrees(i.e., the red dotted line relative to the bottom horizontal line).

To a first order approximation, that is the difference between RC8–YC6 (or C1'–C1') vs. the base-centered mean y-axis of the pair for calculating twist angle. So whenever one has a G–U wobble pair next to a normal Watson-Crick pair, there would be ~11 degrees difference in "calculated" twist angle between the two approaches (NewHelix/CEHS/SCHNAaP/NUPARM vs 3DNA/Curve+). Moreover, when a G–U wobble is next to a U–G wobble pair, the difference would be doubled to ~23 degrees!

It is worth mentioning that the issue here (as in other similar cases) is not which number is "correct" or which is "wrong": a number is a number. It is its interpretation that matters, and it is here that "details" do count.