I think I've (partly) answered my own question. And thought my findings worth sharing. With a bit of trial and error, I've found that if I prescribe all dimensions used to define my Regions as integer values (by converting the length units from meters to millimeters and rounding) the low precision in the returned RegionCentroid values for my 3D region is eliminated and the 3D centroid xy coordinates match those of the equivalent 2D region (see attached notebook).
I have to say, the low precision given by RegionCentroid for my relatively simple 3D region comprising just 7 cylinders is a little surprising and doesn't instill confidence that RegionCentroid can be relied upon when applied to more complex 3D regions in which the geometry is defined by real numbers. I wonder whether there's a bug or whether there's scope for refinement of the RegionCentroid function?
Attachments: