A few minutes later:

Setting up this experiment for you caused me to do a few more for myself. I finally absorbed what you were saying a few messages back, Ajax. This problem is a function of scaling the parent. What happens is that the child acquires the scaling of the parent when it is parented: the effect is much more noticeable and distortionary if the child is tilt-rotated.

Since you did mention this problem before, I'm thinking that it isn't peculiar to my version after all.

And it sounds like it can be worked around. If the parent has to be scaled, maybe exporting it and importing it before adding the child will reset its zeroes. I'll try it and report back.