In doing testing on the fix for RT-38948 under a variety of rotations I discovered a deficiency of the Shape node when dealing with 3D transforms. It snapshots the scale of the first non-2D transform it encounters and caches the node at that scale and doesn't invalidate the cache until the geometry of the Shape node changes. If the first 3D rotation it encounters is nearly perpendicular to the screen (Z=0 plane) then it will snapshot a very tiny scale for one or both axes.

I'm attaching Text3D.java and an example of the output it generates. The UL text is being rotated in the 2D plane, the UR text is rotated about the X axis starting at 0 degrees, and the bottom 2 text objects are being rotated about the X and Y axes starting at close to 90 degrees where they snapshotted a very small X or Y scale.