I think that I see the problem. You are applying the transformations in the wrong order. You are not rotating an ellipsoid in that test at all. Transformations are applied to a point in the opposite direction of what they appear, so in this test, the sphere is rotated and then scaled.
Perhaps this image will help to illustrate.
The blue ellipsoid is generated using the the order in the test (Scaling * Rotation) while the pinkish one is generated with the opposite order (Rotation * Scaling). This demonstrates that you are rotating a sphere in the test and then squishing it to an ellipsoid, not the other way around.
I agree with you that the other way would definitely rotate the normal, but I don’t believe it would in this case. Perhaps maybe Jamis intended for the second case when he made this correction, but inadvertently used this one which has no effect with or without the rotation.