Yes, it is clear that most image processing algorithms would not work correctly with significant shear.
The main reason I would consider allowing it is the convenience of propagating this information along with the image. For example, right after the image comes out of the reader, you could store the transformation matrix, regardless of sheared or not. Returning a separate acquisition transform from the reader would be less convenient and developers might forget about checking it.
Allowing shearing transform at only some places (e.g., in readers and in some filters and mappers) would mean that we may need to add validity checks, more error handling, etc. It is unclear if the convenience of storing shearing transform justify these potential extra complications.