That's easy... They started out with scanned documents in one orientation. It's easier to rotate than to rescan when using them if one needs them in landscape and you scanned them in portrait. There is NO reason for a program to separate out an element, rotate it, reduce it, and save that data in the Meta data, in a different orientation than the original document. None what-so-ever.
This particular point, I will concede, is arguable.
As someone who works with graphics, I see no reason for a human being to work with images in this way.
In fact, there is a LOT wrong with a human being handling images in the way they are handled in this document.
Once again, though, it all fails the fact of the higher resolution document. Unless you can come up with a convincing and satisfying explanation for that, you must concede that the overall point fails.
And even if you should demonstrate that a human being might do things that way, you still haven't demonstrated that the human hypothesis is a better one than the machine hypothesis.
I just don't see any way for you to get there from here. And nobody has suggested a viable path.
By the way, there are also further reasons for failure of the human-editing hypothesis that we haven’t even TOUCHED yet.