A bit of guesswork here, but: back in ancient times if you flipped or rotated an image your computer would actually unpack the compressed image, transform it, then repack it and every time you did that it added to the JPEG Crunchiness, so more modern transform software instead just tags that the image is rotated and the viewer software renders it that way. My guess is that the software that makes the thumbnails doesn't know how to read that tag. Thanks for coming to my TED Talk.