Earlier quoted context omitted.
The transpose is absolutely trivial compared to debayering and compression. It's a lot simpler to do it upfront and not worry about rotation at any later point. And the odds are very high your camera app did already switch memory layouts when you rotated, at least for the UI. Doing that isn't a big deal.
I mean, I get that it isn't incredibly difficult, but it still feels unnecessary. The cynic in me thinks this explains a bit of why the app based cameras are garbage. Do you expect the same when recording video if the user rotates the device while recording? Timestamping an orientation flag is trivial. Why not lean on that?
For anything that has a consistent orientation, storing it that way removes a bunch of annoying edge cases and is very simple to implement.