Thanks for the feedback link Nick. I have now proffered several suggestions there.
As for the "difficulty" of handling the more sensible (if less common) inside of the wrist orientation...it is not about the sensors at all. Quite simply, the sensors can sense everything they need to regardless. The development is in the software that "listens" to the sensors.
These days all of this kind of work is done using machine learning techniques. They simply "record" sensor values for many many motions of the arm/wrist/hand, and "label" them as "yes this is an attempt to make the watch come on" or "no, stay asleep". They then use machine learning techniques to have a computer develop ways of comparing new incoming sensor data to what the software has learned about sensor data for these two cases.
So, while it is not a trivial development to make the auto-on detection work for inside of wrist, it would take a small amount of additional work of the same sort they are already familiar with. They would also need to add this as a new orientation in the UI and have the auto-on detection use that to decide which set of learned responses to use.
