Yuval, you are assuming that *no one* who uses the os byte APIs on Windows is either checking for '?' in returned paths or catching later exceptions. With Google code search, I did find one instance where someone tests paths for '?' after encoding with the file system encoding. It was not an instance of os.xxx output, but it is the same idea.

In any case,
1. Our experience is that any change will affect someone. I was the victim of a 'harmless' micro change introduced in 3.1.2 (an intentional violation of the bugfix-only rule in bugfix releases -- and the last that I know of ;-).
2. The change will introduce an incompatibility between 3.2- and 3.3+.

The justification that mitigates the above is that there is little reason to request os bytes returns. By the same reasoning, the change is hardly worth bothering with as there should be little to no benefit in real code. So I am +-0 on the change.