We have generally made an exception to the "new feature" rule for mimetypes.  That is, we don't really consider a mimetype addition to be a new feature in the sense that our backward compatibility rules mean.  It is true that an application could work on x.y.z and break on x.y.z-1, but this isn't because an *API* present in x.y.z is not there in x.y.z-1.  is more akin to a bugfix (it threw an exception before, now it works).  Think of the absence of the mimetype rule as a bug, rather than its presence as a feature.

And yes, this is a policy evolution.  This way of looking at mimetypes changes has been in effect for....maybe five years now?...and before that we treated them as features.  But then, too, before that we required there be an actual IANA accepted RFC, but that requirement too has had to evolve as mimetype management became more decentralized.