> Also, one at the time controversial decision which worked really well in hindsight was that D3D was an entirely new API in each new major version, which allowed to leave historical baggage behind quickly and keep the API clean (while still supporting those 'frozen' old D3D versions in new Windows versions).

This is interesting. I have always wondered if that is a viable approach to API evolution, so it is good to know that it worked for MS. We will probably add a (possibly public) REST API to a service at work in the near future, and versioning / evolution is certainly going to be an issue there. Thanks!