I had the unfortunate experience of watching them drop support for the GPU part of my APU while it was still one of their headline APUs.
This because they had based it on a previous gen discrete GPU, and they didn't want to support that GPU range any more with future drivers.
It is experiences like these that makes me understand were RMS came from, and it pains me when so many even within the FOSS world poo poo his message and seem all too willing to break things left and right just because it is "old and crufty" code.
Now imagine you're a driver developer: All this software out there assumes your driver is broken. What happens when you fix it? Well, hopefully the app's workarounds don't actually cause it to break or produce incorrect results. What if the workarounds make it really slow? You want it to run fast on your hardware so people won't buy your competitor's hardware. So now you have workarounds for their workarounds...
OpenGL shaders have (iirc) always been versioned, so that's good at least. The introduction of profile versions also helps but at present even those tend to have major bugs depending on OS and driver, so workarounds still end up in games and applications.
Direct3D was historically much more explicit about versioning and feature sets but there's a lot of variability through things like caps flags (hardware can advertise different levels of functionality) and the app explicitly requesting specific features like a given texture format. Shaders in Direct3D are versioned, but it's quite possible that a driver is only built to handle shaders generated by a given shader compiler, so a new compiler targeting the same shader version will still break the driver. Some shaders probably were generated by old broken compilers, and the driver probably has workarounds for those.
Also, modern graphics drivers typically have a bunch of hacks they quietly apply to big-name game titles and applications to make them run faster or suppress crashes. This goes as far as driver developers actually rewriting the shaders that a game uses, in order to improve performance on their hardware. An infamous old example of this is 'quack3'. https://techreport.com/review/3089/how-ati-drivers-optimize-...
On the bright side, a lot of this madness is gone in Vulkan and Direct3D 12. But very little software is going to end up using those APIs because they're much, much harder to use correctly.
fixed
The main issue is probably due to a bug in the old driver. After a while, bugs become a feature, and it's not uncommon to implement bugs to retain backwards compatibility. This is where the politics enter the scene. Some engineers believe that all bugs must be squashed, while others have a more pragmatic approach. IMO, any behavior that is a problematic deviation from the old drivers, should be changed. Otherwise there is no backwards compatibility, only a best effort approach, which is unacceptable for something as crucial as a graphics card driver.