They create the broken pieces for you, which is sort of impressive on an arbitrary mesh, but it's still not a dynamic breakage. Real breakage is probably impossible in real-time with today's hardware.
Show me this amazing fracturing. There is a demo in the SDK folder, but the project file is missing, and I'm not going to mess around trying to compile it. I believe they may have some rudimentary tests working, but I don't believe it is a supported working feature. In any case, I am not that interested in these fake destruction techniques, with any library, because people aren't going to want to deal with the added trouble of making the assets.
From my perspective, implementing Bullet is just as challenging as implementing PhysX, and the two will each have their own problems and complications I can't predict. However, I know with PhysX any problems it has will get overlooked, and with Bullet I will get blamed for anything that doesn't work well. I got criticized so much for using Newton, even when I proved the rigid body simulator was much more stable than Havok or PhysX.
PhysX has a nice logo, therefore the equations it uses are correct. In the absence of any other compelling technical advantage, that's a good decision criteria. It would take six months working with each library before I could choose the right one based on technical qualities.