VR does work when using the back button. As regards why turn VR off if you are shooting at speeds above 1/500th of a second, this was Thom Hogan's explanation some years ago, when asked about it:
"Don't you get some effect from VR even if your shutter speed is above 1/500? After all, the VR elements are probably moving between samples." Yes, sometimes you get a VR-like effect above 1/500, and it's probably because the elements are in near constant motion and the designers have picked a movement frequency and smoothing curve that takes advantage of the known sampling frequency. But the problem with using VR above 1/500 is that you will get clear image degradation often enough that you'll get burned by it. And I believe you get burned by it more often than you'd get burned by having VR off. Again, Nyquist tells us that when we sample something, we can only be "precise" about our data at one-half the frequency. Above that you don't get useful data, and a mechanical system can induce ringing effects as it tries to adjust. Let's see if I can explain it simply (a very gross generalization and simplification coming up): Pretend we're moving enough to impart a different motion that needs new correction ten times a second. Assume we're sampling five times a second. So what if we move Left, Right, Down, Left, Up, Right, Right, Left, Down, Up? The sampling sees Left/Right, Down/Left, Up/Right, Right/Left!, Down/Up! See the problems? If we take images at ten frames a second, the system is lagging us in the first few samples and may settle down while we're still moving in the last two. The problem with Nyquist is that there's a strong chance that the system is going the opposite direction you want it to when you exceed the sampling frequency. But, yes, there's a chance that it's going the right direction, too. Not a good enough chance to use VR, in my opinion. Moreover, I don't know of a working sports or wildlife pro using the long lenses that hasn't discovered the same thing by practice: VR tends to degrade shots above 1/500. "
I'm not clear if the sampling frequency of has been changed to address this problem in newer cameras. If you're shooting at 10 frames a second, designers may have had to change things to accommodate this.
The newest cameras like the D850 are now auto fine tuning and correcting their auto focus systems, so the problem of focus with differing lenses will probably be gone from all cameras in a few years. A good thing too.