The easiest example would be when an earlier step supersampled the data, and you know that nothing above the original Nyqiust could possibly reflect reality. That's one example of when you'd want to zero bins.
Yes, but what's stopping you from using non brick wall windows and filtering nyquist frequencies before sampling? Situations where "unreal" data is present only in higher frequencies is rare, and situations where all data is absent from the higher frequencies are very rare unless you sample in a perfect context where everything is aligned with the sampling period.
What is unreal data? If you have periodic data that's not aligned you're going to have a continuous signal going all the way to the high frequencies.