I understand the issue, and I have done mastering in the past and am generally familiar with the process today. If I get time maybe I'll draw pictures. The problem is the DAC has no way of knowing, barring a predictive decoder before it, that samples are going to be for a signal that's more than full scale. For example, maybe it's a flat-topped wave (square wave) and you don't mean for it to have a rounded top? That said, such decoders exist, and provide the digital equivalent of a limiter circuit. I can think of several ways the DAC itself might deal with the issue, including a few extra lsbs at the top and bottom (over range bits), and filter circuits that allow overshoot in the wave form at the output so the signal is allowed to exceed full-scale after the DAC. I believe the majority of DACs these days are flavors of delta-sigma modulators, and with higher-order loops the signal has to be limited for stability. In those cases, a few samples "over the limit" can be handled. The design will dictate how well they are handled. You can build in a few extra bits in the DAC's modulator to provide headroom (remember the modulator in a DAC is all digital; it is usually sampled-analog in the ADC for a delta-sigma design). I suspect the better ones do, and the corresponding headroom in the output buffers, while the lesser ones do not.
Music gets complex enough to approximate noise signals, and the NPR test is one of the most stringent tests of data converter linearity. Fortunately, we tolerate quite a bit of distortion before it gets noticeable. In fact, I think most of us are more likely to notice its lack (when taken away) than its presence.
Interesting subject! - Don