I took a look at the curves referred to above. The upward explosion in distortion at high power levels is the onset of clipping. Yes, I would expect the "noise" to be signal-independent and constant, while distortion components are of course signal-dependent. Since the curve shows THD + Noise, if distortion is small relative to noise at low power levels, then the curve will rise as power levels become small (i.e., the s/n ratio worsens). If distortion predominated over noise at low power levels, the curve would be more horizontal, at least assuming that crossover distortion
is negligible (which it will be if the amp is operating Class A at small signal levels).
Comparing the curves for the Bryston and the GC, and taking the 20W point for the Bryston so that it can be equated with the THD + Noise figure for the GC at the GC's maximum rated power, it seems clear that the Bryston has distortion performance which is somewhat better, while the GC has lower noise levels.
The slope of the curve for the Zamp is actually not very different than the one for the GC. If you take the cyan-colored (1kHz) curve in Figure 1 of the Zamp review, it goes from -81db for small signal conditions to -87db at 20W. That corresponds roughly to a change from 0.01% to 0.005% (a factor of 2 difference), while the GC numbers are 0.04% and 0.02% (also a factor of 2; the higher absolute numbers may indicate lower feedback).
Perhaps the Bryston's apparently higher noise floor results from the increased gain that is necessary to provide higher maximum output power for the same input voltage range (although based on a quick look I don't think the reviews include sensitivity numbers for the GC or Zamp). In other words, noise generated in its input stages would be amplified to a greater degree than if its output power rating, and consequently its gain, were smaller.
My general feeling concerning distortion specs, fwiw, is similar to Stan's. There are too many other variables, both known and unknown, that enter into the sound of an amp to make these numbers meaningful, aside perhaps from occasionally allowing a poor design to be identified as such. Although I certainly believe that many specs are essential and useful, in that they allow a "short list" of components that may be under consideration to be narrowed down, by eliminating choices that would be poor matches with other components in the system or with the listener's requirements. I'm thinking here of gain mismatches, impedance mismatches, sensitivity and efficiency numbers, power ratings, and things like that.