Linearity by definition requires that addition and scaling are consistent,
Sure that's a perfectly plausible way to describe the definition.
level up or down will produce consistent results and always land on the curve.
yes it will land on the curve since the curve is a function of level and IMD% not sure how that's helpful for the discusion.
However level up or down obviously does not provide consistent results (addition and scaling are not consistent, thus non-linear). for example the distortion DROPS when the level scales up (between -25dBFS and -20 dBFS) while the distortion rises when the level scales up between -5 dBFS and 0 dBFS.
[/spoiler]
Take the THD+N vs. frequency plot. THD+N will vary as a function of frequency. I'm pretty sure that in almost every case the system will be effectively always linear because whatever frequency you as it to reproduce, it will give you that level of THD+N. I think the only devices that will fail here are those that show random THD+N over frequency. I guess that could be due to leakage from some other internal circuit that varies the amount of interference over time. Maybe discussions about certain devices potentially demodulating AM signals into the audible range are relevant, but that's sort of an extraneous factor to what you're talking about.
What term would you use for THD% inconsistencies with regard to frequency?
Btw this phenomenon is way way more common than you think, infact outside of
very well engineered Class A/B, Purifi Eigentakt, Benchmark ABH2, and some Class D GaN transistor impelementations, mainly Orchard Audio, (and well now the Topping PA5) I haven't seen a 'simple' Class D amplifier really tackle this well. Hypex NCore is good but not perfect.