I've found multiple sources (eg. Microchip, Analog) saying something like "An ADC with DNL error less than ±1 LSB guarantees no missing code."
If the DNL is calculated only as a width of the "step" and not with relation to the ideal linear transfer fcn... won't having several wider codes cause a disappearance of some other one?
Unless they only meant to write that DNL > -1 means no output code has 0 width?
Edit: added links. After reading more into the details in the Analog's description,
The key is having -1LSB as the low limit.
It seems like my understanding was correct and that just Microchip worded it poorly.