Yes, I agree it’s an important distinction.
Given the structure of the CommonMark spec, with its embedded test cases, I thing we should think beyond simply tagging paragraphs as normative or not. The some key parts are the examples, and they are a very special case.
Some examples are completely normative, some negative, and there are other flavors of example flavors in between. For example, it could be very useful flag to tag locations in the of the output section of a side-by-side example where additional attributes are permitted, but not required, in the output of a CommonMark processor. In fact I think this would be key in applying the principles of Semantic Versioning to the CommonMark specification.