Did you read the article? The issue isn’t code vs. data, it’s code vs. system state. The system can be in states not model-able by code, e.g. versions 1 and 2 of a piece of software running concurrently. Other examples I can think of are things like nodes going down, database consistency, etc.
Technically, not "not model-able" but "not modeled". As in, the effort was not done, and is easy to omit. And doing it in the general case is a lot of work, hence the expand-contract and only-two-versions designs.
