Cognitive load is what matters

There are separate contexts involved here: the coder, the compiler, the runtime, a person trying to understand the code (context of this article), etc. What's better for one context may not be better for another, and programming languages favor certain contexts over others.

In this case, since programming languages primarily favor making things easier for the compiler and have barely improved their design and usability in 50 years, both coders and readers should employ third party tools to assist them. AI can help the reader understand the code and the coder generate clearer documentation and labels, on top of using linters, test driven development, literate documentation practices, etc.

- Difficult to measure and therefore a hard or impossible to empirically study. (a bad scientific theory)

- Its application to education and learning theory which where a lot of other techniques are more proven.

- The idea that it's a primary mechanism of human learning, which has had a-lot of research showing otherwise.

Though those points seem valid, this article does not concern itself deeply with this concept. The word "mental strain" or "limited short term memory" could have been inserted in place of "cognitive load", and the points raised would be valid. In effect the article argues we should minimize the amount of things that need to be taken into consideration at any given point when reading (or writing) code. This claim is quite reasonable irrespective of the scientific bases of CTL which it takes its wording from.

So i don't think your criticism is entirely relevant to this article, but raising it does help inform others about issues with the used wording if they happen to want to learn more.

I myself find it to be a term that's effectively used as a thought-terminating cliche, sometimes as a way to defend a critic's preferred coding style and organization.