Slack is certainly a factor and having contextually-aware competent people you can call at a moment's notice is important. I would counter by saying the utilization rate is far below an equilibrium that people are capable of working at for sustained periods of time.

Overall, I expect that the answer is that under-utilization is not as extreme as indicated in the article, and that the value of the knowledge and skill sets of the under-utilized people is valuable enough to keep around.

In an R&D context, I would attribute a sizable amount of being "valuable enough to keep around" to (3) talent is finite and firms are paying everyone so someone doesn't start or join a competitor.

Qualitative attribution level assigned to the respective factor as an explanation.

OP here.

Companies in maintenance mode do not need nearly the engineering headcount they employ to keep their systems running. Furthermore, engineers working at these companies are overwhelmingly not incentivized or sometimes even allowed to take actions that would accrue significant equity value.

I think the question for software engineers is the reverse: why do more productive engineers get paid so little (typically no more than 150% of low-productivity engineers)? It seems to be because companies can't tell the difference, and it's not worth paying low-productivity engineers significantly more on average.

This will become more transparent in short order with quantitative tooling to measure output, but people already have a directionally correct sense of who is productive and who isn't, but social and financial incentives keep that knowledge from surfacing and/or being acted on.