This matches my experience every hire adds a coordination tax, so I now default to contractors with a tight interface until I can name the exact bottleneck. I also like the make the interview task look like the real job framing because it forces a clearer bar on what good output actually looks like.
You separated counting arguments from the actual behavior very well we see in modern training, since that gap often gets waved away on both sides. The human learning analogy also felt more grounding than the evolution analogy, but I still worry it hides how brittle things get once you push past the training distribution.