Okay, so... in the coin flipping case, because entropy is already low, I follow the argument that generally attempts to add complexity are likely to make things worse. There's not much room to make things better to begin with.

Not sure that it follow that the same conclusion applies in the cases of other engineering processes you listed, like "configuring a Linux base image for a specific server role" or "setting up a complex cloud environment from scratch" or... "approaching a stranger in a bar"?

Those are already higher entropy processes, with more desirable and more undesirable outcomes, where telling the difference between desirable and undesirable outcomes is much harder to begin with, and so the general advice of 'expend energy to reduce the likelihood of undesirable outcomes' doesn't so immediately suggest that adding what you call 'complexity' is necessarily bad. There are MANY paths to reducing entropy in these sort of situations. Complexity that adds more outcomes can improve the situation in these cases if the new outcomes are both desirable and probable - or at least more probable than the undesirable outcomes they also add.

I just think that if the conclusions from your toy example don't translate obviously to actionable insight into how to improve the toy example, it's unlikely they translate into actionable insight for how to improve real scenarios.