Shalizi usually makes more sense than this

a sign to give it more consideration.

Your response seems to be that Shalizi assumes an ideal observer, while you assume an observer-in-the-system. That's fine, as far as it goes, but often you assume an ideal observer, and statistical mechanics is able to function with some kind of ideal observer. If you can build a model with an ideal observer, you should!

In particular, when you say that knowledge of particles makes something colder, makes it possible to extract work, you've gone back to the ideal observer.

More tangentially: I guess the point of statistical mechanics is that there may (ergodicity) be only a few possible robust measurements, like temperature, and a real observer can draw the same conclusions from such measurements as an ideal observer. I'm annoyed that no one ever spelled that out to me and Shalizi sounds like he's annoyed by Bayesians who don't spell out their models. At the very least, a straw man gives you a chance to say "here's how my model differs."