He must be sorting the labels in reverse order of importance. I notice that if I zoom in near Chicago, Chicago itself is missing, big suburbs are in the smallest font, and small suburbs are in a big font. Looks like the data is right but the sort order is wrong.

I am constantly tweaking at label sizes, and which labels to show at various zoom levels, and am not quite satisfied. You've pointed out an issue that is on my todo list. Problem is, it's easy to get sucked into a full day of TileMill styling, and I have to get some other stuff done over the next month ;)

Am I wrong or does it mean that TileMill, while impressive, lacks an essential feature ?

I just haven't taken the time to fully dial in the details available at each zoom. It's really hard, since in one area that level of detail will look sparse (much of the US), but in the UK, that much detail at a particular zoom will look way overcrowded. It's doable, but takes alot of patience and tweaking, which is my next round of styling. That and US state borders being messed up.