"I imagine the actual integer promotion rules are compiler dependent..."

I don't think so; I think the rules are defined by the ANSI (now ISO) standard - that's why Keil calls it "ANSI Integer Promotion" (my emphasis).

I haven't read the ANSI Standard (either), and I don't have my K&R to hand at the moment, but this is my understanding of the situation!

BTW: Disabling the Promotions in Keil certainly does make a noticeable difference to code size, but will break code that relies upon the rules! (this is most likely to be significant with bought-in code; eg, protocol stacks...)