So it's more like noise recognition, right?
Yes, it's comparing the neighboring pixels for high contrast (which is not necessarily noise!)
Doesn't it get crazy if you have areas of your scene that are intentionally noisy - for example a noisy texture or a glowing lamp with tiny holes in it?
Yes, it increases sampling on contrasty edges, no matter if it's clean already or not - and that's why it's not really useful. A better adaptive algorithm will have to recognize changes in noise (over a couple of passes), not the presence of noise in one single pass.
