You're accusing me of group selectionism? We might disagree on a point of terminology, but come on, I'm not a completely nutter. Anyway, my point in quoting the wikipedia article is that too much dishonest signalling makes signalling completely pointless ('weakens the integrity of the signalling system'), so for signalling to work you need some way of keeping out the cheats. I'm not proposing anything as daft as "groups without cheats will prosper". Indeed, that's why I was making such a big deal about criterion 4 and cost asymmetry, because the analysis of signalling has to work on an individual basis, including the individuals that might be tempted to cheat.

In my limited imagination, the only way I could think of for keeping out the cheats was having an asymmetric cost structure for honest signalling compared to dishonest signalling. Thus cheating wouldn't be worth it. I now realize this is not the only way. ialdaboth called my attention to Batesian Mimicry, where cheaters are "kept out" simply by the fact that mimics are comparatively rare. Doubtless other ways could be invented.

I think I prefer MagnetoHydroDynamics definition of signalling, and would reserve my criteria for describing costly signalling.