Ah, the light bulb is starting to glow. Dimly. Thanks Pfile.
What's really happening (I think) is a pixel-by-pixel determination. That's why they call it PixelMath - Duh. T[0], T[1] and T[3] are decided upon "per pixel" and not as a whole channel necessarily. Mind Blown. More wrapping my head around this to come.
Yes, you've got it. Pixel by pixel. Looks like you were thinking the whole image in each channel was getting adjusted, but it's just each pixel.
G: iif(min($T[0],$T[2])>$T[1],min($T[0],$T[2]),$T[1])
The expression looks at each pixel. If the G isn't strong enough, the pixel is magenta=R+B, boost it to the min of R and B. Keep in mind the intent of this expression is for use with SHO images, where the nebulosity and background are typically greens/yellows/browns, and the magenta is usually just a halo around the stars. So the expression has no effect on the nebulosity, just the magenta around the stars - no need for star masks etc. The effect of increasing the G value to that of R or B in magenta areas is basically to desaturate it, as all 3 channels end up becoming very close values.
It is a sledgehammer approach. So if you use it on a standard RGB image instead of SHO, and there's magenta elsewhere in nebulosity (maybe M42 for example?), it will desaturate the nebulosity, too. So in those instances, maybe a star mask would be wise. Only reason I can think of where you'd be doing this, though, might be for magenta/green chromatic abberations around bright stars, for example.