You can have lossy compression that results in zero difference to the end image on a pixel by pixel basis.

E.g. if the PNG was 32 bit, and had a full color palette but was filled with a single 8 bit color. You could safely, and "loss-ily", convert the PNG to 8 bit and replace the entire color palette with the single entry for the color that is actually used.

That said, PNGQuant uses dithering so there will often be changes apparent if you perform a pixel by pixel comparison in code.

Just like you, I can't visually identify the difference between a PNGQuant image and the 'raw' PNG that was used to create it (at least not on any images that I've seen so far).