I'm failing to understand your argument.

The image he started with was only 8-bit -- so why does GD introduce loss even without performing any operations on the image?

(I could understand if he started with a 24-bit image ... but i don't get why working on 8-bit images, with a tool that only supports 8-bit images, would cause problems.)

Update: ok... at first i thought the confusion was because the orriginal image changed on the remote site (which it did -- the original one was 128x96 -- the current one is 400x300). But i realize now even the 128x96 was 24-bit ("file" lied to me and said it was 8-bit). It looks like after GD is done with an image, it can still be 24bit (at least that's what the gimp says) but instead of being "RGB Color" it's "Indexed Color (40 colors)"
(file says they're all 8-bit, but says the post GD image is "8-bit colormap" not "8-bit/color RGB")