After some work I finally figured out what it was doing.
The image in question is entirely black. So it generated the palette and observed that there was a grand total of one item in it.
Well, the basic formula is "possible_items = 2^bits". If possible_items = 1, bits = 0.
So it concluded that it could encode all the visible data in zero bits per pixel.
The ironic thing is, it's *right*. If there's only one palette entry then it obviously doesn't need to branch to find out which it is - it *knows*. It's that palette entry! There's only one value that's *possible*!
End result . . . it encodes the picture with zero bits per pixel.
Plus the palette, of course.
(It makes sense . . . it's just kind of disturbing when you see it for the first time :P)
Incidentally, it's generally holding at under 2 bits per pixel overall, which isn't bad compression at all.