And even when you mask them out so that they are no longer visible in the "all white" (paper) background, e.g. by messing with the white/black point of the image there's still the possibility that they could be recovered with correlation methods in grey areas where they aren't visible to the naked eye or just by increasing the contrast.
They didn't say "convert to greyscale".
Very good point. But even then, assume that one page of a leaked document contains a large picture with areas around the thrshold value: With the agency being able to recreate a perfect replica of the initially scanned paper version, but without yellow dots, it might be possible to extract the (very few) bits necessary to boil it down to a single printer serial number by statistical methods.
...and they focus on adding fonts of multiple sizes so it can't be shrunk without losing information.