When a zero'ed pixel is near a hot pixel (not zero'ed) interpolation can smear the hot pixel when filling in the zero'ed pixel. This is why reyalp wrote rawconvert.
Pretty much. I originally ran into it playing with raw subtract, because it would produce lots of below blacklevel (I didn't know about blacklevel

) output and dcraw would smear all over it, making the output useless for analyzing the difference images

So I decided I needed a "just the data, ma'am" command.
Note that the DNG 1.3 standard allows you to use opcodes to deal with bad pixels, but it isn't widely supported yet. Adding 1.3 as an option for CHDK would be nice.