A workaround seems to be to run djvm -c bundle.djvu sample.djvu sample.djvu, then djvm -l works on that. From there, it appears superficially to not be too useful. But it actually helped me discover a hack to work around bug 358 (https://sourceforge.net/p/djvu/bugs/358/). djvm -l shows us that page names are assigned as the original filename. These can be used as a handle to make djvused work. I’ll elaborate further in bug 358.
djvm -l cannot handle output from cjb2
djvused -f crashes when feeding it it’s own output (per the man page example)
I wasn't familiar with netpbm. But I would think that the bileveling and the rasterization from a vector image would have to be done in the same step, considering color images seem to need fewer pixels than bitonal images for a given quality. I'm not sure though. However, it's the rasterization that drove me to use ddjvu, which it does in good quality. When I use ImageMagick to rasterize a PDF, it's very poor quality.
Another option is to output in normal mode to a tiff image and use “tiffcp” to recode it with a fax codec. What I've been doing is running ImageMagick on the output of ddjvu to convert to a raw "fax" format, then using fax2tiff -f to convert to a Class-F TIFF, which is specifically designed for faxing. When I look at the tiffcp manpage, I see nothing about Class-F. Is fax encoding implied when using group 3/4 compression? Because for other tools, group 3 and 4 compression is a lossless compression...
If you want to produce a bitonal image for faxing, best is to output to a pbm file. Indeed I've switched to pbm, but that doesn't solve the problem.
I suggest renaming "color" to "whole_image" or "all_layers" and "black" to "foreground stencil". It would also be useful to know the difference between the stencil and the "mask".
The source data, in LaTeX, has no concept of foreground or background AFAIK. I think LaTeX support infinite layers. Is it pdf2djvu that assigned objects to foreground and background? Can that be influenced so everything is in the foreground?