I have to admit I don't quite follow all of your explanations. However, it seems the c_variation things are what would tell the generator how to produce the derived raws, and it might be it (already/mostly?) does what I'm after.
My understanding of your c_variation description is that it directs DF to generate derived creatures without any explicit raws for these creatures?
I'd say a Mussel Man is not dissimilar from a Slug Man or any arthropod Man (although slug or worm is probably the closest, with the imagined result not that dissimilar from a ninja turtle). It's also possible the generation would just say this base isn't valid for generation.
I tried to specifically not block modding with the suggestion in that the conversion should generate new raw descriptions from the base one, but without the need for manual maintenance when the base is adjusted. Toady running the tool would generate a bunch of raws that are then used by DF itself together with the base ones, with DF having no knowledge of where the raws came from, so any modder could modify the generated raws, and as long as they stay within DF's boundaries (i.e. known tags and tag value ranges, etc, as per normal), the modder is free to replace, change, or remove any or all raws as desired.
If the tool was provided to modders, they could create a new base creature, crank the lever, and get the derived raw versions generated. If desired, these could then be further modified (but re-generation of something modified would require manual copy and merge). A modder would be free to ignore the tool, and just manually create the derived creatures (if any) from scratch.
As far as I understand, there's no relation between the various versions in DF itself (such as the Man version having some kind of affinity with the other ones), but if there are/were/will be, the generated versions should be tagged by the tool to point back to the base creature.
Basically, I want to remove tedious manual generation of things that seem to follow fairly simple rules, and that's what computers are very good at (body size of Man version being an average of the animal and a human, or whatever it was, for instance), and not making typos... If it doesn't work because of too many complications: too bad. If it is implemented already: so much the better.
And yes, if the base is missing, nothing would be derived...