My experience with dag/cll is limited to building it a couple of times to strip it for references for an Anki deck. That said I think I can contribute at least one insight:
It's really obnoxiously difficult to set up the CLL build environment.
I wound up having to update my X11 server to make it work, which is one of those red flags for me normally. It also struck me as kind of ridiculous as a requirement.
BUT, I have very little experience with document processing, and especially with docbook. (I have been involved with a project to write a LaTeX parser in Ruby, which is about as fun as you imagine.) Maybe there aren't better tools to complete the docbook to PDF pipeline?
All that said, addressing your specific questions:
1) I currently am experiencing the initial disbelief one has in encountering a new project. "Surely it doesn't *need* to be this complicated." Of course, that's the viewpoint that led Knuth to start TeX as a weekend project in the 70s... Every document processing project I've encountered as requirements that led them to their current workflow, and I don't have the expertise to address them, so... My only cogent thought is: given that LaTeX is a necessary evil, maybe it'd be a better source language? At least things like \selmaho could be written directly?
2) This may already be available, but in general the most useful tool for debugging text transformations I've found are source maps. Failing that, the poor-man's version are source comments (e.g. <!-- chapter2/section3.docbook:33-58 -->)
3) On the one hand, I'm personally more comfortable in Ruby, but on the other: is re-writing that conversion process the best approach? What's it written in now?
Judson