WRT the Debian version, I am somewhat less convinced, mostly because of the Mesa drivers. As I understand it, llvmpipe is evolving quite quickly, so tracking a more recent Mesa version might me more helpful for development. OTOH it's also valuable to check that things didn't break for released software. Maybe this is just a good reason to test on different Debian versions. One could say that in general more variations you test with, more useful the CI is; and that current pipelines are still quite fast, so we still have a few points to spend before the whole thing becomes unwieldy.
My consideration is mostly that "unstable" is not a fixed version, and that may make it harder to reproduce results.
Testing both the latest versions and the oldest supported versions of dependencies makes sense to me; no disagreement from me there. It does probably imply that MRs bumping the minimum supported versions of dependencies are going to require CI changes.