I have a theory that in such cases one might as well just give up and write a Turing-complete language in the first place, as the sort of TC-complete languages we get with this sort of "slowly but surely backed against the wall" situations are way worse than just starting from scratch.

I hypothesize a TC-complete language for something like CSS that included deep tracking under the hood for where values are coming from and where they are going would be very useful, i.e., you would have the ability to point at a particular part of the final output and the language runtime could give you a complete accounting of where it came from and what went into making the decisions, could end up giving us the auditability that we really want from the "declarative" languages while giving us the full power of the programming langauges we clearly want. However I don't have the time to try to manifest such a thing myself, and I don't know of any existing language that does what I'm thinking of. Some of the more powerful languages could theoretically do it as a library. It's not entirely unlike the auditing monad I mention towards the end of https://www.jerf.org/iri/post/2958/ . It's not something I'd expect a general-purpose language to do by default since it would have bad general-purpose performance, but I think for specialized cases of a TC-complete configuration langauge it could have value, and one could always run it as an debugging option and have an optimized code path that didn't track the sources of everything.