Allow me to introduce you to Mozart/Oz and AliceML. Seriously, those two twin languages are the most epic out there.
You would have to hack CL extensively, and bundle various libraries to make it on par with Oz or Alice. Maybe CL + Screamer + FSet + Heresy + pcall + cl-store would be a pale, distant imitation.
For the last 50 years, PL research has been done with minimal kernel languages, implementing few features. Don't know what mental mutation caused the Oz team to go the opposite way and implement "everything imaginable" in a tiny language, but I like it. The core of Oz is no larger than Scheme, btw.
Having said that, I've never actually used it for anything substantial.
I have "ported" my knowledge to Common Lisp and use it there; but I have been actively re-learning ML for the past few months.
And in the TeX case it also depends on what you expect as output of parsing, if you want some meaningful AST, it is even more complex (and probably impossible). Because even if you would remove the ability to redefine character categories, you would still be able to only get stream of tokens without clear semantic meaning.
But whole this discussion is mostly theoretical, because for reasonable inputs (like LaTeX), parser can effectively guess. Also, in CL cases many of these problems are solved by various conventions (like having most of relevant state consistent for whole file and specified in comment that is readable by said tools).
Perl Cannot Be Parsed: A Formal Proof
This statement suggests the author has limited exposure to Common Lisp, which supports the widest variety of programming paradigms in one language that I can think of. Simplicity of syntax seems to be mostly orthogonal to allowing the use of many paradigms, though it's probably hard to have a very complex syntax with a semantically simple language.