As far as I know the main issue there is that LaTeX is a Turing-complete language and no reliable complete export filter can exist, so in particular one cannot export reliably all of LaTeX to HTML
I don't think Turing-completeness implies any issues when converting programs between equally powerful automata. In fact, a lot of results in the theory of automata / early computational complexity take a form of converting all programs of machine A -- in fact the machine itself -- in order to run it on machine B.
In theory there should be no difference between theory and practice, but in practice there often is. TeX's syntax itself being Turing-complete means you can barely start to do static analysis on the code before hitting a wall you can only get over by installing a huge fraction of the ~1GB download that is TeXLive. "Being theoretically possible" is in no way orthogonal to "impractical in many or most circumstances".
