Presumably because there's no real nesting of terms in assembly code? Personally, I find BNF so simple that I'd probably still use it for describing the grammar, but lots of parser generator technology is heavier machinery than required (though a recursive descent parser built from a regular grammar is essentially a DFA if your language/compiler offers tail call elimination).
If there is an ambiguity in the grammar that takes some time to resolve, a recursive descent parser can easily take exponential time. What tail call elimination does is keep you from having a giant call stack, but does not fix the potential (though rare) performance disaster.
I did use a lexer. I wrote an assembler before without using a lexer; using the lexer isn't that much simpler IMO, but it makes it more robust and makes it much easier to systematically catch syntax errors and produce meaningful errors.
Keep in mind it's a personal project, the assembler is written in not the prettiest idiosyncratic Python.