Even in C you can use something like Ian Piumarta's PEG generator[0] that let's you embed semantic actions written in C. For regular grammars, scanners and lexers you have powerful tools like Ragel[1] (which targets 7 languages)
I'm not a Python person, so if the Python libraries proved inadequate I'd probably resort to these tools and then push out a Python module using Cyphon[2].
Personally I'm not a fan of using semantic actions for AST construction. I generally prefer the 'attribute' approach taken by libraries like Spirit (C++) which make the process feel a lot more like deserialization. SAs are still useful however for rejecting or accepting rules based on complex requirements (like "is this IP address routable"?).
[0] http://piumarta.com/software/peg/ [1] http://www.colm.net/open-source/ragel/ [2] http://cython.org/
A perl6 Grammar is a first-class object.
LPeg is a library, Spirit X3 (for C++, not yet complete) is a library. Both achieve pretty seamless integration with your code. I think in a library is where this stuff should be.
The grammar is just a special class with the rules as methods, so rules with non-standard parsing logic can be implemented using Perl 6 code. Reusing standard rules from custom rules is simply a matter of calling the corresponding method.
The best example of this is probably the EXPR rule in the NQP compiler (which is in turn used by Rakudo's parser), which inserts an operator precedence parser into the recursive descent of the rest of the grammar.
In Python when I have some data and it grows just a bit too irregular for regexes, I have to start manually writing state-machines or add a dependency on a third-party library and spend an afternoon explaining parsing theory to my co-workers.