The problem is that there is a data-type reference counting error some
where that is attempting to deallocate the built-in data-type 'l'


That's what I supposed, but I couldn't find the reason why it wanted to do this
 

It's not really a Python error but a logging.  The code won't let you
deallocate the built-ins, but it will tell you that something tried to.

Reference counting on data-types is easy to get wrong (particularly with
Pyrex extension modules) because most calls consume a reference to the
data-type (if they return an object that  contains a reference to the
data-type).

It is a bug, and it would be nice to figure it out, but that would
require the code that caused it.

I've updated my numpy version to the latest svn, the behaviour seems to be different (more warnings), I'll try to give more information about the error, but giving the whole code will not be simple (it uses a big data file that seems to trigger the error as with other data files, the error didn't show up :()

Matthieu