I reproduced the problem on Windows.
The exception shown is the AttributeError('next') raised and caught in
lxml._elementpath.find().
The "args" atribute is cleared in the BaseException_clear, during a call
to gc.collect() (in some tearDown() method).
Unfortunately this exception is still shomehow stored in the thread
state structure...

After some researches, I think that the problem is in Cython. Cython
tries to simulate a python frame without using the interpreter loop.
But at least an invariant is not respected: when a frame exits without
an exception set, the tstate->exc_type (and friends) should be NULL.
For example, at the end of the PyInit_etree() function (called by an
"import lxml.etree"), the tstate->exc_value contains an
AttributeError('module' object has no attribute 'unicode').

Now, the consequence may be that all these exceptions are "implicitely
chained" together. And there may be a subtle refcounting bug that shows
up only if this invariant is not respected.