Martin,

I think this problem came about from the work on PEP
263 (coding spec).  Attached is a patch that corrects a
free memory write.  The problem shows up with valgrind
and test_coding IIRC.

There is a XXX comment in the code which points to
another problem.  It's possible that you could break
and not do a strcpy().  Or perhaps decoding_fgets()
shouldn't call error_ret().  I'm not sure if
error_ret() should free the buffer ever.  I think that
would be my preference.  That way we can deallocate it
where we allocate it.  I think I plugged all the other
leaks.

Let me know what you think.