Ok, I nailed the bug, but i'm not sure what the correct fix is.
Attached tsearch_morph.diff that remedies this problem by avoiding it.
Also there's a debug aid patch if someone would like to know how i
finally found it out :)
There problem in the lemmatize() function is that GETDICT(...) returned
a value not handled (BYLOCALE).
The value (-1) and later used as an index into the dicts[] array.
After that everything went berserk stack went crazy somehow so trapping
the fault sent me to the wrong place, and every time i read the value it
was positive ;)
So now i just return the initial word passed to the lemmatize function,
because i don't know what to do with it.
So you tsearch guys will have to work it out :)
Magnus