homepage

Since I too was bitten by this issue, I'd like to support Gregory's
request, and report my rationale for changing the current behaviour.
With the current behaviour, we could see (and I *have* seen) scenarios
like this:
	
 1. A test exposing a known bug is written, and the test is marked
 as "expected failure".
	
 2. Some refactoring and code improvements follow.
	
 3. They improve the overall behaviour/correctness of the program
 or library under test, to the point of fixing the bug "behind
 the scenes".
	
 4. The developer doesn't notice the fix though, since the testing
 framework doesn't warn him "strongly enough" of the fact that the
 test marked as expected failure has begun to pass. (To reiterate,
 the current behaviour of just printing a warning saying "some test
 passed unexpectedly" on the standard output is not good enough of
 a warning; it's easy to miss, and worse, it's *certain* that it
 will be missed if the tests are run by some CI systems or a similar
 wrapper system -- those would only report failures due to non-zero
 exit statuses.)
	 
 5. Without noticing that the issue has been fixed, the developer does
 some further changes, which again re-introduce the bug, or a
 similar one that the test still marked as "expected failure" could
 easily catch.
	 
 6. That test starts to fail again; but since it has remained marked
 as "expected failure" all along, the fact is not reported to the
 developer in any way. So the bug has managed to sneak back in,
 unnoticed.
In addition to this rationale, another (weaker) reason to change the
existing behaviour would be the "principle of least surprise". Among
other widely used testing framework (for python, or language-agnostic)
many of those which support the concept of "expected failure" will
throw hard errors by default when a test marked as expected failure
starts to pass; among these are:
 * Python libs:
 - py.test (http://pytest.org)
 - Nose (https://pypi.python.org/pypi/nose/1.3.0)
 * Language-agnostic protocols/frameworks:
 - the TAP protocol (the standard in the Perl world)
 (http://en.wikipedia.org/wiki/Test_Anything_Protocol)
 - the Automake "Simple Tests" framework
 (http://www.gnu.org/software/automake/manual/html_node/Tests.html)