> It's a blocker because the fix broke a couple of tests.

That cannot possibly be the explanation why haypo declared
it a blocker on 2012-08-01; the fix was only applied on
2012-08-04.

But I agree that it should block the release now; hence I
propose to roll back the entire set of changes and revert
to the 3.2 state.

> And it's also a regression from 3.1 and 2.7.

Since 3.2 was released with this behavior, this bug cannot manage
to block 3.3.