Maybe patchwatcher could detect new warnings (e.g. using git-blame as I described before) and either flag the patch outright as incorrect, or give it an intermediate state between bad patches (those that don't compile or cause test regressions) and good ones (that compile and don't cause test regressions).
Since Rob was noncommittal about whether -Werror was a good or a bad idea, I'll take a stance: I think it's a bad one. For one thing, different optimization levels produce different warnings. For another, some warnings are just plain wrong. In some cases the "fix" to quiet the warning is uglier than the warning.
Having a code base that has hundreds of warnings is also a problem, but we're fortunate enough not to be in that situation.
I think our current system isn't bad: if one of us makes a mistake and it gets past Alexandre, someone often points out the problem and it gets fixed. With -Werror, a mistake (or a gcc bug) becomes everyone else's problem. I don't think that's a step in the right direction. --Juan