We now test all the code through valgrind; a single memory leak will complain. There are also static analysis, and various other checks.
And BTW, you can't treat warnings as errors. Just try to fix the warnings we have and you might understand why it's not possible easily (might require API changes, compiler false-positive, deprecated warnings that can just be silenced, etc). Of course, you might be able to fix some easily, and the patches will likely be accepted (assuming they are correct).
You are absolutely right that not all the warnings are accurate, and that not all code should be changed to satisfy a every compiler. But I'm arguing that there is a long-term benefit to choosing a single compiler, and making the modifications necessary to allow this compiler to run without warnings, even if the behaviour is provably correct without doing so.
On top of that, sometimes warnings come from parts of the code you don't control, like a scanner generated by flex.
I think it's better to have a strong policy of not introducing warnings and fixing them whenever a particular compiler is foud to emit them, but having -Werror in the default Makefile can quickly inundate the project's mailing list/IRC channel with "doesn't compile" complaints from people with exotic compilers.
I think SQLite has a good attitude: http://www.sqlite.org/faq.html#q17
"Some people say that we should eliminate all warnings because benign warnings mask real warnings that might arise in future changes. This is true enough. But in reply, the developers observe that all warnings have already been fixed in the compilers used for SQLite development (various versions of GCC). Compiler warnings only arise from compilers that the developers do not use on a daily basis (Ex: MSVC)."
The code wasn't merely copying an uninitialized variable, it was branching on the random data that happened to be in the data returned by malloc(). This confused Valgrind, but was thought to be safe due to some later check. To me this was fragile code and unsafe practice. To the decision makers, it was a good way to save a couple cycles and a few bytes. The culture (at least at the time) felt that efficiency trumped clarity and maintainability. This is a blessing and a curse.
The quality of the code base also left something to be desired. It was fast, and generally worked, but they certainly didn't subscribe to the camp that believes in treating compiler warnings as errors. I remember using Valgrind and discovering that some memory was being used uninitialized, and submitting a patch. The patch was rejected with a comment such as "not needed". I still don't know if this was correct, but I do know that being a static initializer my patch wasn't going to slow anything down.