I reverted the change in tests/filter/stream.c because it caused a segfault on Linux on Travis. I'm still seeing it as a leak on Windows though. So we can address this separately later.

Looks like a few places call refspec__free assuming that they can always do so. Might be a safe thing to zero the refpsec so that subsequent free calls succeed.

Yeah, most of the calls to git_refspec__parse() do explicitly free the refspec regardless of the parse result. The leak was because some calls just returned if the parse failed.