Mirko, there is one more report about the edges changing at runtime or
after suspend, can you check if this is what you're experiencing?
Also, finally found the upstream bug about inaccurate edge calculation:
http://bugs.freedesktop.org/show_bug.cgi?id=21214