June mystery: the microdesc consensus is getting 9 sigs, but the ns consensus is getting only 5

Lately (in the past few days) the dir auths are not getting all 9 signatures on the same document. Frequently we're getting 5 that sign one and 4 that sign a different one.

Some more digging shows that it's the ns consensus that they're disagreeing about, and the microdesc consensus is actually doing fine:

Jun 24 18:00:00.337 [info] A consensus needs 5 good signatures from recognized authorities for us to accept it. This ns one has 6 (dannenberg longclaw bastet moria1 dizum Faravahar). 3 (tor26 maatuska gabelmoo) of the authorities we know didn't sign it.
Jun 24 18:00:00.538 [info] A consensus needs 5 good signatures from recognized authorities for us to accept it. This ns one has 6 (dannenberg longclaw bastet moria1 dizum Faravahar). 3 (tor26 maatuska gabelmoo) of the authorities we know didn't sign it.
Jun 24 18:00:00.643 [info] A consensus needs 5 good signatures from recognized authorities for us to accept it. This microdesc one has 9 (dannenberg tor26 longclaw bastet maatuska moria1 dizum gabelmoo Faravahar).
Jun 24 18:00:00.871 [info] A consensus needs 5 good signatures from recognized authorities for us to accept it. This microdesc one has 9 (dannenberg tor26 longclaw bastet maatuska moria1 dizum gabelmoo Faravahar).

I've hacked up moria1 to write out the consensus document that it generates, for each flavor, in legacy/trac#4363 (moved). The theory is that we need to figure out how the ns consensus that we generated differs from the one that the other dir auths generated.

Next step is for moria1 to be the minority voter and me to notice it in time, or for more people to start running the patch in legacy/trac#4363 (moved) and try to catch the bug in action for themselves.