When calculating the mutations that are relevant to a model, any models

found in the new signature are compared to the old signature, and if

their signatures match, the model is discarded from the list of models

that mutations may apply to.
This assumed if the model wasn't present in the old signature, it would

be considered changed. It also assumed that any new models would have a

baseline present in the signature. That was a valid assumption back when

evolution was somewhat driven by syncdb/migrate signals, but Evolver

changes that a bit. It's now possible for a model to not be in the old

signature at the time this computation is occurring.
Models not present in the old signature are now skipped, with the

assumption being that the baseline with all evolutions will be installed

in a later step. This prevents evolutions from incorrectly applying

immediately after a model is installed.

Unit tests pass on all supported versions of Django and Python.
Tested with a customer in production that was hitting an issue when

upgrading from a very old release of Review Board.