fixed sources in replication are now a suggestion
Code used to hard error if a fixed source was missing. That would be useful
if, say, FSCK decides that some copies are bad and it should be re-sourced
from a particular device. However FSCK can just nuke or mask bad rows and
re-run.
So... attempt, then give up. We can always revert the behavior if some feature
requires it. New file uploads no longer use fixed sources, which gets rid of
most of the issue. This change should let existing clusters clean themselves
up.