The solution we ended up using was to setup separate one-way parent to child replicas from the remote sde geodatabase into file geodatabase in the central office. So each one has its own parent, or target, FGDB. We then have a python script that combines the data from all those replica fgdbs into a central sde geodatabase.
That python that combines data isn't changes only, it simply reloads the data in sde from those fgdbs.Takes few hours to run but that is ok as a nightly task.
Advantage of that is it is very simple (the simplest possible replication scenario) and in case anything goes wrong with one replica we can just re-create that without affecting others.