Indeed, syncprov_operational() is right: it duplicates the entry only if contextCSN is already present in the entry, and thus may need to be updated. What's wrong is back-bdb, which gives up when the entry does not have e_private set appropriately, while it could do more to find out about the entry's subordinates. This may explain why in some cases the attribute is present. This occurs whenever contextCSN is not already present in the entry. I'm preparing a fix for back-bdb.