My understanding is that Ancestry treats Marriage as two individual events, one for each spouse, with one citation of the common source for each. And that’s what is trsnsmitted to RM via the API. RM then tries to combine the two individual events into a single family-type Marriage event with one citation of the common source. For TreeShare transmission in the opposite direction, RM has to reverse or deconstruct the family-type Marriage event into two individual-type events, each with a citation.
So this translation complexity carries some risks. Early on with TreeShare, I was startled to find individual-type Marriage events popping up in RM from TreeShare. Not a standard fact-type but a custom one created as a result of TreeShare, maybe residue from a fractured download before combining was completed.