If I remember correctly, the reason was to prevent a research-grade observation (with agreement on species) from appearing as a definitive subspecies ID if there is only a single identifier suggesting the subspecies.
Imagine you have a RG observation of a species (for example, the painted turtle Chrysemes picta) and somebody adds a subspecies ID (western painted turtle, C. picta bellii). If you allow subspecies IDs to change the observation ID (the “headline” of the observation) the same way as other taxonomic levels, you either have to remove RG from a perfectly good observation, with a non-controversial species ID, or keep RG and lend inappropriate weight to that single person’s subspecies ID. If they were wrong, and it’s either clearly a different subspecies, or it’s not possible to tell based on the observation, then that could be an issue.
(This is especially important because although research-grade depends on consensus in the community ID it is the observation ID of RG observations, rather than the community ID, that is shared with scientific partners like GBIF; under most circumstances these are the same, but there are a few cases where RG can be achieved for a community ID at a higher level than the observation taxon. I’m not sure why they don’t just share the community ID!)
Their solution, as far as I understand, is that subspecies IDs never change the observation ID unless they also match the community ID (that is, unless there is consensus at the level of subspecies), which works for solving that problem. Unfortunately, this also has the side-effect that adding a subspecies ID to a broadly-ID’d observation leaves it up in limbo at genus or higher until somebody else comes along and agrees.
The feature request above, to allow subspecies IDs to affect the observation ID as though they are species-level IDs until there is subspecies-level consensus, seems like a good one. But apparently that system would be difficult to implement.