So yeah, I think the general idea was that it doesn't matter much if we have consensus issues on Ropsten (we don't need to wake anybody up at 2 am). And with that reasoning, we figured we could fork it even without all tests being done.
Now, if this is not a correct sentiment, feel free to yell about it.
Propagating to all peers would lead to O(nm) messages being sent, where n is the number of nodes and m the number of peers per node - this devolves to O(n^2) in a fully connected network.
If we propagate to sqrt(m) peers, then in a fully connected network we send O(n * sqrt(n)) messages, which grows much slower.
Is a fully-connected network a desired/desirable goal? It surprised me that that it was a consideration.