I have several partly-formed thoughts.
First, on "is there REALLY a problem" -- I think Ed Felten makes a good argument that, at least in the current world of mining pools, the incentives work to prevent "selfish" mining: https://freedom-to-tinker.com/blog/felten/bitcoin-isnt-so-broken-after-all/
Related, I think it is important to remember that the current structure and behavior of the p2p networking code is NOT part of the fundamental consensus algorithm. It is easy to change, and, as I've said in the past, I would really like to see blocks and transactions being broadcast over some completely different networking protocol.
I think there is a really interesting theoretical question that I don't know the answer to; I'll try to state it clearly:
As the p2p network is currently implemented, nodes only have a partial view of competing best-chains, because nodes only relay the first node they see at a given height.
If there are two blocks at the same height announced at approximately the same time, they propagate exponentially fast across the network in two waves, and only the nodes at the edges of where the waves meet will see both blocks.
Given that network behavior, assuming no selfish mining, there is a strong incentive to announce your block as soon as possible, because in an exponentially fast race any delay is very likely to make you lose.
There are lots of ideas for fixing selfish mining that rely on changing that behavior, and having nodes relay orphan blocks. I think it is important to remember that no matter what rules we SAY the network will follow, we can't stop individual nodes from implementing whatever rules they wish to follow; any solution that begins "If all nodes do THIS..." is not a solution.
That applies to the behavior we have now about not relaying orphan blocks, too, of course. The interesting theoretical question I'm pondering: does the security of the Bitcoin system depend on the relaying behavior, or on how ties are resolved in block races?
My intuition is that we SHOULD be relaying all orphan blocks, and should let each node decide how to resolve races, but before we talk about how to resolve races I think we should consider the incentives that arise from the current relaying behavior.