In comparison to the 25BTC coinbase, all transactions are essentially 'free' - so saying it comes down to orphan risk; it's more complicated than that.

If it came down to minimizing orphan risk due to propagation delay, all miners would mine empty blocks. While we do still see that occasionally, miners also realize that for bitcoin to grow and prosper, and for their horde to increase in value, they need to do a good job actually including the majority of pending transactions in their blocks.

So I think it's more a result of miners "doing what they think they are expected to do" - and also to some extent miners running the software they are given to run. The result is that 0 fee transactions get delayed, because miners believe they're fulfilling their end of the bargain. There's a lot of game theory that keeps the system running as smoothly as it does.

The core team has done a lot of research into block size and propagation delay, and if I understand it correctly, the problem with larger blocks is it will ultimately lead to centralization as mining pools end up co-location and anyone outside the data center gets put at a disadvantage.