Because they can screw you.
They can screw any pay-per-share or any proportional pool where their hash rate is a small fraction of the total. I don't see anyone bothering.
Also stale shares are not being paid in your model, right? As otherwise the "proper" miners would finance them...
Fees can be distributed retroactively proportionally. Stale shares would not be paid as that is the miner's own fault for not using a client with long polling. All pool-side variance including unconfirmed blocks would be absorbed by the pool.
[Tycho] has said that deepbit's PPS isn't profitable unless they keep at least 7% fees. And even at 10% fees the variance causes it to not be profitable on some days. But the fact that there have to be any fees at all over time (fees above the pool's desired profit margin) indicate that there is some level of cheating the pool operator going on, otherwise things would average out and be profitable at a lower percentage.
You can't expect a variable system to produce the same amount every day. The only thing you know for sure are "expected values". If a pool has 0% fees based on "expected values" then it will break even over a long period of time. If a pool has 10% fees based on "expected values" then it will make a 10% profit over a long period of time.
It would be in [Tycho]'s interests for people to be confused about this. 7% is a lot of profit.