What you're talking about is just disagreement about the content of
the memory pool

That's the same thing. Whilst you're mining your double spend tx, it's in your mempool but you don't broadcast it as per normal. Then when you find the block you broadcast it to override everyone elses mempool. So yours and theirs were inconsistent.

The only slight way BitUndo differs is, they provide it as a service, and I don't know if they inform you when they found a block (probably not), so you have to do the purchase and then hope BitUndo finds the next block. Otherwise the purchase clears. But they could certainly add a pre-notification before they broadcast to get back to the exact scheme originally described, they have everything else in place.
 
Oscar himself can be implemented as a majority M parties to further
increase confidence

This just brings us back to square one. Who are these parties and what if I pay them to be corrupt? What if they offer to be corrupt as a service?

Let's say I succeed in finding some parties who are incorruptible no matter how large of a percentage I offer them. At this point, why bother with miners at all? Why pay for double spend protection twice, once to a group of Oscar's who are trustworthy and once to a group of miners who are not?

The point of the broadcast network and mining is so there can be lots of Oscar's and I don't have to know who they are or sign up with them or put any effort into evaluating their reputation.
 
value retail transactions— the fact that any cheating by oscar is
cryptographically provable (just show them the double signatures)
maybe be strong enough alone.

But as you point out, cheating my GHash.io did not result in any obvious negative consequence to them, despite that preventing double spending is their sole task. Why would Oscar be different to GHash.io?

Trying to solve the problem of dishonest miners is effectively trying to solve the "automatically find trusted third parties" problem at scale.