Miners decide the order of transactions within a block - unless the block gets orphaned, this remains fixed. Block ordering in PoW chains is eventual, which implies the deeper a block gets embedded in the chain the higher the probability it won't be reordered. I don't know if etherium scripts start executing after a set number of confirmations, but they almost certainly should do to reduce the probability of the ordering changing.
I expected you to write that. That is why I wrote this rebuttal before you wrote your post:
Do not dismiss my point thinking that the boundary is at block confirmation. That is a naive perspective.
The issue is that since the output state of a contract in one partition could impact the input state of a contract in another partition (and there is no way to prevent this unless no one can write any user data into the block chain), then validators in each partition need to trust validators in all partitions, thus they really need to verify the scripts from all the partitions else the consensus-by-betting doesn't reflect rational incentives which the math depends on for rational (versus randomized noise) convergence, i.e. the game theory of the consensus model is impacted. But if all partitions have to validate all partitions, that destroys the scaling gains from partitions.
The point is there can't be partitions with programmable scripting. Coasian boundaries are subject to destruction by entropy.
I was talking about PoW chains - casper is a horrible idea IMO.
Okay let's replace PoS consensus-by-betting with Satoshi's PoW. So the problem remains every full node needs to validate every script (which can't scale). If we instead split validators into partitions, then they can't all agree on adding the other partitions to a block unless they've all validated all the scripts thus destroying the scaling advantage of having partitions. If we could assume the partitions can't impact each other in terms of incorrect validation impacting the output a script which impacts the input of another script in another partition, then the partitions wouldn't care to validate each other. But I showed that with scripting, it is impossible to know whether scripts in different partitions impact each other. The externalities that the block chain can't know means it can't guarantee the partitions are bounded (i.e. self-contained).
In short, if the partitions can impact each other, then all full nodes must validate/verify all scripts. Thus partitions are useless for scaling.
Whereas with the directed acyclic graphs that are crypto asset transfer, we can keep these self-contained in a partition, thus we can use partitions to do scaling.