It doesn't matter, except that nodes with a too-small limit will frequently have the latest blocks "on probation" waiting for them to reach an acceptance depth. A node which sees a block which is excessive (over its limit) will not relay or process it until it is buried under enough confirmations. BU nodes always track the chain tip with the most PoW.
So it sounds like you are using non-relaying by end user nodes as a form of soft pressure to discourage miners from creating blocks larger than node limits.
How do you deal with the fact that miners will often want to have direct connections with other miners (due to the natural incentive to reduce "orphans"), making relay through end user nodes irrelevant?
Setting aside the previous, have you done simulations or mathematical models to characterize how relaying will behave in practice given some distribution of nodes putting a block "on probation" while other nodes process and relay it?
Can you characterize the cost of creating nodes that deliberately relay larger blocks than most of the end user nodes in an attempt to influence the orphan pressure on miners?
Without doing any simulations or models, it intuitively seems to me that even a small number of such permissive nodes (either malicious or otherwise) can quickly relay blocks to nearly the entire portion of the network willing to process them, though operators of such nodes will incur bandwidth costs.
A miner producing a block larger than emergent network consensus gets its block orphaned.
Only if the block does not reach other miners.