None of all this is going to happen. Miners want small blocks and a fighting fee market. The cartel you are talking about with large blocks only sets in with such incredibly large blocks, that they are out of the question in the next few years ; if you want a mining cartel, the telephone between mining pool bosses is a much more useful device than multi-GB blocks that saturate small miner's network links and most of the user nodes.
You
had a mistake in your concept of propagation through the network. You are thinking a decentralized network is a fully connected mesh topology.
I answered that. The important mining pools (10-20 of them) should of course mesh together in a fully connected mesh topology. I took that for granted, as it is quite evident that this should emerge.
You don't need that much larger blocks to cause an amplification of propagation delay to the smallest miners in order to destroy decentralization and take 51% control.
Important miners will upgrade their mutual links from the moment that the cost of the link upgrade diminishes the time lost on mutual orphaned blocks and wasted hash rate. I don't know how much costs 6 seconds of wasted hash rate for a big mining pool, but I can imagine that that can pay for a better network link to his biggest competitor (who has even more losses during these 6 seconds, and who has also an incentive to upgrade his link to his big competitors).
Bitcoin is already centralized
Yes in that case they don't need huge blocks to destroy decentralization because it is already destroyed. In that case, they need to be able increase blocks to whatever they think is the level of transaction fees that maximizes their revenue (volume x transaction fees).
Look at the hashing distribution over pools: 5 pools have already more than 50%. 10 pools have more than 75%. For these 5 or 10 pools, it is utterly advantageous to have very good links between them, and as they don't trust each other, have a FULL mesh or an almost full mesh.
In either case, I am showing that big blocks are a cartelization paradigm. I am being thorough. Please don't fault me for being thorough, just because the network is already centralized.
I would say that every notion of rewarded mining is a centralisation paradigm, because pooling together your chances to win reduces risk -> pool formation.
For instance, with 10 minute reward lotteries, if you want to have, say, less than 10% income fluctuation during period T, you need to have at least 100 winning events in time T. If you put T equal to a week, there are 1000 blocks to win, and you'd need 10% of total hash rate to do so.
So if people want to have only 10% income fluctuation per week, ideally there are at most 10 pools with each 10% of the hash rate.
The system will evolve towards about 10 pools at most.