If my understanding of IBLT is correct, its main advantage is that the quantity of information transmitted on the wire is constant, whatever the number of transactions inside the block.
IMHO, it's fair to say that propagation of these data is not the whole story. There's a reconciliation to be done between the new block and the local mempool but I think it's also fair to make the hypothesis that with such a mechanism in place, mining pools may have an incentive to keep an exhaustive mempool (in order to avoid the burden of requesting missing transactions and to increase their performances/profits).
...
Right, the reconciliation process will take time (information must be communicated), and the amount of time will have a lower limit proportional to the Shannon Entropy in the new block. One could argue that the Shannon Entropy in a new block will not in general be proportional to the size of the block, but that doesn't make any sense to me. In fact, the only way I can imagine that it would be possible is if the block solutions contain no information about the included transactions
at all (i.e., the case of infinite coding gain). But if there's no information communicated about the transactions included in a block, then in my opinion the miners aren't peers but slaves and the system is already centralized.
If the network is composed of peers who build blocks based on their own volition, then I argue that information (Shannon Entropy) about the transactions contained within the new blocks must be communicated as part of the block solution. If this is the case, the fee market will remain healthy (according to the definition of a healthy fee market given in feemarket.pdf).