Thanks, just a few clarifications:
The purpose of applying a threshold like max_fee_per_gas * n > base_fee_per_gas with, e.g., n=2 is to make manipulation of the normalization step more difficult. Without such a threshold, one could submit transactions with a low max_fee_per_gas to skew the normalization process, even if those transactions are not intended to be included. The thresholding of the mempool is thus specifically for making the normalization step less arbitrary, and is only necessary under normalization. If txs are directly ranked from an equation such as
then txs that are not included will not have any influence on which txs that are included. This reduces the opportunities for the relay to manipulate the outcome (albeit, it can still nudge the registration time). The transacting user that did not have its tx included can simply review included txs to confirm that those score higher (at least the score will not depend on non-included txs, while registration time is still not objective).
And to be clear, my assumption is that the networking constraint imposed for p2p propagation in FOCIL does not really apply for the relay.
Ok, so the inclusion list should be adhered to under full blocks, potentially with a gas threshold. This follows the definition in FOCILR, which differs from FOCIL. In FOCILR, these stronger censorship resistance guarantees are imposed at the protocol level, as opposed to by a single relay. Builders can extract more value when the IL does not constrain which txs it must include. When the harder stance is taken at the protocol level, the proposer cannot avoid them. When the relay takes this harder stance, the proposer has the opportunity to opt out—the option to let the builder control the content of the full block still exists. This will probably have a larger effect on proposer rewards than the latency trade-off discussed in the introduction of the post.
I think the issue is pretty complex given that diverging ILs generally are desirable, and that the dashboard could not reveal collusion between relays. Simply put, IL aggregation in this context seems generally difficult to get right.