(Putting on my devil’s advocate hat on 
.)
I can’t get incentives to align. Before going into details I want to propose a basic model for proposer revenue T-F versus F:
- Marginality: Because of proposer open competition we should expect small profit margins (and small revenue T-F) for proposers. The more proposers join in, the smaller the profit margins. This is in contrast to F which derives from an open auction where the more users means a higher F. In short, in a healthy market with many proposers and many users, T-F should be marginal versus F.
- Unpredictability: Proposer open competition should also yield steady (non-volatile) profit margins. This is in contrast to total transaction fees T being wildly volatile and hard to predict (as we know empirically). In short, the ratio between T-F and F changes widely and is hard to predict.
We now go back incentive alignment. In general, if the design assumes validators do not have access to the collation body before posting the collation header to the VMC (in contrast to the commitment-and-trapping scheme) then either:
- Proposers can grief validators, or
- Validators get full payment (proposer fees and collation subsidy) for collations that contribute to security, even for unavailable collations.
Proposers having the option to grief validators seems bad. The marginality of T-F versus F (or versus F+S, where S is the collation subsidy) means there is a high incentive to grief. In turn, this encourages validators to do “credit scoring” on proposers (see a similar argument here). The greater the demand for transaction fees the higher the incentive to grief, which in turn increases the need for validators to do credit scoring on proposers.
The other option is to give validators full payment, but this seems to introduce other bad things. In the diagram above the proposers for two of the three collations with a virtual score score of 4 (namely the leftmost and rightmost collations) would have to forgo F without any reward. The huge discrepancy between the potential reward (T-F) and the potential penalty (F), combined with the unpredictability of F, may make it hard for proposers to effectively run a business without taking huge risks.
This also opens the possibility for proposers to get griefed by other proposers, or by validators. Notice for example that because validators are guaranteed payment they can grief proposers for free, i.e. the griefing factor is infinite. In general I also dislike the idea of paying validators for unavailable collations because the clean separation between availability (enforced by validators) and validity (enforced by proposers) gets muddied.