The separation of collation proposal and consensus makes a lot of sense, beautiful! 
A few misc remarks:
- While I pushed back on collation proposals being necessarily state-demanding to access the log inclusion fee market, I can see how being aware of state can only help optimising utility/revenue beyond the log inclusion fee market, and this edge could become a point of centralisation.
- Outsourcing collation proposals to proposers also means offloading data availability proofs (see Phase 4 of current roadmap). Those proofs can be very heavy computationally (e.g. based on SNARKs/STARKs) so offloading them to specialised proposers seems appropriate to avoid giving computationally powerful validators an edge.
- One of the thing I liked about having precisely one proposer per period is that (unlike PoW) there is no “waste” with multiple proposers simultaneously building collations and competing. When we separate collation proposal and consensus, we reintroduce this competition with proposers trying to sell to validators the most profitable collation proposals.
- A “zen philosophy” for protocol design is “embrace whatever you don’t control”. In this case, because collation proposal and consensus can be separated via extra-protocol means, we might as well embrace a clean separation at the protocol layer.