I was thinking in three steps:
- just remove the conflicts and related txs and rechunk what’s left of the old clusters, without relinearising O(n)
- combine those chunkings to produce a new diagram and compare it with the current diagram O(n\log(c)). if the diagram’s not better, throw it away, reject the RBF, all done
- if it is better; split and re-linearise the old clusters, and update the mempool O(\frac{n^2}{c})
Relinearising before comparing the diagrams would make RBFing slightly cheaper/easier in complicated cases, but I don’t think we need to do that, and deferring the O(n^2) steps until we know we’re doing something to improve the mempool seems worthwhile?