Unfortunately this PR doesn't actually improve performance. cea215c appears to be the culprit, and with some tweaks to allow more fails it can do better than #6421 but not so much better that I think this is the right approach.
Some very high fee rate transactions up to 400k satoshis/kB get rejected by both pulls. Once you start cleaning stuff out from the mempool, there is a tendency for the bottom to become occupied by large transactions with very low fee rate. Small transactions then have no chance of evicting these even with much higher fee rates because they fail the total fee test. I believe some combination of one of these approaches is necessary:

1. Allowing high fee rate transactions to exceed the mempool cap via a multiplier based on size.
2. Aggregating the fees of small high fee rate transactions to collectively evict large low fee rate transactions
3. Attempting to evict by lowest total fee transactions as well as lowest fee rate.
4. Using a packaging approach which combines dependent chains of transactions for a total feerate (or fee) to sort on.

Closing until something is actually materially better.