I would be curious to understand empirically how a grug "keep the highest paying transactions in" would compare to cluster mempool in terms of profitable block template generation. Again, with a significant amount of ram available.

Me too.
Greedy algorithms usually do extremely well. Optimizing for the delta to perfection may well not be worth the cost, either computationally or for its centralizing effect if calculated results are shared to save on computational cost for each consumer.