My idea had been that even if the custom plans average higher cost than the generic plan, it is still worth considering custom plans. If 1 time in 5, the custom plan is 10% of the cost of the generic plan, it doesn't matter if the average custom plan is worse, you should at least determine the cost of the custom plan before deciding whether to use it or the generic plan. Even if 80% of the time, you don't use the custom plan, the payoff in the good case may be worth it.
In theory, while ignoring planning costs, a generic plan will never outperform a custom plan (the custom plan establishes a floor any custom plan could fall back on). Thus any algorithm that requires computing the custom plan unconditionally amounts to simply setting the GUC to infinity (instead of 5).
The assumption is that at moderate to high reuse volumes it is quite probable that a generic plan will win or at least be acceptable in 999 in 1000 or more executions, not 4 in 5.
