select
Selects a candidate probabilistically from the Pareto frontier.
Weight is proportional to the number of instances where the candidate is the best performer. This encourages selecting underexplored candidates that cover unique instance niches. Falls back to uniform random if no instance-level data exists.