Diverse top-N design picks: farthest-point within an objective slack
When an optimizer offers N 'options' as the N best-scoring sweep samples, they cluster in one corner of the parameter space and look identical; a per-pair 15% difference threshold on one or two params does not fix it. Instead: take the best design, force one pick from each visibly distinct sub-family (e.g. a feature toggle) if any exist, then greedily add the design maximizing min distance to the chosen set in a range-normalized, visibility-weighted parameter space, restricted to designs within a slack factor (e.g. 1.3x) of the best objective. Build the batch in a ProcessPoolExecutor (module-level worker so it pickles) and make the pool class a module attribute so tests can swap in ThreadPoolExecutor, since monkeypatched functions do not reach spawned processes.