feat(strategies): BatchLexicase and UnprunedPareto candidate selectors

[Benzhang2004]

Motivation

On binary metrics with large candidate pools, ParetoCandidateSelector's dominance-pruning step (remove_dominated_programs) can assign zero selection probability to the highest-aggregate candidates:

• A sole solver of any instance is unremovable, regardless of quality: its front contains no other candidate, so is_dominated can never return true for it. With single-rollout evaluation, sole-solver slots are frequently minted by noise.
• A broadly strong candidate typically holds no unique slot (everything it solves, someone else also solves), so it gets pruned as "redundant" and receives zero mass.

Replaying the deployed sampler on nine 15,000-metric-call LiveBench-Math optimization runs (~126-candidate pools, binary scoring): the best-val candidate received exactly zero selection mass in five of nine runs, and lineage analysis showed the resulting search trees grow as stars around a static set of early sole-solver hubs (69-75% of accepted candidates never breed) with max depth at or below what uniform random parent attachment would produce. On a continuous-metric task (deterministic simulator scores, mostly singleton fronts), the same code behaves well: champions keep 2-3x uniform mass. The pathology is specific to large tie sets, consistent with dominance resistance in many-objective optimization (Purshouse & Fleming 2007).

This does not contradict the GEPA paper's Table 3 ablation (Pareto sampling >> SelectBestCandidate): those experiments ran 1.8k-7k rollouts, where pools are small and unique slots are meaningful. The failure emerges at larger pool sizes.

Changes

• BatchLexicaseCandidateSelector(batch_size=8, rng) — batch lexicase selection (Helmuth, Spector & Matheson 2015; Aenugu & Spector 2019, whose batched variant targets noisy binary case scores). Validation instances are shuffled and grouped into batches; the pool is filtered batch-by-batch keeping max-batch-score survivors until one remains. Every candidate wins under some ordering, so no candidate is ever excluded outright; batch_size tunes selection pressure continuously between plain lexicase (1) and argmax-by-aggregate (whole valset). In counterfactual replays on the nine saved runs, batch lexicase at b=8-16 raised the expected val score of the selected parent in every run while keeping 27-66 effective parents (diversity preserved).
• UnprunedParetoCandidateSelector(rng) — frontier-frequency sampling with the dominance-pruning step skipped. Ablation control to isolate the pruning step's contribution.
• Both registered as candidate_selection_strategy strings ("batch_lexicase", "unpruned_pareto") in api.py and gepa_launcher.py.
• Regression test reproducing the pathology in miniature: a best-aggregate champion whose every win is shared with an irremovable sole-solver specialist gets zero mass from ParetoCandidateSelector across 200 draws and majority mass under batch lexicase.

Status

[semanticdiff-com]

Review changes with  

Changed Files

File	Status
src/gepa/gepa_launcher.py	55% smaller
src/gepa/api.py	51% smaller
src/gepa/strategies/acceptance.py	0% smaller
src/gepa/strategies/candidate_selector.py	0% smaller
tests/test_candidate_selector.py	0% smaller