DSPy 3.2.1 BootstrapFewShotWithRandomSearch, which DSPy also exports as BootstrapRS.

Candidate seeds, baseline scheduling, chained labeled-demo sampling, tie ordering, and rounded percentage scores follow DSPy 3.2.1. Seeded draws use Imp's deterministic BEAM-native sampler rather than Python's MT19937 sequence.

Durable invocations accept :max_candidates, :checkpoint_fn, and :resume_state. A candidate is sealed only after both its bootstrap/build stage and full validation evaluation finish; interrupted candidates replay as a unit, while completed candidates do not. Captured metrics require a stable JSON-safe :metric_identity. Complete anonymous-metric runs remain available in process without a checkpoint.