Reflective prompt evolution with Pareto selection and merge.
Each predictor instruction is a component of the search in
Tune.Optimizers.Reflective: parents are chosen from the per-row
validation Pareto frontier, the proposal role reflects on training
trajectories with scores and metric feedback (score-only metrics work too),
children must beat their parent on the same training minibatch before a
full validation pass, and complementary children with a common ancestor
are merged. Lineage and per-row frontier data are kept in the report.
Options: :max_iterations (default 8), :minibatch_size (default 3),
:use_merge (default true), :merge_every (default 3), and
:predictors (names to evolve; default all).
Differences from the gepa package: budgets are managed model calls rather
than metric calls; there is no automatic budget preset, experiment
tracking, or multi-objective metric framework.