Module tuner 

Auto-tuner: search PipelineConfig space to maximize a QualityMetric.

This is the first usable rung of the metalearning ladder. Given a corpus and a scalar objective, the tuner enumerates or samples candidate configurations, builds a full pipeline for each, and records the score. Three strategies ship: exhaustive SearchStrategy::Grid, uniform SearchStrategy::Random, and the axis-parallel TPE-lite SearchStrategy::Bayesian acquisition — all reproducible under a fixed seed, establishing baselines for higher-order tuners (CMA-ES, meta-learning) to beat.

Projections are fit once per distinct fit-affecting hyperparameter tuple from the input corpus and reused across every trial: PCA and Kernel PCA key per kind, Laplacian per (k_neighbors, active_threshold), and UMAP per (n_neighbors, n_epochs, category_weight, min_dist) — with UMAP’s kNN graph additionally cached per n_neighbors (see TuneReport::umap_graph_builds). Only the downstream config knobs (bridge thresholds, inner-sphere gates, domain-group counts, etc.) vary per trial.

Structs

SearchSpace

Discrete candidate values for each tunable knob.

TrialRecord

One trial’s observation.

TuneReport

Full tuner output.

Enums

SearchStrategy

Which enumeration to use over the SearchSpace.

Functions

auto_tune

Run the auto-tuner and return the best pipeline plus a report.