@hackage/shikumi-optimize

Widened the generic-lens bound to >=2.2 && <2.4 (admits 2.3) and upgraded the test suite's baikai dependency to the 0.4 series (>=0.4 && <0.5). Dependency bounds only — no changes to the exported API.

Shared optimizer budget metering helpers, effective-instruction inspection, and exact LLM-call counting for opaque optimizer runs.
Budgeted ensemble search via ensembleSearchWith and exported bootstrapKeptDemos for metered bootstrap demo recovery.

Optimizers now reserve predicted LM-call cost before scoring/proposing and return the best candidate found so far when the next spend would exceed budget.
Instruction search and seeding preserve effective instructions without serializing redundant overrides.
Ensemble and KNN documentation now calls out structure-changing artifacts and how to load saved state against matching templates.
Refreshed internal shikumi, shikumi-compile, shikumi-eval, and shikumi-trace bounds for the current package set.

Renamed exported Candidate record selectors to params, perExample, and aggregate.
Renamed exported PastInstruction record selectors to instruction and score.
Refreshed internal shikumi, shikumi-compile, shikumi-eval, and shikumi-trace bounds for the current package set.

Constrained the test suite's baikai dependency to the 0.2 series and refreshed internal shikumi package bounds.

Initial Hackage release of optimizer APIs for shikumi programs.
Labeled few-shot, bootstrap few-shot, instruction search, random search, GEPA, KNN, COPRO, MIPRO, ensemble, Pareto, and proposal helpers.