v0.3.1

RetrainOpt 0.3.1 fixes input validation and strengthens the tests for approximate planning and online continuation.

• Reject fractional/non-finite class labels and malformed Prototype features before they can produce misleading losses. Preserve valid integer-valued labels and NumPy sample rows.
• Validate the actual model IDs and boolean observation flags returned by custom evaluators.
• Exercise real GP fitting and prediction of unobserved candidates in the BoTorch integration test.
• Verify continuation from positive and negative accumulated staleness against uninterrupted online execution, including actions, costs and model state.
• Clarify Base's label-independent decision rule, the controller's measurement requirements, and the existing approximate-search and initialization conventions.

The public planning and online APIs are unchanged. Inputs that were silently coerced or accepted incorrectly may now raise ValueError; see the migration guide.

python -m pip install --upgrade retrainopt

Validation includes the core/optional-backend matrices, minimum direct dependencies, strict documentation builds, reproducible wheel and source builds, and anonymous index installation with artifact checksum verification.

Full changelog · Changes since v0.3.0