VBPCApy 0.5.0

VBPCApy 0.5 introduces native scikit-learn estimators and selection helpers with samples × features inputs and modern observation semantics by default. Existing 0.4 applications should follow the migration guide or explicitly import the feature-first interface from vbpca_py.legacy. Low-level pca_full keeps its feature-first layout and historical default.

• Conditional projection and optional full latent covariance for new rows, inverse reconstruction, and rowwise Gaussian marginal scoring. Joint training variational means remain in scores_; fit_transform returns conditional projections.
• Explicit constructor parameters, frozen solver options, string-compatible enums, typed result/array contracts, slotted records and documented ownership. Old native pickles/prepared CV folds require the documented migration.
• Missing-aware preprocessing and mask routing before feature expansion, strict numeric schema validation, transactional fitted state, feature names, optional pandas/Polars output and tested Pipeline/CV/GridSearchCV composition.
• Bounded covariance products, reusable sparse views, screened complete-minus-missing noise statistics and explicit endpoint-only PRMS cost diagnostics. Representative thread tuning and advisory performance evidence retain workload/hardware provenance.
• Correctness fixes preserve observed zeros and small values, validate raw sparse infinity before aggregation, retain warm-start ownership/covariance axes and return valid rank-zero native outputs with older Eigen headers.

Six standalone gallery examples execute during strict site builds, alongside a mixed-data notebook and source-collected public API/tutorial examples. Their short fits demonstrate the API; they do not qualify convergence or predictive calibration.

Python 3.11–3.14 and scikit-learn >=1.8 are required. The advertised binary matrix comprises twelve wheels: four Python versions each on Linux x86_64, macOS arm64 and Windows amd64, plus a source archive. Free-threaded Python builds are excluded. Optional extras provide plotting, pandas and Polars 2.

The 0.5 qualification record describes the required numerical, type, native allocation, actual Octave, installed-example and artifact gates. Release commit b8405f93777b37db27d63a5f2709af569c5457dc has the identical source tree to independently reviewed 7cf9358b4ff70813101bd34ee980616adea18ab2. All fifteen fresh release-head CI checks pass. The build-only qualification passes all twelve installed platform wheels, actual source-archive tests, matrix completeness and strict metadata validation. The extracted source archive passes 3,515 tests with zero skips and builds a verified strict documentation site. All thirteen downloaded artifacts match reviewed package sources; Windows line endings and the generated DLL-loading bootstrap were checked separately.

The protected production workflow completed successfully. All thirteen PyPI files match the production artifact checksums. The actual published Linux wheel and source archive were downloaded and verified. A fresh installation with sklearn 1.8.0 matched all 61 shipped source/type/header files, imported all six native modules and passed six rank-zero native regression cases, 74 public API doctests and seven documentation blocks.

Frozen scientific installations retain their registered versions and policies until their owners record an adoption decision.