Validation and release engineering

The Python port uses layered validation rather than treating API-name parity as proof of scientific parity.

Layer 1 β€” public surface

The frozen R 2.3.0 namespace contains 278 exports. Every export must resolve to an introspectable Python callable, while the Python package may expose a small number of Python-specific utility names.

Layer 2 β€” semantic call contracts

R spellings, required arguments, aliases and legacy Python convenience interfaces are tested together. Where a canonical R return shape conflicts with an earlier Python convenience return shape, semantic dual dispatch preserves both rather than silently breaking one interface.

Layer 3 β€” frozen behavioral oracles

Deterministic R fixtures are flattened into typed assertions and compared against Python. Numerical tolerances are explicit. Frozen oracle files are hash checked during the completion run. R1–R3 are executed as their permanent oracle test suites. For R4, the completion runner executes the permanent canonical target/structure tests and preserves the immutable 3,786-row oracle; a full regenerated 3,786-row scorer is a distinct evidence layer and is not implied by hash verification.

Layer 4 β€” release integrity

The repository must pass linting, compilation, full tests, ≥90% coverage, documentation build, examples, wheel/sdist construction and isolated wheel installation.

Scientific interpretation

BACKEND_EQUIVALENT does not mean byte-for-byte numerical equality. For model families implemented with different statistical libraries, the contract is the analysis family, inputs, outputs, diagnostics and documented interpretation.