v0.5.0

New features

• EBMInsuranceWrapper: actuarial facade for InterpretML EBM with balance checks, calibration diagnostics, and relativity extraction — replaces raw EBM usage with an interface designed around insurance pricing workflows (v0.2.0)
• PostSelectionGLM: post-selection inference for Poisson GLMs, correcting for the bias introduced by variable selection before fitting
• DebiasedGLM: bias-corrected confidence intervals for Lasso-selected GLM coefficients using debiasing via the score equation
• PenalizedGLMInference: fit-once, explore-both interface combining penalised estimation with valid post-selection CIs — avoids refitting when switching between confidence interval methods
• freMTPL2 external dataset benchmark comparing EBM against GLM and GBM baselines

Improvements

• 165+ new tests across GLM inference modules, bringing the suite to comprehensive coverage of penalised and post-selection paths
• EBM docstrings rewritten: Raises sections, typed returns, and prose descriptions of actuarial assumptions
• References section added covering Hastie & Tibshirani (1990), Nori et al. (2019), and NAM literature
• Databricks-format quickstart notebook committed to notebooks/
• CI via GitHub Actions (Python 3.11+); numpy floor raised to >=2.0

Bug fixes

• Fixed EBM offset handling for Poisson exposure
• Resolved CI collection failure caused by missing torch in tests.yml
• Corrected pl.Utf8 deprecation (renamed to pl.String in Polars >=0.19)
• Bumped numpy minimum to >=1.25 for numpy.exceptions compatibility