v1.1.0

Added

• TVAE-GAN promoted to supported: moved from katabatic.experimental.models.tvaegan to katabatic.models.tvaegan, with artifact persistence (load_from_ref), a tvaegan extra, CI integration tests and benchmark scripts for all five datasets. It now uses the shared six-dimension evaluate(); its reconstruction loss is available as evaluate_loss().
• TabKDE promoted to supported: katabatic.experimental.models.tabkde_updated renamed and moved to katabatic.models.tabkde, now subclassing Model with artifact persistence (load_from_ref) and CI integration tests.

Fixed

• The privacy dimension failed when a categorical column held integers in the real data and floats in the synthetic data (e.g. education-num on adult); values now compare by number, so 13 and 13.0 match.
• TabKDE: numeric class labels were returned as interpolated fractions (e.g. 0.37) by both train() and fit(x, y); labels now round-trip exactly.
• TabKDE: seed did not reach the GMM radius draw, so seeded draws were not reproducible and a fixed random_state returned identical rows on every sample() call.