What's New in v0.5.0

Q² Blindfolding — Stone-Geisser Q² via omission loop (default D=7). Benchmarks: none <0.02, small ≥0.02, medium ≥0.15, large ≥0.35.

PLSpredict — k-fold cross-validation (default k=10) comparing model RMSE/MAE against a naive LM baseline per indicator of each endogenous LV.

CVPAT (Cross-Validated Predictive Ability Test) — one-sample t-test on per-observation loss difference (LM − model).

CMB Marker Variable Analysis — POST /cmb — Lindell & Whitney (2001) method. Flags cmb_concern = True when max |r| > 0.20.

Predictive tab in UI — Q² table, PLSpredict table, CVPAT table, CMB panel. All data downloadable via Downloads tab.

CVPAT significant: true when statistic ≈ 0 — added mean_diff > 1e-6 guard to prevent spurious significance on near-zero loss differences.