[Caleb Robinson (calebrob6)]

Closes #11.

PPI now normalizes weights within each correction fold and averages the fold mean vectors equally, matching deployment imputation. The delta covariance sums the within-fold covariances and divides by the squared number of models. IWCV uses the same fold weighting, and effective sample size follows those normalized weights.

Bootstrap resampling preserves every correction fold. It samples rows or nested clusters within folds; clusters that span folds remain intact, and draws that omit a correction fold are redrawn. This keeps each replicate's imputation and correction on the same fixed-model target.

Direct cross-fit calls to ppi_estimate and ppi_macro_f1 with multiple deployment models now require fold_l. audit supplies it automatically. Single-model calls retain their existing behavior.

The issue's reproduction now returns accuracy 0.50 with a 95% interval of [0.448349, 0.551651], rather than 0.38 with an interval that excludes the truth.

The regression cases cover accuracy and pooled macro-F1 with unequal fold sizes, fold-specific inclusion weights, IWCV, delta variance, and row, nested-cluster and cross-fold-cluster bootstrap intervals. All 18 tests pass on Python 3.10 and 3.13 with pinned dependencies. Ruff lint passes, and a grouped audit with real logistic trainers produces identical bootstrap results for n_jobs=1 and n_jobs=2.

The README revision is separate in #12. This PR changes only the estimator, pipeline and regression tests.