Validation script uses train AUC — bounds are not meaningful for generalization

[shaypal5]

Problem

scripts/validate_v4_dataset.py fits a logistic regression on the full 1000 rows and evaluates AUC on the same 1000 rows (train AUC). The 0.65–0.90 bounds and leakage trap boost threshold (0.03) are calibrated against this inflated metric.

Train AUC is a function of feature count and sample size, not just signal strength. With 17 features (several categorical, expanding the effective dimensionality) and only 1000 rows, train AUC > 0.65 is almost guaranteed even with weak signal.

Suggested fix

Replace train AUC with k-fold cross-validated AUC (e.g. 5-fold stratified). This requires:

1. Switching _fit_lr to use cross_val_score or manual fold logic
2. Recalibrating AUC_LOWER, AUC_UPPER, and AUC_TRAP_BOOST against the new metric
3. Updating docs/v4/validation_spec.md with the new bounds

Context

Identified in self-review of PR #21. The current check still catches gross failures, so this is not blocking v4 release.

[shaypal5]

Resolved by PR #27 — the canonical validation module (leadforge/validation/lead_scoring.py) uses stratified hold-out splits (not train AUC) with properly calibrated bounds.