Audit Portable and Relative Context Agreement

This no-refit audit compares the completed bounded portable-matchup and bounded relative-context fits on the same observed 2025-26 regular-season RAPM stints. It separates the net player-rating edge from the net contextual edge and weights every summary by stint possessions.

uv run --group docs nba-build-portable-relative-context-agreement-audit

The command writes immutable paired predictions and agreement metrics under artifacts/analysis/portable_relative_context_agreement/2025-26/, updates the latest pointer, renders aggregate and feature-level correlation figures, and refreshes the audit section of Forward Bounded Hierarchical Portable-Matchup Contextual RAPM.

It is a model-agreement diagnostic, not a predictive leaderboard evaluation: both models have already fit the 2025-26 season before scoring its observed lineup matchups.

Feature-level results compare each model's total attribution for the same original contextual feature. In the portable model that is its composition contribution plus its matchup-residual contribution; it is not a comparison of the portable-only composition score, which has no counterpart in the relative model.