Train Forward Decomposed Contextual RAPM

MPLCONFIGDIR=/private/tmp uv run nba-train-forward-decomposed-contextual-rapm

This command recursively fits the forward exposure-gated RAPM state and an identifiable contextual side function through 2025-26. Each season's contextual forecast is h_(t-1)(home) - h_(t-1)(away), using exactly the same player profiles and composition features as Forward Contextual RAPM. It writes an immutable run under artifacts/models/forward_decomposed_contextual_rapm/2025-26/<run_id>/.

The run retains season_context_models.joblib, context metadata, player priors, historical coefficients, frozen target-season predictions, and all frozen evaluation tables. The completed target-season context model is not used for target-season evaluation.

Run the focused contract tests with:

MPLCONFIGDIR=/private/tmp uv run pytest \
  tests/test_contextual_features.py \
  tests/test_decomposed_contextual.py \
  tests/test_forward_decomposed_contextual_rapm.py

Refresh the shared full-game table after a successful run:

MPLCONFIGDIR=/private/tmp uv run nba-report-frozen-game-outcomes --season 2025-26

The model is currently an interpretability ablation rather than the promoted predictive benchmark; see Forward Decomposed Contextual RAPM for the exact constraint and frozen comparison.