Train Forward Contextual RAPM

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

This command recursively fits player RAPM and contextual states through 2025-26. It uses the published season-specific forward-RAPM lambda schedule as a fixed control and carries each completed contextual model forward as the next season's lineup-level offset. The output is written to artifacts/models/forward_contextual_rapm/2025-26/<run_id>/.

The regression suite includes a synthetic two-season leakage test. It asserts that the frozen 2025-26 evaluator receives g_2024-25, while g_2025-26 is only retained as the completed state for the next forecast season:

MPLCONFIGDIR=/private/tmp uv run pytest tests/test_forward_contextual_rapm.py

Publish The Next-Season Rankings

After a forward contextual run completes, create the immutable completed-state ranking artifact and refresh the sortable top-100 table on the model page:

MPLCONFIGDIR=/private/tmp uv run nba-build-forward-contextual-rankings

For the current run, these are 2026-27 player priors from the completed 2025-26 additive state. The corresponding g_2025-26 remains a lineup-level term rather than a player ranking column.

Publish The Frozen Lineup Context Case Study

This publishes established 2025-26 five-man units scored only with the frozen g_2024-25 contextual state. It writes an immutable artifact and regenerates the positive and negative examples on the model page:

MPLCONFIGDIR=/private/tmp uv run nba-build-forward-contextual-case-study

The generated case study compares the frozen g_(t-1) forecast with the completed additive and contextual state for the target season. The completed columns are retrospective, in-sample explanations and are not leaderboard evaluation metrics.

The default eligibility rule is at least 250 shared possessions and 20 games. The standardized score uses a possession-weighted distribution of established opponent units and is not an intrinsic player or lineup value.

The artifact also contains top_lineup_attribution.parquet and response_curves.parquet. The latter records the frozen, orientation- symmetrized spline responses for relative usage and defensive rebounding; the command publishes its SVG to docs/assets/images/forward-contextual-rapm/context-response-curves.svg.