Evaluate Frozen No-Prior Window RAPM

Build the zero-prior controls used to compare a single completed season with a pooled three-season window. Both commands fit regular-season stints only, select lambda chronologically inside their completed window, and evaluate the fixed coefficients on the entire 2025-26 regular season and playoffs.

uv run nba-evaluate-frozen-one-year-rapm --season 2025-26
uv run nba-evaluate-frozen-three-year-rapm --season 2025-26

The first command trains on 2024-25. The second trains on 2022-23 through 2024-25. Neither command accepts a player prior or uses 2025-26 outcomes for training, lambda selection, or player-coefficient fitting.

The fit artifacts are written to artifacts/models/frozen_one_year_rapm/ and artifacts/models/frozen_three_year_rapm/. Their immutable evaluation outputs are written to artifacts/models/frozen_prior_evaluation/, alongside the other frozen preseason candidates. The completed run is also recorded in MLflow.

See Frozen No-Prior Window RAPM for the model contract and Frozen Preseason Leaderboard for the common evaluation table.