Train Student-t Talent-Prior Contextual RAPM

Train the combined Student-t player-prior and lagged contextual-offset model:

uv run nba-train-student-t-talent-contextual-rapm --through-season 2025-26

The command fixes the Student-t prior at \(\nu=3\) and a three-point scale, keeps Gaussian stint errors, and uses the completed Gaussian forward exposure-gated lambda schedule. Each season subtracts its prior-season context function before fitting the Student-t player state, then fits the new context function from the completed player residuals.

The trainer checkpoints after each completed season at artifacts/models/student_t_talent_contextual_rapm/<season>/.checkpoint.joblib. Re-run the same command after an interruption to resume. Use --max-seasons 3 for a deliberate short batch.

See Student-t Talent-Prior Contextual RAPM for the transition equations and frozen evaluation.