Train Student-t Talent-Prior RAPM

Train the Gaussian-error, heavy-tailed-talent version of the recursive forward RAPM state through the latest completed season:

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

The first specification fixes the coefficient prior at a Student-t with three degrees of freedom and a three-point RAPM tail scale. Stint errors remain Gaussian. The trainer keeps the exact completed Gaussian forward run's per-season ridge lambda schedule, so this is a coefficient-prior ablation.

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

See Student-t Talent-Prior RAPM for the equations and current evaluation.