Train NAIL-RAPM v1.4 Kalman Profiles

Build the assisted-shot profile mart first:

uv run nba-build-assisted-shot-taxonomy --workers 4

Train the 30-season recursive candidate with an append-only log:

uv run nba-train-nail-v14-kalman-additive-profiles \
  --log-path artifacts/logs/nail-v14-kalman-additive-profiles.log

Each bar advances only after a seasonal context state has been stored. Follow the run with:

tail -f artifacts/logs/nail-v14-kalman-additive-profiles.log

Replay the three frozen evaluation seasons after training:

uv run nba-evaluate-nail-v14-kalman-additive-profiles \
  --log-path artifacts/logs/nail-v14-kalman-additive-profiles-frozen.log

Then produce the required 10,000-draw paired-bootstrap non-promotion gate and the all-season coefficient audit:

uv run nba-bootstrap-nail-v14-kalman-additive-profiles
uv run nba-audit-nail-v14-kalman-additive-weights

The initial Kalman experiment uses a diagonal random-walk process multiplier of 1.0: the next-season prior variance equals the prior posterior variance plus one additional copy of that diagonal posterior variance. The current season's weighted residual variance calibrates the measurement update.