TOML-based per-model training profiles replacing hardcoded hyperparameters. 8 config files (DQN/PPO/supervised × production/smoketest + hyperopt + walk-forward). 3-tier loading: env var > filesystem > embedded defaults. Full CLI coverage for all profile fields. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>