Add RLlib sweeps for token-guess cycle 2

[Al-does]

Summary

• add RLlib Tune grids for learning rate, predictive-loss coefficient, and Kelly-loss coefficient
• add and run a predictive coefficient 1.0 extension for direct comparison with Kelly 1.0
• write compact summaries using final-window token accuracy and condition-specific metrics
• record completed GPU results and findings

Results

• PPO LR 1e-4: 68.8839% token accuracy
• predictive coefficient 1.0: 69.1461% (versus 0.3 at 68.6077%)
• Kelly coefficient 1.0: 66.9872%
• predictive 1.0 exceeds Kelly 1.0 by 2.1589 percentage points, but its final window also exceeds the Bayes ceiling, so multi-seed fixed-rollout validation is recommended

Testing

• uv run pytest -q -m "not slow" (68 passed, 1 skipped, 1 deselected)
• local predictive 1.0 RLlib smoke run completed
• full predictive 1.0 GPU run completed at 2,504,905 environment steps
• B2 durability verified; all cloud instances self-terminated