docs: publish real SFT log_history and add provenance note

docs: add final training evidence plots

fix(eval): add phase-aware hard and cascade action selection

fix(grpo): align hard and cascade training with final eval

docs(evidence): add multi-agent eval, long-horizon trace, and constrained eval logs

feat(rl): GRPO continuation from SFT adapter with multi-step task rollouts

test(reward): make reward audit run in-process when server is offline

Use constrained action scoring for model evaluation

Improve GRPO rewards and add model evaluation

feat(eval): add seed variance reporting script

feat(eval): add benchmark harness and reward integrity artifacts