PERMANENCE: reversibility-aware RL environment for training LLM agents

Run 6.1: env precondition fix (destructive DB ops on missing tables short-circuit)

Run 8: disable unlikeliness shaping (β_rank=0.0) — fix classification-task reward inversion that collapsed Run 7 accuracy to 46%

Run 7: R4/R5 calibration traces — fix git context conflation + destructive DB R4 distinction

Run 6.1: env precondition fix (destructive DB ops on missing tables short-circuit)

Run 6: forced variants (eps 50%→70%), β_rank=0.25, R-level bonus, μ=2 PPO epochs, balanced R1-R5 warmup traces

Run 4: tech-only curriculum, 3B model, integrated deploy task

Run 4: tech-only curriculum, 3B model, integrated deploy task

Sync: new curriculum, 6 tasks, composable rubric, latent dynamics