fixes : rewards and training

Fix reward consistency and timeout progress

fix: rebalance reward range + unblock enrollment->monitoring->analysis phases

fix: remove duplicated curriculum block, dedup FDA table, accumulate total_reward for invalid steps

fix(N1-N6): unblock SUBMIT_TO_FDA, harden episode lifecycle, add tests

V4: fix 5 critical + 8 major + 2 moderate issues

Push 3: Curriculum controller, hidden-state pipeline, phase detector, trial judge, and full EpisodeManager wiring

Push 2: Rule-Engine, Reward Components, Episode Logging, Noise Model

fix(structure): align file layout with ARCHITECTURE.md — rules/ and root models.py