Fix oversight reward: expose oversight_reward on IncidentObservation

Save full training logs to JSON + fix oversight obs.metadata crash

Reduce to 4 gens/200 tokens for ~60s/step on T4 (120 steps ≈ 2h)

Print env crash traceback in reward fn to diagnose -0.3 path

Add reward diagnostic logging to identify why rewards are flat

Tune GRPO for T4 speed: 4 gens, 256 tokens, temp 1.2

Fix GRPO rollout collapse: temperature 0.7→1.0, generations 4→8, completion length 300→512

Citadel v2.0.0 — Multi-Agent AI Defense Council