feat: Module C v5/v6 training complete, ablations, SOTA baselines, paper updates

- Module C: BC+PPO training v5/v6 done; eval results in experiments/eval_intervention_v{5,6}.json
- Reward: v5 label-aligned constrained reward (code/src/rl/reward.py)
- Ablations: Module B (history_r, response_only, full) + Module C (wo_category_reward)
- SOTA baselines: WildGuard and ShieldGemma2b eval scripts and results
- Paper: update sections 05–08 (Module B/C description, experiments table, discussion)
- Docs: add record.md (change log), update state.md and exp.md; retire change.md
- Tools: add html-to-ppt utilities and run_shieldgemma2b.sh
- Configs: add ablation YAML configs for Module B and C
- Cleanup: remove stale reference/ PNG screenshots

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

code/scripts/train_intervention.py

@@ -295,8 +295,10 @@ def main():
             torch.distributed.broadcast(obs_tensor, src=0)
             torch.distributed.broadcast(action_tensor, src=0)
 
-        obs_tensor = obs_tensor.to(accelerator.device)
-        action_tensor = action_tensor.to(accelerator.device)
+        # Keep tensors on CPU: DataLoader(pin_memory=True) requires CPU tensors.
+        # accelerator.prepare() moves batches to the correct device during training.
+        obs_tensor = obs_tensor.cpu()
+        action_tensor = action_tensor.cpu()
 
         agent = InterventionAgent(
             detector_hidden=detector_hidden,
@@ -355,6 +357,7 @@ def main():
             detector_hidden=detector_hidden,
             reward_weights=cfg.get("reward"),
             max_turns=env_cfg.get("max_turns", 20),
+            enable_category_reward=cfg.get("reward", {}).get("enable_category_reward", True),
         )
 
         output_cfg = cfg["output"]