[P2] Reward function + rollout MDP

[duckyquang]

Phase: P2
Owner(s): Davin

---

Goal: Define the RL reward and rollout interface Quang's GRPO imports.

Where: marc/train/reward.py, marc/train/rollout.py, tests/test_reward.py.

Parallel track: Day 1 of P2. Build + test with fake trajectories before GRPO trainer exists.

1. Implement compute_reward(trajectory, G, checker, cas) → float.

2. Implement run_rollout(policy, G, K) → trajectory wrapping solve steps.

3. Unit test: fake trajectory with energy decreasing → positive shaping reward.

4. Unit test: checker accept → terminal reward fires.

5. Document API in module docstring for Quang.

Done when: pytest tests/test_reward.py passes; Quang imports without changes.

Push to: PR davin/p2-reward → main (merge before or same day as GRPO).

---

Auto-created from the MARC task board (index.html). Update the board if scope changes.