scripts/benchmark/rl/runtime.py:490

• evaluate_checkpoint() binds an evaluation rollout buffer via env.set_rollout_buffer(), but the new evaluate_episodes() loop does not populate policy-side fields (notably previous actions) into that buffer. This can corrupt rewards/metrics for environments whose managers read prev actions from env.rollout_buffer (e.g. action_smoothness_penalty reads buffer["action"]). Prefer not binding a rollout buffer in this evaluation path unless the evaluator also writes the buffer fields each step.
  scripts/benchmark/rl/runtime.py:505
• The on_step callback currently only compacts eval_rollout_buffer, but never writes action/reward/done into the buffer. After removing rollout-buffer binding for evaluation, this callback can be dropped and evaluate_episodes() called without on_step.
  embodichain/learning/rl/train.py:98
• _build_learning_policy() computes policy_name but never uses it. This is dead code and can trigger lint/type-check warnings; remove it (or use it for validation) to keep the entry point clean.

    obs_dim = int(env.single_observation_space.shape[-1])
    action_dim = int(env.single_action_space.shape[-1])
    policy_name = policy_block["name"].lower()
    actor_cfg = policy_block.get("actor")
    critic_cfg = policy_block.get("critic")

embodichain_tasks/embodichain_tasks/rl/basic/point_mass.py:33

• PointMassEnv is newly introduced, but it registers with override=True. This silently replaces any existing registration for "PointMassRL" and can mask accidental duplicate IDs during imports. Prefer the default (override=False) and rely on the registry's nested-package shadow detection for editable installs.