fix: complete reward config injection for all algos and add verification tests

- Fix APPO: pass env_cfg_override and sim_backend through runner → worker → registry.make
- Fix MultiGPU offpolicy: add missing env_cfg_override to collector_kwargs
- Fix offpolicy play mode: extract and pass env_cfg_override
- Fix MLX PPO: add reward extraction in both train and play paths
- Add reward: default to PPO and MLX PPO config defaults
- Add sim_backend to APPO config
- Add PPO reward configs for Go1/Go2/G1 × mujoco/motrix (5 new files)
- Add PPO task reward overrides for all 3 locomotion robots
- Add 67 config system verification tests covering Hydra composition,
  reward injection end-to-end, YAML↔dataclass consistency, and extraction logic

Author: TATP-233

README.md

@@ -133,10 +133,13 @@ UniLab 支持两种仿真后端：
 
 ```bash
 # 训练
+uv run python scripts/train_rsl_rl.py task=go1_joystick
+
+# 指定motrix后端
 uv run python scripts/train_rsl_rl.py task=go1_joystick training.sim_backend=motrix
 
 # 回放（交互式可视化）
-uv run python scripts/train_rsl_rl.py task=go1_joystick training.sim_backend=motrix training.play_only=true
+uv run python scripts/train_rsl_rl.py task=go1_joystick training.play_only=true
 ```
 
 ## 训练与回放指南

conf/appo/config.yaml

@@ -50,6 +50,7 @@ training:
   device: null
   collector_device: null
   logger: tensorboard
+  sim_backend: mujoco
   log_dir: null
   play_only: false
   no_play: false

conf/ppo/config.yaml

@@ -1,5 +1,6 @@
 defaults:
   - _self_
+  - reward: default
   - task: go1_joystick
 
 algo:

conf/ppo/config_mlx.yaml

@@ -1,5 +1,6 @@
 defaults:
   - _self_
+  - reward: default
   - task: go1_joystick
 
 algo:

conf/ppo/reward/g1_ppo_motrix.yaml

@@ -0,0 +1,19 @@
+# @package _global_
+reward:
+  scales:
+    tracking_lin_vel: 2.4
+    tracking_ang_vel: 0.3
+    feet_phase: 1.2
+    lin_vel_z: -1.2
+    ang_vel_xy: -0.3
+    base_height: -600.0
+    orientation: -6.0
+    action_rate: -0.012
+    pose: -0.12
+  tracking_sigma: 0.25
+  gait_frequency: 1.5
+  feet_phase_swing_height: 0.09
+  feet_phase_tracking_sigma: 0.008
+  base_height_target: 0.754
+  min_base_height: 0.55
+  max_tilt_deg: 25.0

conf/ppo/reward/g1_ppo_mujoco.yaml

@@ -0,0 +1,19 @@
+# @package _global_
+reward:
+  scales:
+    tracking_lin_vel: 2.0
+    tracking_ang_vel: 0.2
+    feet_phase: 1.0
+    lin_vel_z: -1.0
+    ang_vel_xy: -0.25
+    base_height: -500.0
+    orientation: -5.0
+    action_rate: -0.01
+    pose: -0.1
+  tracking_sigma: 0.25
+  gait_frequency: 1.5
+  feet_phase_swing_height: 0.09
+  feet_phase_tracking_sigma: 0.008
+  base_height_target: 0.754
+  min_base_height: 0.55
+  max_tilt_deg: 25.0

conf/ppo/reward/go1_ppo_motrix.yaml

@@ -0,0 +1,13 @@
+# @package _global_
+reward:
+  scales:
+    tracking_lin_vel: 1.2
+    tracking_ang_vel: 0.3
+    lin_vel_z: -6.0
+    ang_vel_xy: -0.15
+    base_height: -120.0
+    action_rate: -0.006
+    similar_to_default: -0.12
+    contact: 0.28
+  tracking_sigma: 0.25
+  base_height_target: 0.3

conf/ppo/reward/go2_ppo_motrix.yaml

@@ -0,0 +1,17 @@
+# @package _global_
+reward:
+  scales:
+    tracking_lin_vel: 1.2
+    tracking_ang_vel: 0.3
+    lin_vel_z: -6.0
+    ang_vel_xy: -0.03
+    base_height: -120.0
+    action_rate: -0.006
+    similar_to_default: -0.12
+    alive: 0.0
+    foot_lift_reward: 0.24
+    foot_drag_penalty: -0.0
+  tracking_sigma: 0.25
+  base_height_target: 0.3
+  target_foot_height: 0.08
+  foot_clearance_sigma: 0.02

conf/ppo/reward/go2_ppo_mujoco.yaml

@@ -0,0 +1,17 @@
+# @package _global_
+reward:
+  scales:
+    tracking_lin_vel: 1.0
+    tracking_ang_vel: 0.2
+    lin_vel_z: -5.0
+    ang_vel_xy: -0.02
+    base_height: -100.0
+    action_rate: -0.005
+    similar_to_default: -0.1
+    alive: 0.0
+    foot_lift_reward: 0.2
+    foot_drag_penalty: -0.0
+  tracking_sigma: 0.25
+  base_height_target: 0.3
+  target_foot_height: 0.08
+  foot_clearance_sigma: 0.02

conf/ppo/task/g1_joystick.yaml

@@ -1,4 +1,7 @@
 # @package _global_
+defaults:
+  - override /reward: g1_ppo_mujoco
+
 training:
   task_name: G1JoystickFlatTerrain
 algo: