docs: document reward logging contract

docs/sphinx/source/en/2-user_guide/1-training/3-logging.md

@@ -6,9 +6,10 @@ refreshes on a fixed 2 Hz clock and shows a two-second sliding average;
 `training.logger=tensorboard` (the default) or `training.logger=wandb` selects the
 persistent backend, which keeps each unsmoothed iteration value.
 
-This page first covers the log directory shared by all algorithms, then documents the
-off-policy terminal used by SAC / TD3 / FlashSAC and APPO. Every terminal field in the
-tables maps directly to one backend key; an `_ms` suffix always means milliseconds.
+This page first covers the log directory shared by all algorithms and the
+Manager-Based reward metric contract, then documents the off-policy terminal used by
+SAC / TD3 / FlashSAC and APPO. Every terminal field in the tables maps directly to one
+backend key; an `_ms` suffix always means milliseconds.
 
 ## Log Directory and Backend
 
@@ -47,6 +48,24 @@ Shared fields include `training.wandb_project`, `training.wandb_entity`,
 `run_config.json` and `run_summary.json`; RSL-RL PPO also connects its writer in W&B
 mode. A MuJoCo `play_video.mp4` is uploaded when the run produces one.
 
+## Manager-Based Reward Metrics
+
+Manager-Based environments place structured metrics in `info["log"]`. The RSL-RL PPO
+logger forwards slash-prefixed entries in that dictionary to TensorBoard / W&B and also
+writes its own native aggregate training metrics. The reward-related keys have the
+following contract:
+
+| Key | Owner | Meaning |
+| --- | --- | --- |
+| `reward/<term>` | UniLab `RewardManager` | Per-term diagnostic for the current reward computation: the mean across environments of `raw_value * weight`, before `dt` scaling |
+| `Train/mean_reward` | Native RSL-RL PPO logger | Aggregate mean completed-episode return; use this for overall training progress |
+| `Train/mean_episode_length` | Native RSL-RL PPO logger | Aggregate mean completed-episode length |
+| `Episode_Termination/<term>` | UniLab `TerminationManager` | Count of reset environments for which that termination term fired; this is a distinct diagnostic and is not a reward |
+| `Episode_Reward/<term>` | Removed | Historical normalized cumulative per-term value; it is intentionally no longer emitted because it duplicated `reward/<term>` with a different meaning |
+
+Historical TensorBoard event files are immutable, so runs made before this change can
+still display `Episode_Reward/<term>`. New Manager-Based runs should not emit it.
+
 ## Off-policy Terminal
 
 The bottom of the terminal has three columns:

docs/sphinx/source/zh_CN/2-user_guide/1-training/3-logging.md

@@ -5,8 +5,9 @@ training logger 提交一次指标。终端面板按固定 2 Hz 时钟刷新，
 `training.logger=tensorboard`（默认）或 `training.logger=wandb` 决定持久化 backend，
 backend 保留每个 iteration 未经时间平滑的值。
 
-本文先说明所有算法共用的日志目录，再详细说明 SAC / TD3 / FlashSAC 与 APPO 共用的
-off-policy 终端视图。表中的“终端字段”与 backend key 一一对应；后缀 `_ms` 均为毫秒。
+本文先说明所有算法共用的日志目录和 Manager-Based reward 指标契约，再详细说明
+SAC / TD3 / FlashSAC 与 APPO 共用的 off-policy 终端视图。表中的“终端字段”与
+backend key 一一对应；后缀 `_ms` 均为毫秒。
 
 ## 日志目录与 backend
 
@@ -45,6 +46,23 @@ uv run train --algo ppo --task go2_joystick_flat --sim mujoco \
 `run_config.json` 与 `run_summary.json`；RSL-RL PPO 在 W&B 模式下还会接管
 RSL-RL writer。MuJoCo run 若产生 `play_video.mp4`，该视频会上传至 W&B。
 
+## Manager-Based reward 指标
+
+Manager-Based 环境把结构化指标放在 `info["log"]` 中。RSL-RL PPO logger 会把该字典里
+带斜杠的条目转发到 TensorBoard / W&B，同时写出自己原生的聚合训练指标。reward 相关
+key 的契约如下：
+
+| Key | 归属 | 含义 |
+| --- | --- | --- |
+| `reward/<term>` | UniLab `RewardManager` | 当前 reward 计算的分项诊断：`raw_value * weight` 在所有 env 上的平均值，不经过 `dt` 缩放 |
+| `Train/mean_reward` | RSL-RL PPO 原生 logger | 已完成 episode return 的聚合平均值；查看整体训练进度使用它 |
+| `Train/mean_episode_length` | RSL-RL PPO 原生 logger | 已完成 episode 长度的聚合平均值 |
+| `Episode_Termination/<term>` | UniLab `TerminationManager` | 本次 reset 的 env 中触发该 termination term 的数量；这是独立诊断，不是 reward |
+| `Episode_Reward/<term>` | 已移除 | 历史版本中的按 term 归一化累计值；它与 `reward/<term>` 语义不同却形成重复树，因此被有意移除 |
+
+历史 TensorBoard event 文件不可重写，因此变更前的旧 run 仍可能显示
+`Episode_Reward/<term>`；新的 Manager-Based run 不应再产生它。
+
 ## Off-policy 终端视图
 
 终端底部并排显示三列：