chore: prune redundant tests and remove legacy task-factory seam

docs/sphinx/source/en/2-user_guide/4-tasks/2-motion_tracking.md

@@ -65,8 +65,9 @@ actor keeps the command and anchor-orientation terms at one step while the
 `base_ang_vel`, `joint_pos`, `joint_vel`, and `actions` terms declare
 `history_length: 5`. `ObservationManager` owns and flattens those per-term
 histories; the actor uses the configured encoder-biased joint-position term while
-the critic keeps the clean term. Per-term oldest-first ordering is guarded by
-`tests/scripts/test_obs_alignment_g1_wbt.py`; the hardware-side contract is
+the critic keeps the clean term. Per-term oldest-first ordering is guaranteed by
+the `ObservationManager` per-term history buffers
+(`tests/managers/test_observation_buffers_noise.py`); the hardware-side contract is
 documented in the sim-to-real deployment guide. When a Motrix sim2sim replay needs
 a checkpoint from another log root, pass the absolute path through `uv run eval`:
 

docs/sphinx/source/en/3-deployment/1-sim_to_real/2-g1_whole_body.md

@@ -97,7 +97,8 @@ within the term, and terms are concatenated in declaration order:
 
 The motion command contributes the reference joint position and velocity
 (`29 + 29`) ahead of the observation terms. Per-term oldest-first ordering is
-guarded by `tests/scripts/test_obs_alignment_g1_wbt.py`; mirror that ordering
+guaranteed by the `ObservationManager` per-term history buffers
+(`tests/managers/test_observation_buffers_noise.py`); mirror that ordering
 on hardware or the policy reads a permuted vector.
 
 ## 3. Actuator interface

docs/sphinx/source/en/3-deployment/1-sim_to_real/8-latency_budget.md

@@ -10,7 +10,7 @@ budgets as robot-specific measurements, not UniLab defaults.
 | --- | --- | --- |
 | One-step action delay | Manager action term `simulate_action_latency` declarations in task owners | Executes the previous action instead of the current action. |
 | G1 WBT observation history | Per-term `history_length` in `src/unilab/conf/sac/task/g1_wbt_obs/mujoco.yaml` | Per-term history for `base_ang_vel`, `joint_pos`, `joint_vel`, and `actions`. |
-| Obs history ordering guard | `tests/scripts/test_obs_alignment_g1_wbt.py` | Asserts per-term oldest-first flatten for the G1 WBT actor obs. |
+| Obs history ordering | `ObservationManager` per-term history buffers (`tests/managers/test_observation_buffers_noise.py`) | Per-term oldest-first flatten for the G1 WBT actor obs. |
 
 ## Action Latency
 

docs/sphinx/source/zh_CN/2-user_guide/4-tasks/2-motion_tracking.md

@@ -61,7 +61,8 @@ uv run train --algo sac --task g1_wbt_obs --sim mujoco training.use_amp=true
 orientation term 保持单步，`base_ang_vel`、`joint_pos`、`joint_vel` 和 `actions` term
 分别声明 `history_length: 5`。这些逐项历史由 `ObservationManager` 维护并展开；actor
 使用配置中的 encoder-biased joint-position term，critic 则保留 clean term。逐项最旧
-优先顺序由 `tests/scripts/test_obs_alignment_g1_wbt.py` 守护；硬件侧契约见仿真到真机
+优先顺序由 `ObservationManager` 的逐项历史缓冲实现保证
+（`tests/managers/test_observation_buffers_noise.py`）；硬件侧契约见仿真到真机
 部署指南。当 Motrix sim2sim 回放需要引用其他日志根目录下的 checkpoint 时，用
 `uv run eval` 透传绝对路径：
 

docs/sphinx/source/zh_CN/3-deployment/1-sim_to_real/2-g1_whole_body.md

@@ -91,7 +91,8 @@ Actor 观测宽度是 `env.observations.actor.terms` 下各项 `dim * history_le
 ```
 
 motion command 在观测项之前贡献参考关节位置与速度（`29 + 29`）。逐项的最旧优先
-顺序由 `tests/scripts/test_obs_alignment_g1_wbt.py` 守护；硬件侧必须镜像该顺序，
+顺序由 `ObservationManager` 的逐项历史缓冲实现保证
+（`tests/managers/test_observation_buffers_noise.py`）；硬件侧必须镜像该顺序，
 否则策略读到的是被置换过的向量。
 
 ## 3. 执行器接口

docs/sphinx/source/zh_CN/3-deployment/1-sim_to_real/8-latency_budget.md

@@ -9,7 +9,7 @@
 | --- | --- | --- |
 | 单步动作延迟 | task owner 中 Manager action term 的 `simulate_action_latency` 声明 | 执行上一步动作而非当前动作。 |
 | G1 WBT 观测历史 | `src/unilab/conf/sac/task/g1_wbt_obs/mujoco.yaml` 中逐 term 的 `history_length` | 为 `base_ang_vel`、`joint_pos`、`joint_vel` 与 `actions` 提供逐项历史。 |
-| 观测历史顺序守护 | `tests/scripts/test_obs_alignment_g1_wbt.py` | 断言 G1 WBT actor 观测按逐项最旧优先展平。 |
+| 观测历史顺序 | `ObservationManager` 逐项历史缓冲（`tests/managers/test_observation_buffers_noise.py`） | G1 WBT actor 观测按逐项最旧优先展平。 |
 
 ## 动作延迟
 

src/unilab/tasks/migration_matrix.py

@@ -11,7 +11,7 @@
 from typing import Literal
 
 MigrationStatus = Literal["Compatible", "Adapted"]
-MigrationTarget = Literal["complete", "mba", "compatibility"]
+MigrationTarget = Literal["complete", "mba"]
 
 
 @dataclass(frozen=True)

tests/algos/test_appo_runner.py

@@ -26,23 +26,6 @@
 from unilab.structured_configs import APPOConfig
 
 
-@pytest.mark.slow
-def test_appo_runner_init_no_crash(mock_env_name):
-    cfg = APPOConfig().to_dict()
-    cfg["num_envs"] = 4
-    cfg["steps_per_env"] = 4
-
-    runner = APPORunner(
-        env_name=mock_env_name,
-        env_factory=registry_env_factory(mock_env_name, "mujoco"),
-        env_cfg_overrides={},
-        rl_cfg=cfg,
-        num_envs=4,
-        steps_per_env=4,
-    )
-    runner.close()
-
-
 @pytest.mark.slow
 @pytest.mark.parametrize("env_name", ["Go2JoystickFlat"])
 def test_appo_runner_learn_two_iterations(env_name):

tests/algos/test_offpolicy_double_buffer_runner.py

@@ -11,7 +11,6 @@
 import pytest
 from hydra import compose, initialize_config_dir
 from hydra.core.global_hydra import GlobalHydra
-from hydra.errors import ConfigCompositionException
 from uni_rl.ipc.dp_launcher import UNILAB_DP_LOG_DIR, UNILAB_DP_RANK, UNILAB_DP_WORLD_SIZE
 
 _ROOT = Path(__file__).parent.parent.parent
@@ -83,35 +82,6 @@ def test_offpolicy_config_has_one_replay_path():
     cfg = _offpolicy_cfg()
     assert cfg.training.replay_prefetch_mode == "one_tick"
     assert cfg.training.env_steps_per_sync == 1
-    assert "env_steps_per_sync" not in cfg.algo
-    assert "inference_owner" not in cfg.training
-    assert "collector_infer_device" not in cfg.training
-    assert "no_sync_collection" not in cfg.training
-    assert "replay_pipeline" not in cfg.training
-    assert "verbose_metrics" not in cfg.training
-    assert "replay_pack_layout" not in cfg.training
-    assert "replay_pack_executor" not in cfg.training
-    assert "replay_h2d_submitter" not in cfg.training
-
-
-@pytest.mark.parametrize(
-    "override",
-    [
-        "training.replay_pipeline=cpu_pinned_double_buffer",
-        "training.verbose_metrics=true",
-        "training.num_gpus=2",
-        "training.multi_gpu_sync_mode=sync_sgd",
-        "training.multi_gpu_sync_interval=2",
-        "training.device=cuda",
-        "training.inference_owner=collector",
-        "training.collector_infer_device=cpu",
-        "training.no_sync_collection=true",
-        "algo.env_steps_per_sync=2",
-    ],
-)
-def test_removed_offpolicy_options_fail_hydra_compose(override: str):
-    with pytest.raises(ConfigCompositionException, match="Could not override"):
-        _offpolicy_cfg([override])
 
 
 @pytest.mark.parametrize("mode", ["invalid_mode", "same_tick"])
@@ -160,11 +130,6 @@ def test_sac_dispatch_constructs_unique_runner(monkeypatch: pytest.MonkeyPatch):
     assert runner.kwargs["algo_type"] == "sac"
     assert runner.kwargs["device"] == "cuda:0"
     assert runner.kwargs["replay_prefetch_mode"] == "one_tick"
-    assert "inference_owner" not in runner.kwargs
-    assert "collector_infer_device" not in runner.kwargs
-    assert "sync_collection" not in runner.kwargs
-    assert "replay_pipeline" not in runner.kwargs
-    assert "verbose_metrics" not in runner.kwargs
     assert runner.kwargs["learner"].kwargs == {
         "device": "cuda:0",
         "obs_dim": 4,
@@ -252,10 +217,6 @@ def test_td3_dispatch_constructs_unique_runner(monkeypatch: pytest.MonkeyPatch):
     assert isinstance(runner, _FakeRunner)
     assert runner.kwargs["algo_type"] == "td3"
     assert runner.kwargs["device"] == "cuda:0"
-    assert "inference_owner" not in runner.kwargs
-    assert "collector_infer_device" not in runner.kwargs
-    assert "sync_collection" not in runner.kwargs
-    assert "replay_pipeline" not in runner.kwargs
     assert runner.kwargs["replay_prefetch_mode"] == "one_tick"
     nan_guard_cfg = runner.kwargs["nan_guard_cfg"]
     assert nan_guard_cfg.enabled is True
@@ -326,10 +287,6 @@ def test_flashsac_dispatch_constructs_unique_runner(monkeypatch: pytest.MonkeyPa
     assert runner.kwargs["algo_type"] == "flashsac"
     assert runner.kwargs["device"] == "cuda:0"
     assert runner.kwargs["replay_prefetch_mode"] == "one_tick"
-    assert "inference_owner" not in runner.kwargs
-    assert "collector_infer_device" not in runner.kwargs
-    assert "sync_collection" not in runner.kwargs
-    assert "replay_pipeline" not in runner.kwargs
 
 
 def test_flashsac_n_step_is_rejected():

tests/algos/test_offpolicy_dp_sync.py

@@ -153,15 +153,6 @@ def test_learner_without_initial_sync_tensors_fails_with_type_error():
         runner._dp_init_broadcast()
 
 
-def test_close_closes_dp_sync_idempotently():
-    dp_sync = _FakeDpSync()
-    runner = _runner_with(_SyncLearner(), dp_sync)
-    # Avoid the full AsyncRunner.close(); only the dp_sync branch is under test.
-    runner.dp_sync.close()
-    runner.dp_sync.close()
-    assert [name for name, _ in dp_sync.calls] == ["close", "close"]
-
-
 def test_close_restores_terminal_and_ipc_before_destroying_process_group(monkeypatch):
     from uni_rl.offpolicy.runner import OffPolicyRunner
 
@@ -499,11 +490,6 @@ def __init__(self, *args, **kwargs):
     return runner.kwargs
 
 
-def test_offpolicy_config_has_no_periodic_parameter_sync_interval():
-    cfg = _offpolicy_cfg()
-    assert "dp_sync_interval" not in cfg.training
-
-
 def test_build_runner_single_rank_keeps_dp_sync_none(monkeypatch: pytest.MonkeyPatch):
     monkeypatch.delenv(UNILAB_DP_RANK, raising=False)
     kwargs = _build_sac_runner_with_dp_fakes(monkeypatch, [])