docs(tasks): task_template demonstrates _process_action instead of overriding step (#799)

* docs(tasks): task_template demonstrates _process_action instead of overriding step

The canonical task template showed an (illustrative, no-op) step() override as
the way to transform actions — teaching the exact anti-pattern the contract
discourages. Replace it with a _process_action example (unnormalise / delta /
ee-mapping), the sanctioned action hook. The template no longer overrides step,
so it leaves KNOWN_STEP_OVERRIDES (20->19) and new tasks copied from it model
the unified contract. 7/7 guardrail checks pass; ruff clean.

* docs(tasks): clarify task_template _process_action examples are copy-safe

Note that unnormalise_action is RLTaskEnv-only (BaseTaskEnv does not define it)
and that _last_action/_action_scale are user-defined, so a contributor copying
the MinimalTask (BaseTaskEnv) template won't hit AttributeError. Addresses review
nit on #799; docstring-only.

Author: geng-haoran

roboverse_pack/tasks/task_template.py

@@ -114,21 +114,24 @@ def _terminated(self, states) -> torch.Tensor:
         return torch.zeros(self.num_envs, dtype=torch.bool, device=self.device)
 
     # ========================================
-    # 4. (Optional) Override step and reset
+    # 4. (Optional) Transform actions, and reset custom state
     # ========================================
-    def step(self, actions: torch.Tensor):
-        """Execute one simulation step.
-
-        Optional customizations:
-        - Action normalization/unnormalization
-        - End-effector control mapping
-        - Custom control logic
-        - Other logic
+    def _process_action(self, actions: torch.Tensor):
+        """Transform actions before they are applied (the sanctioned action hook).
+
+        Prefer this over overriding ``step``: ``step`` calls ``_process_action``
+        first, then runs the rest of the shared pipeline (for ``RLTaskEnv`` that
+        includes auto-clamping to joint limits). Example patterns to adapt:
+        - unnormalisation:  ``return self.unnormalise_action(actions)``
+          (``unnormalise_action`` is provided by ``RLTaskEnv``; subclass that if
+          you need it — ``BaseTaskEnv`` does not define it)
+        - delta control:    ``return self._last_action + actions * self._action_scale``
+          (define ``self._last_action`` / ``self._action_scale`` on your task)
+        - end-effector / custom control mapping.
+
+        Default is identity (no transform).
         """
-        # Example: unnormalize actions from [-1,1] to joint limits
-        # actions = (actions + 1.0) / 2.0 * (action_high - action_low) + action_low
-
-        return super().step(actions)
+        return actions
 
     def reset(self, states=None, env_ids=None, seed=None):
         """Reset environment.

tests/test_task_reset_seed_contract.py

@@ -87,7 +87,6 @@
     "simpler_env/_metasim/base.py",
     "simpler_env/_metasim/coke_task.py",
     "simpler_env/_metasim/place_task.py",
-    "task_template.py",
 })
 KNOWN_CLOSE_OVERRIDES = frozenset({
     "robosuite/robosuite_env.py",