feat(sim): add semantic skill IR and compiler

.agents/skills/add-atomic-action/SKILL.md

@@ -20,15 +20,19 @@ Inspect only the files relevant to the requested skill:
 |---|---|
 | Base action and descriptors | `embodichain/lab/sim/atomic_actions/core.py` |
 | Goals and dynamic pose references | `embodichain/lab/sim/atomic_actions/goals.py` |
-| Role-to-resource binding | `embodichain/lab/sim/atomic_actions/bindings.py` |
+| Skill endpoint requirements | `embodichain/lab/sim/atomic_actions/requirements.py` |
+| Resolved endpoint bindings and targets | `embodichain/lab/sim/atomic_actions/bindings.py` |
 | Invocation, options, and resolved request | `embodichain/lab/sim/atomic_actions/invocation.py` |
 | Control-part semantic commands | `embodichain/lab/sim/atomic_actions/control.py` |
 | Invocation policies | `embodichain/lab/sim/atomic_actions/policies.py` |
 | Robot/task/scene state | `embodichain/lab/sim/atomic_actions/state.py` |
 | Dynamic scene provider contract | `embodichain/lab/sim/atomic_actions/scene.py` |
 | Effects and plans | `embodichain/lab/sim/atomic_actions/effects.py`, `plans.py` |
+| Runtime command frames and payloads | `embodichain/lab/sim/atomic_actions/runtime_commands.py` |
+| Endpoint command transports | `embodichain/lab/sim/atomic_actions/transports.py` |
 | Trajectory helpers | `embodichain/lab/sim/atomic_actions/trajectory_ops.py` |
 | Engine-owned planning resources | `embodichain/lab/sim/atomic_actions/runtime.py` |
+| Declarative robot resources and adapters | `embodichain/lab/sim/skills/profiles.py` |
 | Reference implementations | `embodichain/lab/sim/atomic_actions/primitives/` |
 | Static compiler and execution session | `engine.py`, `execution.py` |
 | Controller-facing execution ports | `runner.py`, `sim_adapter.py` |
@@ -92,24 +96,33 @@ class PushOptions(ActionOptions):
     push_distance: float = 0.05
 ```
 
-Do not put arm/hand names, hand qpos, or named robot postures in options. Bind
-participants with `ActionBinding`. Register embodiment-specific commands such
-as `open`, `grasp`, or `ready` on `ControlPartCommandProfile`; use
+Do not put arm/hand names, hand qpos, or named robot postures in options.
+Declare robot-independent participant slots and endpoints with
+`SkillBindingContract`; the engine or a bound robot skill profile produces the
+engine-owned `ActionBinding`. Register embodiment-specific commands such as
+`open`, `grasp`, or `ready` on `ControlPartCommandProfile`; use
 `ActionControlOverrides` only for one invocation revision.
 
 ## 3. Implement the planner
 
-Inherit `AtomicAction[PushGoal, PushOptions]` directly. Declare stable metadata and resolve
-resources from semantic binding roles.
+Inherit `AtomicAction[PushGoal, PushOptions]` directly. Declare stable metadata
+and an explicit, robot-independent endpoint contract. Every concrete action
+class must declare `binding_contract` in its own class body; use
+`SkillBindingContract()` for a skill that consumes no robot resource.
 
 ```python
 from typing import ClassVar
 
 from embodichain.lab.sim.atomic_actions import (
-    ResolvedActionRequest,
     ActionPlan,
     AtomicAction,
+    CARTESIAN_POSE_CAPABILITY,
+    JointPositionTarget,
     PlanningContext,
+    ResolvedActionRequest,
+    SkillBindingContract,
+    SkillEndpointRequirement,
+    SkillResourceSlot,
     StateDelta,
 )
 from embodichain.lab.sim.atomic_actions.trajectory_ops import (
@@ -122,7 +135,19 @@ class Push(AtomicAction[PushGoal, PushOptions]):
     skill_id: ClassVar[str] = "push"
     GoalType: ClassVar[type] = PushGoal
     OptionsType: ClassVar[type] = PushOptions
-    manipulator_roles: ClassVar[tuple[str, ...]] = ("primary",)
+    binding_contract: ClassVar[SkillBindingContract] = SkillBindingContract(
+        slots=(
+            SkillResourceSlot(
+                slot_id="primary",
+                endpoints=(
+                    SkillEndpointRequirement(
+                        endpoint_id="motion",
+                        capabilities=frozenset({CARTESIAN_POSE_CAPABILITY}),
+                    ),
+                ),
+            ),
+        ),
+    )
 
     def __init__(self, default_options: PushOptions | None = None) -> None:
         super().__init__(default_options)
@@ -133,11 +158,13 @@ class Push(AtomicAction[PushGoal, PushOptions]):
         context: PlanningContext,
     ) -> ActionPlan:
         goal = self.require_goal(request)
-        options = request.skill_options
-        manipulator = request.binding.manipulator("primary")
-        control_part = manipulator.name
-        joint_ids = list(manipulator.joint_ids)
+        motion_target = request.binding.endpoint(
+            "primary", "motion"
+        ).require_target(JointPositionTarget)
+        control_part = motion_target.control_part
+        joint_ids = list(motion_target.joint_ids)
         start_qpos = context.robot.qpos[:, joint_ids]
+        target_poses = goal.contact_pose
 
         # Build planner states and generate controlled-joint motion using
         # request.motion_policy. Embed it into full robot DoF.
@@ -168,6 +195,11 @@ Follow these invariants:
 
 - Let the engine supply `self.robot` and `self.motion_generator`; use
   `_on_bind()` only for robot/device-dependent setup.
+- Keep slot and endpoint IDs semantic and robot-independent. Declare all-of
+  capabilities, required typed commands, and disjointness constraints in the
+  `SkillBindingContract`; do not infer resources from endpoint names.
+- Resolve an endpoint with `request.binding.endpoint(slot_id, endpoint_id)` and
+  call `require_target(ExpectedTarget)` before using target-specific fields.
 - Import pure target-shaping, interpolation, pose-translation, and full-robot
   embedding helpers directly from `atomic_actions.trajectory_ops`; keep
   stateful planning inside `MotionGenerator`.
@@ -176,8 +208,8 @@ Follow these invariants:
   `plan()` method; the latter injects the latest dynamic obstacle poses into a
   copied planner policy.
 - Plan from `context.robot.qpos`, never an implicit live robot start state.
-- Return full-robot `(B, N, robot.dof)` motion as a tensor or
-  `TimedTrajectory` with matching `env_ids`.
+- For joint-backed motion, return full-robot `(B, N, robot.dof)` motion as a
+  tensor or `TimedTrajectory` with matching `env_ids` through `build_plan()`.
 - Preserve row-local planner success. `build_plan()` normalizes the mask and
   replaces unsuccessful trajectory rows with the context's observed qpos.
 - Preserve backend timing/derivatives when available.
@@ -196,7 +228,60 @@ Follow these invariants:
   `collision_entity_ids`; supported planners receive those entity poses through
   the framework-owned `plan()` entry point.
 
-## 4. Register and invoke
+## 4. Emit generic runtime commands when needed
+
+Use `build_command_plan()` when a skill targets a mobile base, whole-body
+controller, tool, or another non-joint transport. Build immutable endpoint
+commands; keep live controller and device handles in the transport:
+
+```python
+target = request.binding.endpoint("primary", "tool").require_target(ToolTarget)
+frames = tuple(
+    RuntimeCommandFrame(
+        commands=(EndpointCommand(target=target, payload=ToolPayload(value)),),
+        active_mask=torch.ones(
+            context.batch_size,
+            dtype=torch.bool,
+            device=context.robot.qpos.device,
+        ),
+        env_ids=context.env_ids,
+        hold_duration=torch.full(
+            (context.batch_size,),
+            request.motion_policy.control_dt,
+            device=context.robot.qpos.device,
+        ),
+    )
+    for value in command_values
+)
+return self.build_command_plan(
+    request,
+    context,
+    success=success,
+    commands=TimedCommandSequence(frames=frames, env_ids=context.env_ids),
+)
+```
+
+For a new transport kind:
+
+1. Define an immutable `RuntimeEndpointTarget` and `RuntimeCommandPayload` with
+   the same stable `transport_id`; both must return independently owned
+   snapshots. Payloads also expose `batch_size` and `device`. If target-specific
+   addressing or safe hold depends on fields beyond the exact target type,
+   `transport_id`, and `target_id`, override `address_fingerprint` to include
+   those immutable fields; frames, replans, and revisions preserve it.
+2. If declarative robot profiles select it, define a `ResourceEndpoint` and an
+   exact-type `ResourceEndpointAdapter` that returns `EndpointResolution` with
+   the runtime target and physical claim metadata.
+3. Implement `EndpointCommandTransport.send()`, `hold()`, and `cancel()`, then
+   register it in `EndpointCommandRouter` used as the `ExecutionRunner` command
+   sink. The router validates payload types before dispatch.
+
+The default command-plan feedback mode is timed and `joint_trajectory` is
+optional. Use joint-position feedback only when a matching full-robot
+`joint_trajectory` is supplied. Test target/payload snapshot ownership, frame
+batch/device consistency, routing, acknowledgement, hold, and cancel behavior.
+
+## 5. Register and invoke
 
 Register an instance by its class-level `skill_id`:
 
@@ -213,10 +298,14 @@ register_action(Push)
 Construct a grounded invocation explicitly:
 
 ```python
+binding = engine.bind_control_parts(
+    "push",
+    {"primary": {"motion": "left_arm"}},
+)
 invocation = ActionInvocation(
     skill_id="push",
     goal=PushGoal(contact_pose),
-    binding=ActionBinding(manipulators={"primary": "left_arm"}),
+    binding=binding,
     motion_policy=MotionPolicy(sample_count=60),
     recovery_policy=RecoveryPolicy(max_replans=2),
 )
@@ -228,26 +317,33 @@ For dynamic scene updates or online error recovery, create a session with
 through `ExecutionRunner`. Use non-blocking `runner.step()` in an existing event
 loop or `runner.run_until_blocked()` in a simple application.
 
-## 5. Export and document
+`engine.bind_control_parts()` is the explicit direct-core path for joint-backed
+control parts. When a `RobotSkillProfile` is installed, prefer
+`engine.skill_profile.resolve("push", selections).action_binding` so capability,
+command, resource-claim, and custom-adapter validation remain declarative.
+
+## 6. Export and document
 
 Export the goal, options, and action from:
 
 1. `embodichain/lab/sim/atomic_actions/primitives/__init__.py`
 2. `embodichain/lab/sim/atomic_actions/__init__.py`
 
-Add the stable skill ID, goal, roles, and effect to
+Add the stable skill ID, goal, binding slots/endpoints, and effect to
 `docs/source/overview/sim/atomic_actions/builtin_actions.md`. Update API docs for
 new public classes. Do not create a compatibility re-export module or a closed
 built-in-goal union.
 
-## 6. Test behavior
+## 7. Test behavior
 
 Add pure pytest tests under `tests/sim/atomic_actions/`. Cover:
 
-- descriptor `skill_id`, `GoalType`, and required roles;
-- invalid goal and missing binding rejection;
+- descriptor `skill_id`, `GoalType`, and explicit binding contract;
+- invalid goal, wrong binding owner, and missing/extra endpoint rejection;
 - per-environment planning success/failure masks;
 - full-robot trajectory shape, `env_ids`, timing, and failed-row hold behavior;
+- generic command target/payload ownership, frame batch/device consistency, and
+  optional `joint_trajectory` behavior when the skill emits command frames;
 - side-effect-free context handling;
 - masked `StateDelta` application for task effects;
 - `SceneEntityPose` replanning when the action accepts a dynamic goal;
@@ -264,9 +360,12 @@ then use the `pre-commit-check` skill before committing.
 |---|---|
 | Inherit another action | Inherit `AtomicAction` directly; compose helpers. |
 | Add one generic target with many optional fields | Define a narrow action-owned goal. |
-| Put hardware names in the goal | Bind semantic roles through `ActionBinding`. |
-| Put arm/hand control-part names in skill options | Use `ActionBinding` as their only source. |
-| Bind a joint, link, TCP frame, or arbitrary name | Every binding value must be a key in `RobotCfg.control_parts`. |
+| Put hardware names in the goal | Declare semantic slots/endpoints and resolve an engine-owned binding. |
+| Put arm/hand control-part names in skill options | Read typed runtime targets from bound endpoints. |
+| Declare legacy role tuples on the action | Declare a class-local `SkillBindingContract`. |
+| Use role-specific binding accessors | Use `binding.endpoint(...).require_target(...)`. |
+| Construct a binding from role dictionaries | Use a bound skill profile, or `engine.bind_control_parts()` for the direct joint path. |
+| Pass an arbitrary joint/link/TCP name to the direct path | `bind_control_parts()` values must be keys in `RobotCfg.control_parts`; add an endpoint adapter for another resource kind. |
 | Put hand qpos or named robot postures in skill options | Register semantic commands on the concrete control-part profile. |
 | Put planner/recovery knobs in skill options | Move them to invocation policies. |
 | Pass a motion generator to each action | Pass it once to `AtomicActionEngine`; construct actions from default options only. |
@@ -277,4 +376,6 @@ then use the `pre-commit-check` skill before committing.
 | Mutate held state after planning | Declare a `StateDelta`. |
 | Treat `plan_success` as physical success | Verify effects during execution. |
 | Step the simulator from the action | Emit plans; connect execution through `ExecutionRunner`. |
+| Put live controller handles in targets or payloads | Keep immutable addressing/data in values and own handles in the transport. |
+| Force a non-joint endpoint into a fake trajectory | Emit typed frames with `build_command_plan()` and install its transport. |
 | Override public `plan()` | Implement `_plan()` so scene binding cannot be bypassed. |