Name the sim and mock components as the real ones

[T-K-233]

The mock and MuJoCo backends collapsed every bus into a single component, so a robot was LiteHardware in simulation and LiteLeftArm + LiteRightArm on hardware.

Anything keyed by component name therefore meant a different thing per backend — a controller's safety_components, the hardware_spawner, the manager's hardware_components_initial_state. The concrete cost: SafetyMonitorController names the real components, so it could not load in simulation at all — the one place the fault-handover path can be exercised without a robot.

Approved as item 3 of the structural audit.

What changed

                        before                    after
default / mock    LiteHardware              LiteLeftArm LiteRightArm
use_mock:=false   LiteLeftArm LiteRightArm  LiteLeftArm LiteRightArm
sim_mujoco:=true  LiteHardware              LiteLeftArm LiteRightArm

The combined macro emits one block per bus using the real backend's own block_name. Nothing in the vendored mujoco_ros2_control needs changing — it already iterates every <ros2_control> block and imports each.

Mock also gains a <gpio> per component carrying safety_level and safety_flags. The real drivers export those from export_unlisted_state_interface_descriptions(), which mock_components has no equivalent for. The base class parses gpio state interfaces straight from the description (hardware_component_interface.cpp:163) and GenericSystem overrides only the command side, so declaring them is sufficient.

MuJoCo is deliberately excluded. Our MujocoSystem overrides the deprecated by-value export_state_interfaces(), which bypasses the base class's gpio handling — a gpio there would promise interfaces that never appear. Full MuJoCo parity needs a fork patch and stays out of scope.

combined_block_name is deleted from every cad/ros2_control.json: with the real names in use, nothing reads it.

An edge case the tests caught

An IMU with a real component now gets a matching sensor block in sim. A sim-only IMU has no name to match and stays nested — test_imu_without_real_plugin_emits_no_sensor_component pins exactly that, and caught my first attempt, which invented a component for it.

Verification

1618 tests pass. Two new ones lock the parity in: component names equal across sim and real, mock a subset since it cannot back an IMU; and every mock system block carrying both safety interfaces.

Both were negative-controlled against the committed artifacts, which is what they actually read — sabotaging the generator alone leaves them green, so a generator-only control would have been worthless. Renaming a mock component gives sim ['LiteHardware', 'LiteRightArm'] != real ['LiteLeftArm', 'LiteRightArm']; deleting a gpio gives mock component LiteLeftArm declares no safety gpio.

Also caught in review: my first generated comment contained --, which is illegal inside an XML comment. The repo's own well-formedness test found it.

Not verified end to end: I did not get a clean live run of safety_monitor activating against the mock gpios in this session — my test harness kept losing the controller_manager. The mechanism is confirmed by expansion (the gpios are present with the right names) and by reading the upstream source, and an earlier structural audit did observe it activate. Worth a single ros2 launch … lite_mujoco.launch.py to confirm before relying on it in CI.

🤖 Generated with Claude Code

https://claude.ai/code/session_01EotyWG9YmYvg5xE2v3Zkpt