feat(robots): add opt-in direct action primitive

[Elessar123]

Agents can use VLA and scripted primitives such as move_to, but cannot submit a native environment action as a planner tool. This adds execute_action(values=[...]) alongside those primitives for LIBERO, RoboCasa, and RoboTwin, exposed only when the run starts with --enable-direct-action.

Environment-owned action contract

• The connected environment supplies native action layouts and per-coordinate bounds through the read-only env.get_action_spec RPC. The shared client generates the tool schema and validates inputs from that specification; the primitive has no fixed action dimensions or blanket [-1, 1] limit.
• LIBERO and RoboCasa read the active simulator's native action_spec. RoboTwin describes native joint/pose waypoint layouts from its configured robot. Typed environments advertise their supported action_type values and default mode.
• Existing environment clients, cancellation, observation updates, recordings, and recipe export are reused. RoboCasa invalidates VLA history after a direct action so subsequent VLA calls reseed from the current scene.
• Default runs do not register the new tool or query its specification. No new dependencies. External environment servers must be updated with the client before enabling this optional tool.
• Paired README documentation and offline tests cover enablement, action forwarding, different environment dimensions and bounds, invalid inputs, cancellation, state records, and VLA history continuity.

This makes the direct-action primitive independent of a particular robot's vector size. It does not broaden the set of robot/controller configurations supported by the existing execution backends.

Validation

• pre-commit run --all-files and pre-commit run --files rpent/robots/components/action_spec.py — passed.
• .venv/bin/python -m pytest tests/unit_tests/robots/test_direct_action.py -q --tb=short — 36 passed.
• .venv/bin/python -m pytest tests/unit_tests -q --tb=short — 613 passed, 3 skipped, 4 failed. All four failures are in test_pi05_vla_server_contracts.py because the test environment lacks torch; the same four failures reproduced on unmodified main.
• Manual simulation smoke tests on the revised implementation used the changed checkout, production robot runtime hooks with only the environment component, real RPC clients, and Toolkit.execute_tool("execute_action", ...) on GPU 3. Each test fetched the live environment action specification:
  • LIBERO-PRO, libero_object_swap, task 2, seed 0: five actions moved the end effector 16.42 mm; six state records, updated images, and an episode video were produced.
  • RoboCasa, OpenDrawer, target split, seed 0: eight actions moved the base 1.59 mm; nine state records, updated images, and an episode video were produced.
  • Both owned environment processes shut down successfully.

The simulation checks establish action execution and observation/artifact updates, not autonomous planner performance or benchmark task success. RoboTwin simulation execution and live VLA/direct-action alternation remain unverified. RoboTwin native-layout tests require its optional runtime and were skipped locally; RoboCasa's VLA-history reseeding is covered offline.