feat(sim): add BEHAVIOR-1K challenge support

[AdrianLlopart]

What changed

• add the official BEHAVIOR WebSocket evaluation bridge and pinned GR00T N1.7 sidecar adapter
• package the turning_on_radio B1K checkpoint as an R1Pro rSkill
• add the R1Pro manifest, OmniGibson scene sidecar, sim run, and full deploy sim configs
• preserve the native 61-D policy state through WorldState and atomically commit six safety-approved slots as one 23-D OmniGibson action

Why

OpenRAL needs a reproducible path to evaluate and deploy its policies on the 2026 BEHAVIOR-1K household tasks while retaining the normal safety-kernel and observability graph.

How tested

• just lint
• just test — 4,000 passed, 46 skipped
• openral sim run --config scenes/sim/behavior_turning_on_radio.yaml --rskill rskills/gr00t-n17-b1k-turning-on-radio --dry-run
• openral deploy sim --config scenes/deploy/behavior_r1pro.yaml --initial-task "turn on the radio" --no-object-detector --dry-run
• rSkill publisher dry-run validation

Live validation (RTX 4070 Laptop, 8 GB)

Commit c521377 adds whole-model NF4 in the GR00T sidecar (plus dropping the unused Qwen3-VL lm_head, whose full-vocab logits were the largest single inference allocation) and RGBA->RGB slicing in the WebSocket bridge. With that, the full official-evaluator loop runs on one 8 GB card:

• policy sidecar: 2.70 GiB loaded, 2.77 GiB inference peak (BF16 needed ~6.13 GiB and OOM'd next to OmniGibson); ~320-540 ms per inference
• full rollout, public instance 0 (ID 301): 3225 steps to timeout, evaluator MP4 + metrics JSON written; total GPU ~7.0/8.2 GiB, no OOM
• result: success=false, q_score 0.0 — infrastructure reproduction only; whether the zero score is NF4 degradation or checkpoint zero-shot behavior is not yet isolated, so no success-rate claim is made
• revalidated after the change: unit suite 4057 passed / 49 skipped, ruff + mypy --strict clean

[AdrianLlopart]

Code review

Found 3 issues:

1. WorldState.policy_state's docstring claims it is "never inferred", but SimAttachedHAL.read_policy_state() silently falls back to the generic obs["state"] key when a backend exposes no policy_state — and since HALLifecycleNodeBase._capture_proprio publishes it unconditionally, every sim backend that populates obs["state"] (LIBERO, RoboCasa, …) now feeds /openral/policy_state, not just r1pro/BEHAVIOR (CLAUDE.md says "Explicit beats implicit. No hidden retries, fallbacks, or magic globals"). The fallback branch is also untested. Either drop the fallback or make the docstring/tests match the real behavior.

openral/python/core/src/openral_core/schemas.py

Lines 2205 to 2207 in c521377

	checkpoint whose state includes more than joint positions/velocities.
	Populated through a typed HAL -> WorldState topic; never inferred.
	ee_poses: End-effector poses keyed by EE name.

openral/python/hal/src/openral_hal/sim_attached.py

Lines 1733 to 1735 in c521377

	return None
	raw = self._last_obs.get("policy_state", self._last_obs.get("state"))
	if raw is None:

2. _make_policy_adapter_skill's docstring documents exactly one obs["state"] substitution path (tf_lookup + state_contract.layout), but the PR adds a second, earlier path (policy_extras.use_world_state_policy_state → world_state.policy_state) that also sets state_assembled=True and skips the same downstream conversion (CLAUDE.md says "Docs travel with the code. Every PR updates all affected docs in the same commit range").

openral/packages/openral_rskill_ros/openral_rskill_ros/rskill_runner_node.py

Lines 2490 to 2492 in c521377

	use_policy_state = bool(
	getattr(self.manifest, "policy_extras", {}).get("use_world_state_policy_state")
	)

3. The ImportError belt-and-braces handler still calls the family-only model_family_install_hint(family), so a BEHAVIOR-1K manifest (policy_extras.implementation == "behavior_b1k_sidecar") that fails to import gets the generic gr00t hint — which this PR itself rewrote to say GR00T "runs in-process … no sidecar" — instead of just sync --group behavior-groot. The manifest-aware manifest_install_hint(manifest) added in policy_deps.py (and already used by deploy_sim.py's preflight) is the correct call here (bug due to install_hint = model_family_install_hint(family) in rskill_runner_node.py).

openral/packages/openral_rskill_ros/openral_rskill_ros/rskill_runner_node.py

Lines 1894 to 1896 in c521377

	family = manifest.model_family
	install_hint = model_family_install_hint(family)
	raise ROSRuntimeError(

🤖 Generated with Claude Code

_{- If this code review was useful, please react with 👍. Otherwise, react with 👎.}

[AdrianLlopart]

The 3 review findings above are addressed in 29a2b7ae4f4a4574f652659e5e0f1a55665da9f9 (fallback dropped + test, manifest-aware hint, both docstrings).