feat(behavior1k): add BEHAVIOR-1K benchmark integration

[MilkClouds]

Summary

Adds the BEHAVIOR-1K (OmniGibson + NVIDIA Isaac Sim 4.5.0) benchmark, plus a zero-action baseline and a demo-replay model server used to verify env wiring against the released LeRobot v2.1 trajectories.

Replaces #56 (auto-closed when its stacked base feat/rlbench-license-guard was deleted on merge of #55). Rebased onto main.

What's in here

Benchmark module

• src/vla_eval/benchmarks/behavior1k/benchmark.py — Behavior1KBenchmark(StepBenchmark), R1Pro robot, 23-D action, RGB head + L/R wrist cameras. The async bridge is overridden so reset / step / cleanup run on a worker thread (Isaac Sim's SimulationApp.__init__ calls signal.signal, which has to be monkey-patched while the worker thread is set up — signal.signal assumes the main thread).
• Lazy imports for OmniGibson (heavy startup, can't be at module level — registry resolves the class without loading the sim).
• gm.HEADLESS=True set before og.launch (required to avoid an XR-extension segfault on the cluster).

Model servers

• behavior1k_baseline.py — zero-action 23-D baseline. Smoke-test sanity check.
• behavior1k_demo_replay.py — plays back recorded actions from a LeRobot v2.1 parquet episode. Used to verify the dataloader / action-space wiring matches the released dataset.

Docker

• Dockerfile.behavior1k — installs Isaac Sim 4.5.0 from pypi.nvidia.com (26 omni-* / isaacsim-* wheels), bddl3 + OmniGibson[eval] + joylo from BEHAVIOR-1K v3.7.2.
• Build is gated behind ARG ACCEPT_NVIDIA_EULA=YES (NVIDIA Omniverse EULA — https://docs.omniverse.nvidia.com/eula/). Surfaced via docker/build.sh --accept-license behavior1k, dispatched through the EULA_GATED map added in docker: opt-in license gate for rlbench builds #55.
• Listed under NO_REDIST in docker/push.sh, so the image is build-locally-only.

Configs / docs

• configs/behavior1k_eval.yaml — turning_on_radio, task instance 1, max 2000 steps.
• configs/model_servers/behavior1k/baseline.yaml — zero-action server.
• configs/model_servers/behavior1k/demo_replay.yaml — demo-replay server (parquet path placeholder).
• docs/reproductions/behavior1k.md — full repro write-up. Result data archived under docs/reproductions/data/.

Verification

# Demo replay (task = turning_on_radio, instance 1)
vla-eval run -c configs/behavior1k_eval.yaml \
            -m configs/model_servers/behavior1k/demo_replay.yaml
# → success=True, finished at step 1364/2000, wall 2933.8s

# Zero-action baseline
vla-eval run -c configs/behavior1k_eval.yaml \
            -m configs/model_servers/behavior1k/baseline.yaml
# → success=False at max_steps (expected)

The demo-replay success step (1364) falls inside the human-annotated press-skill window [1162, 1434] from the BEHAVIOR Dataset annotations, which gives reasonable confidence that the env wiring (action space, observation cameras, success detection) matches the upstream evaluation.

Skill notes

.claude/skills/add-benchmark and .claude/skills/add-model-server gain a short note not to add tests/test_<name>_benchmark.py or tests/test_<name>_server.py with mocked sim / model libraries. tests/ is for harness mechanics, not per-sim integration; mocked omnigibson / sapien / mujoco modules drift from upstream and miss the real bugs (import paths, action encoding, physics determinism). Verification is done via the smoke-test commands above.

Checklist

• I have read the relevant contributing guide (CONTRIBUTING.md)

Code changes:

• make check passes (ruff + ty)
• make test passes (pytest) — 295 passed, 1 skipped
• Smoke-tested affected configs (demo replay + baseline runs above)

Smoke test commands run:

make check
make test
docker/build.sh behavior1k --accept-license behavior1k
vla-eval run -c configs/behavior1k_eval.yaml -m configs/model_servers/behavior1k/baseline.yaml
vla-eval run -c configs/behavior1k_eval.yaml -m configs/model_servers/behavior1k/demo_replay.yaml