Summary

Implements #107: a benchmark-agnostic evaluation protocol so LIBERO, Meta-World,
RoboSuite, ManiSkill, and user-authored tasks can plug into the SimEngine eval
loop without committing the core to any single benchmark's conventions.

Adapters stay thin and ship in follow-up extras (#108 Meta-World, #109 RoboSuite,
#110 LIBERO). The existing success_fn path is kept working for backcompat.

What's in scope (from #107)

1. BenchmarkProtocol ABC + StepInfo dataclass in strands_robots/simulation/benchmark.py
2. PolicyRunner.evaluate widened to accept spec: BenchmarkProtocol + seed, alongside the existing success_fn path
3. Tool actions + tool_spec.json entries: list_benchmarks, register_benchmark_from_file, evaluate_benchmark
4. Named-predicate library in strands_robots/simulation/predicates.py (11 built-ins: body_above_z, joint_above, distance_less_than, inside_region, contact_between, contact_any, distance_neg, joint_progress, constant, …)
5. Declarative YAML/JSON loader (register_benchmark_from_file) restricted to the named-predicate DSL - no eval, no exec, safe for LLM-authored specs
6. Reference adapter: DeclarativeBenchmark (the DSL-driven adapter) serves as the end-to-end reference. A full MetaWorldAdapter is tracked as a separate PR (Benchmark adapter: Meta-World (MetaWorldAdapter) #108) since it needs the metaworld env + real task validation.
7. Tests: 110 new tests covering protocol contract, cumulative reward, per-episode seed reproducibility, DSL compile + validation, backcompat, and MuJoCo dispatch

Architecture notes

• Robot compatibility is first-class metadata. Default on_episode_start validates loaded robot data_config against supported_robots and auto-loads default_robot when the sim is empty. Mismatches surface as structured error dicts.
• Registry mirrors register_urdf. Module-level dict[str, BenchmarkProtocol] guarded by an RLock. Re-registration is idempotent-overwrite with a warning.
• Predicate library is a closed registry. YAML/JSON specs can only reference predicates in PREDICATE_REGISTRY. No eval. Sandboxed by construction.
• Seed reproducibility. evaluate(spec=…, seed=42) seeds a master RNG, derives a per-episode child RNG, and threads it through spec.on_episode_start(sim, rng).
• Signature-driven dispatch stays intact. No router code changes; just three new enum entries and two new top-level properties (benchmark_name, spec_path).

Verification

```bash
hatch run lint # ruff + mypy, all clean
hatch run pytest tests/simulation/test_benchmark.py \
tests/simulation/test_benchmark_predicates.py \
tests/simulation/test_benchmark_dsl.py \
tests/simulation/test_policy_runner_benchmark.py \
tests/simulation/mujoco/test_benchmark_dispatch.py

→ 110 passed

```

Full suite: 1347 passed, 31 skipped. 3 pre-existing failures all require OpenGL/OSMesa (confirmed failing on `main` too).

Example

Python:

```python
sim.register_benchmark_from_file(benchmark_name='drawer-open', spec_path='specs/drawer.yaml')
sim.evaluate_benchmark(benchmark_name='drawer-open', robot_name='arm', policy_provider='mock', n_episodes=10, seed=42)
```

Spec (`specs/drawer.yaml`):

```yaml
name: drawer-open
max_steps: 300
supported_robots: [panda]
default_robot: panda
success:
all:
- {predicate: joint_above, joint: drawer_slide, value: 0.15}
failure:
any:
- {predicate: body_below_z, body: gripper, z: -0.1}
dense_reward:

• {predicate: distance_neg, body_a: gripper, body_b: drawer_handle, weight: 1.0}
• {predicate: joint_progress, joint: drawer_slide, target: 0.2, weight: 5.0}
  ```

Out of scope (tracked as follow-ups)

• Meta-World (Benchmark adapter: Meta-World (MetaWorldAdapter) #108), RoboSuite (Benchmark adapter: RoboSuite (RoboSuiteAdapter) #109), LIBERO + BDDL (Benchmark adapter: LIBERO (LiberoAdapter + BDDL parser) #110) adapters
• Dense-reward curriculum tooling, RL training harness
• Replacing the existing `success_fn` path (kept working)

Closes #107.