Port OpenAgentSafety benchmark to Harbor

[neubig]

Context

benchmarks/openagentsafety does not appear to have a corresponding Harbor registry dataset/adapter yet. To remove benchmark-specific execution code from this repository, OpenAgentSafety needs a Harbor-native task representation and verifier.

Proposed scope

• Model OpenAgentSafety scenarios as Harbor tasks.
• Preserve NPC/workplace interaction setup and safety scoring semantics.
• Validate Oracle/reference behavior where available.
• Compare aggregate results with the current OpenHands benchmark harness.

Acceptance criteria

• OpenAgentSafety tasks can be run through Harbor.
• Verifier/scoring parity is documented.
• The OpenHands wrapper can delegate OpenAgentSafety execution to Harbor.