Port HybridGym benchmarks to Harbor

Context

The HybridGym benchmarks (hybridgym_funclocalize, hybridgym_funcgen, hybridgym_depsearch, and hybridgym_issuelocalize) do not appear to have corresponding Harbor registry datasets/adapters yet. These should be ported as a benchmark family while keeping per-task scoring behavior intact.

Proposed scope

• Create Harbor adapters/tasks for each HybridGym variant.
• Preserve prompts, environment setup, and evaluator semantics for each variant.
• Validate parity separately for function localization, function generation, dependency search, and issue localization.
• Document any shared adapter code and per-variant differences.

Acceptance criteria

• All four HybridGym variants can be run through Harbor.
• Per-variant parity results are recorded against the current OpenHands benchmark workflows.
• OpenHands wrappers can delegate HybridGym execution to Harbor.

Start by locating the Harbor registry and adapter/task entry points for benchmark integrations, then inspect the current OpenHands benchmark workflows for the four HybridGym variants. Add coverage for function localization, function generation, dependency search, and issue localization while preserving each variant’s prompts, environment setup, and evaluator semantics. Done means all four run through Harbor, parity results are recorded, and OpenHands wrappers can delegate execution to Harbor.