[FreakyAdy]

Summary

Adds a first-party adapter for running local Harbor task directories from EvalGuard without importing Harbor's Python dependency tree.

• Maps stable EvalGuard task IDs to local Harbor task directories and validates task.toml before execution.
• Invokes harbor run with an explicit agent, model, jobs directory, unique job name, and one attempt using argument-list subprocess execution (no shell).
• Reads per-trial verifier_result.rewards and derives pass/fail from a configurable reward key and threshold; a zero exit alone is not treated as success.
• Reports timeout, startup, CLI-exit, and missing-reward errors; retains result files for audit snapshots.
• Adds mocked CLI tests for success, reward threshold, CLI error, missing reward, timeout, and invalid task mapping, plus adapter documentation.

Validation

Tests are mocked and do not require Harbor credentials, Docker, or installation of Harbor in EvalGuard's environment. GitHub Actions will validate lint, typing, unit tests, and coverage.

Scope

This contributes the Harbor half of #4. NeMo Gym is intentionally left for a separate adapter because its Python requirement and server protocols need a separate compatibility boundary.