Add shared rule tests, calibration and agent guidance

[mrubens]

Summary

Add shared rule-example testing and calibration to Judgement. judgement test evaluates configured thresholds and fails on wrong or incomplete results. judgement calibrate --rule <id> and --all compare candidates using repeated full checks in disposable Git repositories. Library APIs support existing application inference adapters.

Use .judgement/rules.json as the only policy location and .judgement/examples/<id>.json for fixtures. Base rules still govern policy edits. Suites validate and snapshot every selected fixture before inference, bound concurrency, disable verdict caching, and include input hashes in JSON reports. Generated reports are intended for local output or CI artifacts.

Add an agent guide shipped in the npm package, linked from the README, covering rule design, labels, commands, uncertainty, threshold selection, held-out validation, and credential handling. Repository AGENTS.md documents implementation ownership and validation requirements. Applications keep rules, examples, and thin inference adapters rather than copying the runner.

Includes the 0.2.0 version bump for the configuration change. Not published.

Validation

• npm run check: 47 tests passed, including suite thresholds and exit semantics, preflight validation, input snapshots, cancellation, Git isolation, and CLI dry runs.
• npm pack --dry-run --json: examples and agent guide included.
• Earlier live custom-evaluator smoke test completed without operational failures and correctly declined to recommend an unsupported threshold.