[Ker102]

Harneloop can now compare a frozen candidate against its baseline without silently mixing models, budgets, task sets or edited evidence. New benchmark commands pin source/grader revisions and candidate contents, require distinct artifact-backed task/trial runs, and report paired wins, losses, ties and unavailable evaluations.

Compatible candidates can be composed with fresh validation, and operators can retain lessons linked to finished-run artifacts. New run/evidence fingerprints are checked again before evidence-backed promotion, so changing the overlay or captured result requires fresh evidence. Legacy records remain readable.

Includes a CLI workflow guide and schema additions. Executors, model-provider clients, benchmark data, videos and raw outputs remain outside the installed framework. These checks verify local integrity and operator-attested conditions; they do not authenticate remote model execution or enforce a solver sandbox.

Validation:

• python -m unittest discover -s tests: 102 tests, successful, one POSIX executable-mode check skipped on Windows.
• python -m compileall -q src tests and CLI smoke checks passed.
• A separate frozen 16-task HumanEval+ Luna pilot scored 16/16 in both arms: no measured uplift and no promotion. MiMo results are not part of this implementation claim.

CodeRabbit review is not requested during this early prototype, following the repository PR template.

Closes #3.