[Mansutti081]

Purpose

Bring the standalone RestoreBench updates from Mansutti081/RestoreBench@647ca89 into the existing PowerAgentBench Steady-State Level 3 integration. This makes the published campaign configuration match the intended three-run evaluation design, improves corpus-aware standalone scoring, and adds a run-preserving extraction workflow for auditable episode/event CSVs.

Changes and scope

• Set all six published RestoreBench campaigns to three repetitions (rep0–rep2), for 5,796 expected cells across both corpora.
• Add --data-dir support to restorebench-score for IEEE118 and PEGASE89, including corpus-matched batch stratification.
• Write one attempt_NNNN_score.json per batch input in addition to benchmark_report.json.
• Replace the private IEEE118 S0081 witness with a one-maneuver solution that converges at the primary 1e-8 tolerance without recovery.
• Add restorebench.sweeps.extract_three_run_data, which exports episode-level and event-level CSVs while validating three-run coverage and preserving missing observations.
• Update Level 3 documentation and tests. Existing PowerAgentBench-specific README installation guidance is preserved.

The extractor is a documented manual workflow and does not fabricate data: when result stores are absent, it emits header-only CSVs plus a report identifying all missing cells.

Validation

• Tested revision and Python version: 3583c5f, Python 3.11.15.
• Commands executed and results:
  • uv sync --frozen — passed.
  • .venv/bin/ruff check . — passed.
  • .venv/bin/pytest -q -m 'not slow' — 551 passed, 1 skipped, 5 deselected.
  • .venv/bin/pytest -q tests/environment/test_orchestrator.py::test_convergent_maneuver_is_not_success_when_slack_p_is_infeasible — 1 passed.
  • .venv/bin/restorebench-verify --dataset-dir dataset/ieee118 — 46/46 scenarios verified.
  • End-to-end smoke: batch scoring succeeded for IEEE118 S0008 and PEGASE89 S0001; per-attempt and aggregate reports were written; extractor validated 5,796 expected cells and handled absent result stores without inventing rows.
• CI run and status: GitHub Actions passed all five jobs (root tests on Python 3.10, 3.11, and 3.12; Level 2 tests on Python 3.11 and 3.12).
• Validation not performed and reason: four opt-in live-provider tests were not run because they issue real external model calls. No full 5,796-cell LLM campaign was run, and no experimental result stores are included.

Benchmark and compatibility impact

• Published campaign size changes intentionally from one to three repetitions. Existing result-cell naming remains compatible because repetition identity was already encoded as repN / repetition_index.
• The standalone scorer gains a backward-compatible --data-dir option; IEEE118 remains the default.
• Batch scoring now creates per-attempt JSON reports in addition to the existing aggregate report.
• The private S0081 witness changes, but the public scenario, action space, solver policy, and success criterion do not.
• No dependency or public/hidden-boundary changes are introduced.
• Reference result outputs were not regenerated because PowerAgentBench does not ship the underlying experimental result stores; the extractor reports this limitation explicitly.

Review checklist

• The change addresses a concrete problem and excludes unrelated work.
• Relevant tests cover the changed behavior, and CI includes them or manual-validation limits are explained.
• Documentation and commands match the implementation.
• Public/hidden data boundaries are preserved, and no secrets or licensed simulator data are included.
• The description and validation evidence reflect the current revision.