G0~G8 성과·동맹 측정 OS 작업 일괄 고정

8월 7일까지 워킹트리에만 남아 있던 미커밋 작업을 커밋한다. 여러 사본
폴더(worktree·clone)에 흩어져 있던 중간 스냅샷을 정리하기 전에 원본을
git 이력으로 고정하는 것이 목적이다.

- contracts/routes/services: measurement, outcome_trajectory, rupture_repair,
  deliberate_practice, calibration_transfer, supervision_research,
  multimodal_alliance, continuous_improvement 계열 신규 모듈과 테스트
- infra/db/init: 07~16 마이그레이션(측정 기반~calibration transfer 실행)
- apps/web: 세션 리뷰 카드·관리 화면·E2E 스펙 추가
- docs/ops: G0~G8 라이브 통합·배포·롤백 증거 문서와 evidence JSON/PNG
- scripts: smoke·ledger·릴리스 에이전트·NAS 프리뷰 운영 스크립트

engine.public 로그 .bak과 apps/web/test-results 산출물은 커밋에서 제외했다.

apps/api/app/test_alliance_calibration.py

@@ -0,0 +1,126 @@
+"""Alliance calibration/drift report contract tests."""
+
+from __future__ import annotations
+
+import json
+import unittest
+from pathlib import Path
+
+from pydantic import ValidationError
+
+from .contracts.measurement import BenchmarkCase
+from .services.alliance_calibration import (
+    AllianceCalibrationPrediction,
+    AllianceCalibrationSnapshot,
+    compare_alliance_snapshots,
+    evaluate_alliance_snapshot,
+    gold_reference_snapshot,
+    render_alliance_comparison_markdown,
+)
+
+
+BENCHMARK_PATH = (
+    Path(__file__).resolve().parent
+    / "data"
+    / "outcome_alliance_benchmark_g0.v1.json"
+)
+
+
+def _cases() -> tuple[BenchmarkCase, ...]:
+    payload = json.loads(BENCHMARK_PATH.read_text(encoding="utf-8"))
+    return tuple(BenchmarkCase.model_validate(item) for item in payload["cases"])
+
+
+def _prediction(
+    case_id: str,
+    perspective: str,
+    dimension: str,
+    value: float,
+    evidence: tuple[int, ...],
+) -> AllianceCalibrationPrediction:
+    return AllianceCalibrationPrediction.model_validate(
+        {
+            "case_id": case_id,
+            "perspective": perspective,
+            "dimension": dimension,
+            "value": value,
+            "confidence": 0.8,
+            "evidence_turn_indices": evidence,
+        }
+    )
+
+
+class AllianceCalibrationTest(unittest.TestCase):
+    def test_gold_reference_is_explicit_and_complete(self) -> None:
+        reference = gold_reference_snapshot(_cases())
+        report = evaluate_alliance_snapshot(reference, _cases())
+        self.assertEqual(reference.provider, "human_gold")
+        self.assertEqual(len(reference.predictions), 9)
+        self.assertEqual(report["direction_accuracy"], 1.0)
+        self.assertEqual(report["evidence_recall"], 1.0)
+
+    def test_failure_predictions_cannot_hide_behind_neutral_scores(self) -> None:
+        with self.assertRaises(ValidationError):
+            AllianceCalibrationPrediction(
+                case_id="oas-g0-001",
+                perspective="independent_observer",
+                dimension="goal",
+                value=0.5,
+                status="error",
+                error_code="provider_timeout",
+            )
+
+    def test_snapshot_scores_each_expected_axis_without_total(self) -> None:
+        snapshot = AllianceCalibrationSnapshot(
+            run_id="candidate",
+            provider="fake",
+            model="fake-v1",
+            prompt_bundle_version="1.0.0",
+            predictions=(
+                _prediction("oas-g0-001", "independent_observer", "goal", 0.2, (0, 1, 2)),
+                _prediction("oas-g0-002", "independent_observer", "task", 0.2, (0, 1, 2)),
+                _prediction("oas-g0-003", "independent_observer", "bond", 0.2, (1, 2)),
+                _prediction("oas-g0-005", "client_agent_report", "bond", 0.2, (1,)),
+                _prediction("oas-g0-006", "client_agent_report", "task", 0.8, (2, 3)),
+                _prediction("oas-g0-007", "client_agent_report", "bond", 0.2, (2,)),
+                _prediction("oas-g0-008", "client_agent_report", "bond", 0.8, (0, 1)),
+                _prediction("oas-g0-008", "independent_observer", "goal", 0.2, (2, 3)),
+                _prediction("oas-g0-008", "independent_observer", "task", 0.2, (2, 3)),
+            ),
+        )
+        report = evaluate_alliance_snapshot(snapshot, _cases())
+        self.assertEqual(report["expected_count"], 9)
+        self.assertEqual(report["ready_count"], 9)
+        self.assertEqual(report["direction_accuracy"], 1.0)
+        self.assertEqual(report["evidence_recall"], 1.0)
+        self.assertNotIn("total", report["dimension_means"])
+
+    def test_comparison_reports_version_drift_and_errors(self) -> None:
+        baseline = AllianceCalibrationSnapshot(
+            run_id="baseline",
+            provider="reference",
+            model="gold-v1",
+            prompt_bundle_version="1.0.0",
+            predictions=(
+                _prediction("oas-g0-001", "independent_observer", "goal", 0.2, (0, 1, 2)),
+            ),
+        )
+        candidate = AllianceCalibrationSnapshot(
+            run_id="candidate",
+            provider="gateway",
+            model="candidate-v2",
+            prompt_bundle_version="1.1.0",
+            predictions=(
+                _prediction("oas-g0-001", "independent_observer", "goal", 0.35, (1, 2)),
+            ),
+        )
+        report = compare_alliance_snapshots(baseline, candidate, _cases())
+        self.assertAlmostEqual(report["drift"]["mean_absolute_score_delta"], 0.15)
+        markdown = render_alliance_comparison_markdown(report)
+        self.assertIn("candidate-v2", markdown)
+        self.assertIn("교육용 합성 장면", markdown)
+        self.assertIn("oas-g0-001", markdown)
+
+
+if __name__ == "__main__":
+    unittest.main()