diff --git a/docs/integration-tests.md b/docs/integration-tests.md
index f77dcff2..c10b1e27 100644
--- a/docs/integration-tests.md
+++ b/docs/integration-tests.md
@@ -134,3 +134,9 @@ PostgreSQL·Redis 경계를 증명하지 않습니다. Full suite 실패는 변
 
 Ref 호환성 판정은 [차등 검증](architecture/turn-state-differential-testing.md),
 UI는 [프론트엔드 호환 검증](frontend-legacy-parity.md)을 함께 사용합니다.
+
+전투 아이템 전체 차등 검증은 공격자·방어자 각각 145개 fixture를 모두 Core와
+Ref에서 실행하고 event, RNG, outcome, 전체 log bucket을 항목별로 비교합니다.
+Ref harness는 fixture마다 새 PHP process를 띄우지 않고 `--jsonl` 입력을 한
+process에서 일괄 처리합니다. `ITEM_PARITY_FILTER`는 특정 항목을 진단할 때만
+사용하며, 전체 검증의 fixture 수나 assertion을 줄이는 최적화로 사용하지 않습니다.
diff --git a/tools/integration-tests/test/battleDifferential.test.ts b/tools/integration-tests/test/battleDifferential.test.ts
index 8b0a918d..9f0cf569 100644
--- a/tools/integration-tests/test/battleDifferential.test.ts
+++ b/tools/integration-tests/test/battleDifferential.test.ts
@@ -582,7 +582,28 @@ const runReferenceTraceBatch = (workspaceRoot: string, fixtureLines: string[]):
     }
     const compareSourceRoot = process.env.REF_COMPARE_SOURCE_ROOT;
     if (!compareSourceRoot) {
-        throw new Error('BATTLE_CORPUS_PATH requires REF_COMPARE_SOURCE_ROOT with the JSONL-capable ref harness.');
+        const stdout = execFileSync(
+            'docker',
+            ['compose', 'exec', '-T', 'php', 'php', '/var/www/html/hwe/compare/battle_trace.php', '--jsonl'],
+            {
+                cwd: path.join(workspaceRoot, 'docker_compose_files/reference'),
+                input: normalizeJsonlForManifest(fixtureLines),
+                encoding: 'utf8',
+                stdio: ['pipe', 'pipe', 'pipe'],
+                maxBuffer: 512 * 1024 * 1024,
+            }
+        );
+        const traces = stdout
+            .split(/\r?\n/u)
+            .filter(Boolean)
+            .map((line) => JSON.parse(line) as ReferenceTrace);
+        if (traces.length !== fixtureLines.length) {
+            throw new Error(`ref batch returned ${traces.length} traces for ${fixtureLines.length} fixtures`);
+        }
+        traces.forEach((trace, index) =>
+            assertReferenceFixtureIdentity(trace, fixtureLines[index]!, `compose trace[${index}]`)
+        );
+        return traces;
     }
     const stdout = runReferenceSourceScript({
         workspaceRoot,
@@ -2370,6 +2391,14 @@ describeWithReference('ref ↔ core2026 battle differential', () => {
             },
         };
         const failures: string[] = [];
+        const cases: Array<{
+            itemKey: string;
+            base: BattleSimRequestPayload & { startYear: number };
+            coreEvents: WarBattleTraceEvent[];
+            coreLogs: CoreLogCapture;
+            coreRng: TracingRng | null;
+            coreOutcome: WarBattleOutcome | null;
+        }> = [];
 
         const itemFilter = process.env['ITEM_PARITY_FILTER'];
         for (const [itemKey, slot] of [...itemSlots].sort(([lhs], [rhs]) => lhs.localeCompare(rhs))) {
@@ -2423,7 +2452,16 @@ describeWithReference('ref ↔ core2026 battle differential', () => {
                     return coreRng.createRandUtil();
                 },
             });
-            const reference = runReferenceTrace(workspaceRoot!, JSON.stringify(base));
+            cases.push({ itemKey, base, coreEvents, coreLogs, coreRng, coreOutcome });
+        }
+
+        const references = runReferenceTraceBatch(
+            workspaceRoot!,
+            cases.map(({ base }) => JSON.stringify(base))
+        );
+        for (let index = 0; index < cases.length; index += 1) {
+            const { itemKey, base, coreEvents, coreLogs, coreRng, coreOutcome } = cases[index]!;
+            const reference = references[index]!;
             try {
                 assertTraceParity(coreEvents, reference, coreRng, coreOutcome);
                 assertAllLogBucketsParity(coreLogs, reference, base, `item.attacker.${itemKey}`);
@@ -2487,6 +2525,14 @@ describeWithReference('ref ↔ core2026 battle differential', () => {
             },
         };
         const failures: string[] = [];
+        const cases: Array<{
+            itemKey: string;
+            base: BattleSimRequestPayload & { startYear: number };
+            coreEvents: WarBattleTraceEvent[];
+            coreLogs: CoreLogCapture;
+            coreRng: TracingRng | null;
+            coreOutcome: WarBattleOutcome | null;
+        }> = [];
 
         const itemFilter = process.env['ITEM_PARITY_FILTER'];
         for (const [itemKey, slot] of [...itemSlots].sort(([lhs], [rhs]) => lhs.localeCompare(rhs))) {
@@ -2538,7 +2584,16 @@ describeWithReference('ref ↔ core2026 battle differential', () => {
                     return coreRng.createRandUtil();
                 },
             });
-            const reference = runReferenceTrace(workspaceRoot!, JSON.stringify(base));
+            cases.push({ itemKey, base, coreEvents, coreLogs, coreRng, coreOutcome });
+        }
+
+        const references = runReferenceTraceBatch(
+            workspaceRoot!,
+            cases.map(({ base }) => JSON.stringify(base))
+        );
+        for (let index = 0; index < cases.length; index += 1) {
+            const { itemKey, base, coreEvents, coreLogs, coreRng, coreOutcome } = cases[index]!;
+            const reference = references[index]!;
             try {
                 assertTraceParity(coreEvents, reference, coreRng, coreOutcome);
                 assertAllLogBucketsParity(coreLogs, reference, base, `item.defender.${itemKey}`);