운영 안정화 변경을 최신 main과 통합한다

.env.example

@@ -91,6 +91,9 @@ DAEMON_REQUEST_TIMEOUT_MS=5000
 # realtime은 벽시계 경과를 game tick으로 투영합니다. manual은 daemon이
 # 다음 월 snapshot을 대기 없이 실행하며 테스트/장기 시뮬레이션에 사용합니다.
 GAME_CLOCK_MODE=realtime
+# Optional Windows host boot identity. Key is a private file, never a VITE_* value.
+# TURN_HOST_BOOT_URL=http://windows-private-address:17781/v1/boot-id
+# TURN_HOST_BOOT_KEY_FILE=/run/secrets/host_boot_key
 TRPC_PATH=/trpc
 
 # Frontend public URLs

.env.ops.example

@@ -16,6 +16,9 @@ GAME_API_HOST=0.0.0.0
 GAME_API_PORT=14000
 PROFILE=hwe
 SCENARIO=default
+# Optional Windows observation; unavailable or unconfigured service never blocks gameplay.
+# TURN_HOST_BOOT_URL=http://windows-private-address:17781/v1/boot-id
+# TURN_HOST_BOOT_KEY_FILE=/run/sammo-host-recovery/host_boot.key
 KAKAO_REST_KEY=your-kakao-rest-key
 KAKAO_ADMIN_KEY=your-kakao-admin-key
 KAKAO_REDIRECT_URI=https://example.com/oauth/kakao/callback

.gitignore

@@ -89,6 +89,7 @@ web_modules/
 .env.local
 .env.ci
 .env.ops
+tools/host-recovery/*.key
 tools/legacy-db-migration/migration-plan.json
 tools/legacy-db-migration/secrets/
 

app/game-api/src/realtime/outboxWorker.ts

@@ -76,7 +76,7 @@ export class ReadModelOutboxWorker implements ReadModelOutboxWakeup {
         this.leaseMs = normalizePositiveInteger(options.leaseMs, 30_000);
         this.retentionMs = normalizePositiveInteger(options.retentionMs, 24 * 60 * 60 * 1_000);
         this.pruneIntervalMs = normalizePositiveInteger(options.pruneIntervalMs, 60_000);
-        this.pruneLimit = normalizePositiveInteger(options.pruneLimit, 100);
+        this.pruneLimit = Math.min(500, normalizePositiveInteger(options.pruneLimit, 500));
         this.owner = options.owner ?? `game-api:${profileName}:${process.pid}:${randomUUID()}`;
         this.now = options.now ?? (() => new Date());
         this.onError = options.onError ?? (() => undefined);
@@ -109,11 +109,17 @@ export class ReadModelOutboxWorker implements ReadModelOutboxWakeup {
 
         const now = this.now();
         if (now.getTime() >= this.nextPruneAt) {
-            this.nextPruneAt = now.getTime() + this.pruneIntervalMs;
+            // backlog가 있으면 다음 dispatch에서도 이어서 정리한다. 평상시에는 낮은 빈도로 확인한다.
-            await pruneDeliveredReadModelOutbox(this.db, {
+            // 한 호출은 최대 10 batch/50ms이며 각 SQL도 최대 500행이다.
-                deliveredBefore: new Date(now.getTime() - this.retentionMs),
+            const started = performance.now();
-                limit: this.pruneLimit,
+            for (let batch = 0; batch < 10; batch += 1) {
-            });
+                const deleted = await pruneDeliveredReadModelOutbox(this.db, {
+                    deliveredBefore: new Date(now.getTime() - this.retentionMs),
+                    limit: this.pruneLimit,
+                });
+                this.nextPruneAt = deleted < this.pruneLimit ? now.getTime() + this.pruneIntervalMs : 0;
+                if (deleted < this.pruneLimit || performance.now() - started >= 50) break;
+            }
         }
         return result;
     }

app/game-api/src/router/playAudit/decisions.ts

@@ -1,6 +1,6 @@
 import { TRPCError } from '@trpc/server';
 import { z } from 'zod';
-import type { GamePrisma } from '@sammo-ts/infra';
+import { decodeAuditDecisionSteps, type GamePrisma } from '@sammo-ts/infra';
 import { auditProcedure, monthOrdinal, readAudit, readAuditWorld, zAuditMonth } from './shared.js';
 
 const zEffectivePolicy = z.object({
@@ -275,7 +275,7 @@ export const decisionDetail = auditProcedure
                 decision: project(row),
                 chunks: chunks.slice(0, input.limit).map((chunk) => ({
                     ordinal: chunk.ordinal,
-                    steps: z.array(zStep).max(128).parse(chunk.steps),
+                    steps: z.array(zStep).max(128).parse(decodeAuditDecisionSteps(chunk.steps)),
                 })),
                 nextCursor:
                     chunks.length && chunks.at(-1)!.ordinal + 1 < Math.ceil(row.stepCount / 128)

app/game-api/test/npcPolicyLifecycle.integration.test.ts

@@ -2,7 +2,12 @@ import { afterAll, beforeAll, expect, it } from 'vitest';
 import fastify from 'fastify';
 import { fastifyTRPCPlugin } from '@trpc/server/adapters/fastify';
 import { asRecord } from '@sammo-ts/common';
-import { createGamePostgresConnector, createRedisConnector, GamePrisma } from '@sammo-ts/infra';
+import {
+    decodeAuditDecisionSteps,
+    createGamePostgresConnector,
+    createRedisConnector,
+    GamePrisma,
+} from '@sammo-ts/infra';
 import { createTurnDaemonRuntime, seedScenarioToDatabase, type TurnDaemonRuntime } from '@sammo-ts/game-engine';
 import { appRouter } from '../src/router.js';
 import { DatabaseTurnDaemonTransport } from '../src/daemon/databaseTransport.js';
@@ -255,12 +260,10 @@ integration(
                 include: { chunks: { orderBy: { ordinal: 'asc' } } },
             });
             const procedures = nationDecision.chunks.flatMap((chunk) =>
-                Array.isArray(chunk.steps)
+                decodeAuditDecisionSteps(chunk.steps)
-                    ? chunk.steps
+                    .map(asRecord)
-                          .map(asRecord)
+                    .filter((step) => step.kind === 'PROCEDURE_START')
-                          .filter((step) => step.kind === 'PROCEDURE_START')
+                    .map((step) => step.procedure)
-                          .map((step) => step.procedure)
-                    : []
             );
             expect(procedures).toEqual(['NPC포상']);
             expect(configuration(await call('getPolicy', {}))).toEqual(configuration(withoutRecruitment));

app/game-api/test/readModelOutboxWorker.test.ts

@@ -45,6 +45,28 @@ const createFixture = (rows: readonly object[]) => {
 };
 
 describe('ReadModelOutboxWorker', () => {
+    it('drains more than one full retention batch rather than leaving a permanent 100-per-minute backlog', async () => {
+        let now = new Date('2026-11-09T00:00:00Z');
+        const fixture = createFixture([]);
+        fixture.queryRaw
+            .mockReset()
+            .mockResolvedValueOnce([])
+            .mockResolvedValueOnce([{ id: 1n }, { id: 2n }])
+            .mockResolvedValueOnce([{ id: 3n }, { id: 4n }])
+            .mockResolvedValueOnce([]);
+        const worker = new ReadModelOutboxWorker(fixture.db, fixture.redis, 'che:default', {
+            intervalMs: 60000,
+            pruneLimit: 2,
+            now: () => now,
+        });
+        now = new Date('2026-11-09T00:01:00Z');
+        worker.start();
+        await vi.waitFor(() => expect(fixture.queryRaw).toHaveBeenCalledTimes(4));
+        await worker.stop();
+        const query = fixture.queryRaw.mock.calls[1]?.[0] as { sql: string };
+        expect(query.sql).toContain('FOR UPDATE SKIP LOCKED');
+        expect(query.sql).toContain('ORDER BY "delivered_at", "available_at", "id"');
+    });
     it('publishes a survey-style global front-status invalidation without a turn daemon', async () => {
         const fixture = createFixture([{ id: 10n, payload: payload('front.global'), attempts: 1 }]);
         const worker = new ReadModelOutboxWorker(fixture.db, fixture.redis, 'che:default', {

app/game-api/test/securityTransport.integration.test.ts

@@ -13,6 +13,7 @@ import { afterAll, beforeAll, beforeEach, describe, expect, it } from 'vitest';
 import { encryptGameSessionToken, type GameSessionTokenPayload } from '@sammo-ts/common/auth/gameToken';
 import {
     createGamePostgresConnector,
+    encodeAuditDecisionSteps,
     GamePrisma,
     hashAuditDiplomacyDocument,
     createRedisConnector,
@@ -2345,50 +2346,52 @@ integration('game API security over HTTP transport', () => {
                     {
                         decisionId: decisionIds[0]!,
                         ordinal: 0,
-                        steps: Array.from({ length: 128 }, (_, sequence) =>
+                        steps: encodeAuditDecisionSteps(
-                            sequence === 127
+                            Array.from({ length: 128 }, (_, sequence) =>
-                                ? {
+                                sequence === 127
-                                      ...step,
+                                    ? {
-                                      sequence,
+                                          ...step,
-                                      kind: 'DECISION_START',
+                                          sequence,
-                                      reservedAction: '휴식',
+                                          kind: 'DECISION_START',
-                                      effectivePolicy: {
+                                          reservedAction: '휴식',
-                                          schemaVersion: 1,
+                                          effectivePolicy: {
-                                          general: { priority: ['징병'], flags: { 징병: true, 출병: false } },
+                                              schemaVersion: 1,
-                                          nation: {
+                                              general: { priority: ['징병'], flags: { 징병: true, 출병: false } },
-                                              priority: ['천도'],
+                                              nation: {
-                                              flags: { 천도: true },
+                                                  priority: ['천도'],
-                                              values: {
+                                                  flags: { 천도: true },
-                                                  reqNationGold: 4321,
+                                                  values: {
-                                                  reqNationRice: 100,
+                                                      reqNationGold: 4321,
-                                                  reqHumanWarUrgentGold: 100,
+                                                      reqNationRice: 100,
-                                                  reqHumanWarUrgentRice: 100,
+                                                      reqHumanWarUrgentGold: 100,
-                                                  reqHumanWarRecommandGold: 100,
+                                                      reqHumanWarUrgentRice: 100,
-                                                  reqHumanWarRecommandRice: 100,
+                                                      reqHumanWarRecommandGold: 100,
-                                                  reqHumanDevelGold: 100,
+                                                      reqHumanWarRecommandRice: 100,
-                                                  reqHumanDevelRice: 100,
+                                                      reqHumanDevelGold: 100,
-                                                  reqNpcWarGold: 100,
+                                                      reqHumanDevelRice: 100,
-                                                  reqNpcWarRice: 100,
+                                                      reqNpcWarGold: 100,
-                                                  reqNpcDevelGold: 100,
+                                                      reqNpcWarRice: 100,
-                                                  reqNpcDevelRice: 100,
+                                                      reqNpcDevelGold: 100,
-                                                  minimumResourceActionAmount: 100,
+                                                      reqNpcDevelRice: 100,
-                                                  maximumResourceActionAmount: 100,
+                                                      minimumResourceActionAmount: 100,
-                                                  minNpcWarLeadership: 100,
+                                                      maximumResourceActionAmount: 100,
-                                                  minWarCrew: 100,
+                                                      minNpcWarLeadership: 100,
-                                                  minNpcRecruitCityPopulation: 100,
+                                                      minWarCrew: 100,
-                                                  safeRecruitCityPopulationRatio: 100,
+                                                      minNpcRecruitCityPopulation: 100,
-                                                  properWarTrainAtmos: 100,
+                                                      safeRecruitCityPopulationRatio: 100,
-                                                  cureThreshold: 100,
+                                                      properWarTrainAtmos: 100,
+                                                      cureThreshold: 100,
+                                                  },
+                                                  combatForce: { '1': [2, 3] },
+                                                  supportForce: [4],
+                                                  developForce: [5],
+                                                  secret: 'decision-secret',
                                               },
-                                              combatForce: { '1': [2, 3] },
-                                              supportForce: [4],
-                                              developForce: [5],
                                               secret: 'decision-secret',
                                           },
-                                          secret: 'decision-secret',
+                                      }
-                                      },
+                                    : { ...step, sequence }
-                                  }
+                            )
-                                : { ...step, sequence }
                         ),
                     },
                     {

app/game-engine/src/index.ts

@@ -13,6 +13,7 @@ export * from './scenario/databaseUrl.js';
 export * from './scenario/mapLoader.js';
 export * from './scenario/scenarioSeeder.js';
 export * from './scenario/gameCancellation.js';
+export * from './playAudit/retention.js';
 export * from './turn/types.js';
 export * from './turn/worldLoader.js';
 export * from './turn/inMemoryWorld.js';

app/game-engine/src/lifecycle/databaseTurnDaemonLease.ts

@@ -2,6 +2,7 @@ import { randomUUID } from 'node:crypto';
 import { performance } from 'node:perf_hooks';
 
 import { createGamePostgresConnector, GamePrisma, type GamePrismaClient } from '@sammo-ts/infra';
+import type { HostBootMonitor } from '../turn/hostBootMonitor.js';
 
 export interface TurnDaemonLeaseToken {
     profile: string;
@@ -14,6 +15,7 @@ export interface DatabaseTurnDaemonLeaseOptions {
     ownerId?: string;
     leaseDurationMs?: number;
     heartbeat?: boolean;
+    hostBootMonitor?: HostBootMonitor;
 }
 
 export class TurnDaemonLeaseUnavailableError extends Error {
@@ -45,6 +47,7 @@ export class DatabaseTurnDaemonLease {
     private readonly ownerId: string;
     private readonly leaseDurationMs: number;
     private readonly heartbeatEnabled: boolean;
+    private readonly hostBootMonitor?: HostBootMonitor;
     private token: TurnDaemonLeaseToken | null = null;
     private heartbeatTimer: NodeJS.Timeout | null = null;
     private expiryTimer: NodeJS.Timeout | null = null;
@@ -63,6 +66,7 @@ export class DatabaseTurnDaemonLease {
         this.ownerId = options.ownerId ?? randomUUID();
         this.leaseDurationMs = normalizeLeaseDuration(options.leaseDurationMs);
         this.heartbeatEnabled = options.heartbeat ?? true;
+        this.hostBootMonitor = options.hostBootMonitor;
     }
 
     static async connect(
@@ -75,6 +79,8 @@ export class DatabaseTurnDaemonLease {
     }
 
     async acquire(): Promise<TurnDaemonLeaseToken | null> {
+        await this.checkHostBoot();
+        if (this.lost) throw this.getLossError();
         const requestStartedAt = performance.now();
         const rows = await this.db.$queryRaw<LeaseRow[]>(GamePrisma.sql`
             INSERT INTO "turn_daemon_lease" (
@@ -145,6 +151,12 @@ export class DatabaseTurnDaemonLease {
         return new TurnDaemonLeaseLostError(this.profile, this.lossReason);
     }
 
+    async checkHostBoot(): Promise<void> {
+        if (!this.lost && (await this.hostBootMonitor?.hasChanged())) {
+            this.markLost('Host boot identity changed; restart with a fresh database snapshot.');
+        }
+    }
+
     async renew(): Promise<boolean> {
         const token = this.token;
         if (!token || this.lost || this.renewalInFlight) {
@@ -153,6 +165,8 @@ export class DatabaseTurnDaemonLease {
         const requestStartedAt = performance.now();
         this.renewalInFlight = true;
         try {
+            await this.checkHostBoot();
+            if (this.lost) return false;
             const rows = await this.db.$queryRaw<LeaseRow[]>(GamePrisma.sql`
                 UPDATE "turn_daemon_lease"
                 SET
@@ -180,6 +194,7 @@ export class DatabaseTurnDaemonLease {
     }
 
     async assertActive(transaction?: GamePrisma.TransactionClient): Promise<void> {
+        await this.checkHostBoot();
         const token = this.token;
         if (!token || this.lost) {
             throw this.getLossError();
@@ -207,12 +222,13 @@ export class DatabaseTurnDaemonLease {
         this.stopExpiryWatchdog();
         const token = this.token;
         this.token = null;
-        if (!token || this.lost) {
+        if (!token) {
             return;
         }
         await this.db.$executeRaw(GamePrisma.sql`
             UPDATE "turn_daemon_lease"
-            SET "lease_until" = CURRENT_TIMESTAMP AT TIME ZONE 'UTC',
+            SET "clock_ready" = FALSE,
+                "lease_until" = CURRENT_TIMESTAMP AT TIME ZONE 'UTC',
                 "heartbeat_at" = CURRENT_TIMESTAMP AT TIME ZONE 'UTC'
             WHERE
                 "profile" = ${token.profile}

app/game-engine/src/lifecycle/turnDaemonLifecycle.ts

@@ -45,6 +45,7 @@ export interface TurnDaemonLifecycleDeps {
     commandResponder?: TurnDaemonCommandResponder;
     pauseGate?: () => Promise<boolean>;
     stateManager?: Pick<EngineStateManager, 'transaction'>;
+    runErrorRetry?: { wait(error: unknown): Promise<boolean>; reset(): void };
 }
 
 export class TurnDaemonLifecycle {
@@ -59,6 +60,7 @@ export class TurnDaemonLifecycle {
     private readonly commandResponder?: TurnDaemonCommandResponder;
     private readonly pauseGate?: () => Promise<boolean>;
     private readonly stateManager?: Pick<EngineStateManager, 'transaction'>;
+    private readonly runErrorRetry?: TurnDaemonLifecycleDeps['runErrorRetry'];
     private readonly options: TurnDaemonLifecycleOptions;
 
     private status: TurnDaemonStatus;
@@ -79,6 +81,7 @@ export class TurnDaemonLifecycle {
         this.commandResponder = deps.commandResponder;
         this.pauseGate = deps.pauseGate;
         this.stateManager = deps.stateManager;
+        this.runErrorRetry = deps.runErrorRetry;
         this.options = options;
         this.status = {
             state: 'idle',
@@ -192,8 +195,9 @@ export class TurnDaemonLifecycle {
             // 수동 실행 요청도 가오픈·정지·재조정의 턴 실행 gate를 통과해야 한다.
             // 사용자 명령 처리는 루프 시작에서 계속하되 시간 진행은 여기서 분리한다.
             if (this.pendingRun) {
-                await this.runOnce(this.pendingRun);
+                const pending = this.pendingRun;
                 this.pendingRun = null;
+                await this.runOnce(pending);
                 continue;
             }
 
@@ -447,6 +451,15 @@ export class TurnDaemonLifecycle {
             };
             result = this.stateManager ? await this.stateManager.transaction(runAndFlush) : await runAndFlush();
         } catch (error) {
+            // savepoint 복원이 성공한 획득 충돌만 대기한다. 복원 실패/AggregateError나
+            // 저장 중 소유권 상실을 자동 재실행하면 중복 효과를 숨길 수 있다.
+            if (this.stateManager && (await this.runErrorRetry?.wait(error))) {
+                this.status.running = false;
+                this.status.state = 'idle';
+                this.pendingRun = pending;
+                return;
+            }
+            this.runErrorRetry?.reset();
             this.status.running = false;
             this.status.state = 'paused';
             this.status.paused = true;
@@ -459,6 +472,7 @@ export class TurnDaemonLifecycle {
         }
 
         await this.applyRunResult(result, startMs);
+        this.runErrorRetry?.reset();
         this.status.state = 'idle';
         try {
             await this.hooks?.publishEvents?.(result);

app/game-engine/src/playAudit/decisionPersistence.ts

@@ -1,4 +1,4 @@
-import type { GamePrisma, InputJsonValue } from '@sammo-ts/infra';
+import { encodeAuditDecisionSteps, type GamePrisma, type InputJsonValue } from '@sammo-ts/infra';
 import { auditPolicyHash } from './policy.js';
 import type { PendingAuditDecision } from './decision.js';
 
@@ -50,7 +50,9 @@ export const persistAuditDecisions = async (
                 chunks.push({
                     decisionId: decision.id,
                     ordinal: start / AUDIT_DECISION_CHUNK_STEPS,
-                    steps: decision.steps.slice(start, start + AUDIT_DECISION_CHUNK_STEPS) as InputJsonValue,
+                    steps: encodeAuditDecisionSteps(
+                        decision.steps.slice(start, start + AUDIT_DECISION_CHUNK_STEPS) as InputJsonValue[]
+                    ),
                 });
                 if (chunks.length === BATCH) {
                     await tx.playAuditDecisionChunk.createMany({ data: chunks, skipDuplicates: true });

app/game-engine/src/playAudit/retention.ts

@@ -1,5 +1,5 @@
 import { asRecord } from '@sammo-ts/common';
-import type { GamePrismaClient } from '@sammo-ts/infra';
+import { GamePrisma, type GamePrismaClient } from '@sammo-ts/infra';
 
 export const AUDIT_RETENTION_BATCH_SIZE = 200;
 export type AuditRetentionResult = { status: 'progress' | 'complete' | 'busy' | 'identityChanged'; deleted: number };
@@ -13,6 +13,7 @@ export const prunePreviousAuditBatch = async (
     return db.$transaction(
         async (tx) => {
             await tx.$executeRaw`SET LOCAL statement_timeout = '2000ms'`;
+            await tx.$executeRaw`SET LOCAL lock_timeout = '250ms'`;
             // seeder와 동일한 잠금이다. RESET을 기다리게 하지 않고 다음 batch에서 재시도한다.
             const [lock] = await tx.$queryRaw<{ locked: boolean }[]>`
             SELECT pg_try_advisory_xact_lock(hashtextextended(current_schema(), 0)) AS locked
@@ -32,29 +33,43 @@ export const prunePreviousAuditBatch = async (
                 });
                 return { status: 'progress', deleted: deleted.count };
             }
-            const [decision] = await tx.$queryRaw<{ id: string }[]>`
+            // 이미 있는 (server_id,nation_id,tick,id) 인덱스 순서로 여러 부모를 잠근다.
-                SELECT id FROM play_audit_decision WHERE id = COALESCE(
+            // chunk 총량과 빈 header 삭제량의 합도 BATCH 이하로 제한한다.
-                    (SELECT id FROM play_audit_decision WHERE server_id < ${expectedServerId}
+            let decisions = await tx.$queryRaw<{ id: string }[]>`
-                     ORDER BY server_id, general_id, tick, id LIMIT 1),
+                SELECT id FROM play_audit_decision WHERE server_id < ${expectedServerId}
-                    (SELECT id FROM play_audit_decision WHERE server_id > ${expectedServerId}
+                ORDER BY server_id, nation_id, tick, id LIMIT ${AUDIT_RETENTION_BATCH_SIZE} FOR UPDATE
-                     ORDER BY server_id, general_id, tick, id LIMIT 1)
-                ) FOR UPDATE
             `;
-            if (decision) {
+            if (!decisions.length) {
-                const chunks = await tx.playAuditDecisionChunk.findMany({
+                decisions = await tx.$queryRaw<{ id: string }[]>`
-                    where: { decisionId: decision.id },
+                    SELECT id FROM play_audit_decision WHERE server_id > ${expectedServerId}
-                    orderBy: { ordinal: 'asc' },
+                    ORDER BY server_id, nation_id, tick, id LIMIT ${AUDIT_RETENTION_BATCH_SIZE} FOR UPDATE
-                    take: AUDIT_RETENTION_BATCH_SIZE,
+                `;
-                    select: { ordinal: true },
+            }
+            if (decisions.length) {
+                const ids = decisions.map(({ id }) => id);
+                const [chunks] = await tx.$queryRaw<{ deleted: number }[]>(GamePrisma.sql`
+                    WITH batch AS (
+                        SELECT decision_id, ordinal FROM play_audit_decision_chunk
+                        WHERE decision_id IN (${GamePrisma.join(ids)})
+                        ORDER BY decision_id, ordinal LIMIT ${AUDIT_RETENTION_BATCH_SIZE}
+                    ), removed AS (
+                        DELETE FROM play_audit_decision_chunk c USING batch b
+                        WHERE c.decision_id=b.decision_id AND c.ordinal=b.ordinal RETURNING 1
+                    ) SELECT count(*)::integer AS deleted FROM removed
+                `);
+                const count = chunks?.deleted ?? 0;
+                const remaining = AUDIT_RETENTION_BATCH_SIZE - count;
+                if (!remaining) return { status: 'progress', deleted: count };
+                const empty = await tx.playAuditDecision.findMany({
+                    where: { id: { in: ids }, chunks: { none: {} } },
+                    select: { id: true },
+                    orderBy: { id: 'asc' },
+                    take: remaining,
                 });
-                if (chunks.length) {
+                const headers = await tx.playAuditDecision.deleteMany({
-                    const deleted = await tx.playAuditDecisionChunk.deleteMany({
+                    where: { id: { in: empty.map(({ id }) => id) } },
-                        where: { decisionId: decision.id, ordinal: { in: chunks.map((row) => row.ordinal) } },
+                });
-                    });
+                return { status: 'progress', deleted: count + headers.count };
-                    return { status: 'progress', deleted: deleted.count };
-                }
-                await tx.playAuditDecision.delete({ where: { id: decision.id } });
-                return { status: 'progress', deleted: 1 };
             }
             const policies = await tx.playAuditPolicy.findMany({
                 where: { serverId: { not: expectedServerId } },

app/game-engine/src/playAudit/retentionWorker.ts

@@ -17,7 +17,7 @@ export const startAuditRetentionWorker = (options: {
         timer.unref();
     };
     const run = async () => {
-        let delay = 1000;
+        let delay = 100;
         try {
             const result = await options.prune();
             if (result.status === 'complete' || result.status === 'identityChanged') stopped = true;

app/game-engine/src/turn/cli.ts

@@ -8,6 +8,7 @@ import { createTurnDaemonMemoryReporter } from './turnDaemonMemoryReporter.js';
 import { createGatewayProfileGate } from './gatewayProfileGate.js';
 import { retryTurnDaemonLeaseStartup } from './leaseStartupRetry.js';
 import { TurnDaemonLeaseUnavailableError } from '../lifecycle/databaseTurnDaemonLease.js';
+import { loadHostBootMonitor } from './hostBootMonitor.js';
 
 export interface TurnDaemonCliOptions {
     profile?: string;
@@ -80,6 +81,7 @@ export const runTurnDaemonCli = async (options: TurnDaemonCliOptions = {}): Prom
         throw new Error(`GAME_CLOCK_MODE must be realtime or manual: ${rawGameClockMode}`);
     }
     const gameClockMode = rawGameClockMode as GameClockMode | undefined;
+    const hostBootMonitor = await loadHostBootMonitor(env);
 
     const runtime = await retryTurnDaemonLeaseStartup(() =>
         createTurnDaemonRuntime({
@@ -95,6 +97,7 @@ export const runTurnDaemonCli = async (options: TurnDaemonCliOptions = {}): Prom
             pauseGateIntervalMs,
             adminActionIntervalMs,
             gameClockMode,
+            hostBootMonitor,
         })
     ).catch(async (error: unknown) => {
         // 중복 starter가 정상 owner를 멈추면 안 된다. 그 밖의 초기화 실패는

app/game-engine/src/turn/hostBootMonitor.ts

@@ -0,0 +1,142 @@
+import { createHmac, randomBytes, timingSafeEqual } from 'node:crypto';
+import { readFile } from 'node:fs/promises';
+import { performance } from 'node:perf_hooks';
+import { isRecord } from '@sammo-ts/common';
+
+export interface HostBootMonitor {
+    hasChanged(): Promise<boolean>;
+}
+
+const mac = (key: Buffer, value: string): string => createHmac('sha256', key).update(value).digest('hex');
+
+/** 부팅 세대는 첫 인증 응답에 고정한다. 통신 실패는 기존 lease 복구를 막지 않는다. */
+export const createHostBootMonitor = (options: {
+    url: string;
+    key: Buffer;
+    pollMs?: number;
+    timeoutMs?: number;
+    now?: () => number;
+    wallNow?: () => number;
+    onUnavailable?: () => void;
+    onVerified?: () => void;
+}): HostBootMonitor => {
+    const url = new URL(options.url);
+    if (
+        !['http:', 'https:'].includes(url.protocol) ||
+        url.username ||
+        url.password ||
+        url.search ||
+        url.hash ||
+        url.pathname !== '/v1/boot-id'
+    ) {
+        throw new Error('Host boot URL must use HTTP(S) and /v1/boot-id without credentials or query.');
+    }
+    if (options.key.length !== 32) throw new Error('Host boot key must contain exactly 32 bytes.');
+    const now = options.now ?? (() => performance.now());
+    const wallNow = options.wallNow ?? (() => Date.now());
+    let baseline: string | null = null;
+    let changed = false;
+    let lastAttempt = -Infinity;
+    let lastWall = 0;
+    let lastWarning = -Infinity;
+    let inFlight: Promise<boolean> | null = null;
+    const sample = async (): Promise<boolean> => {
+        const nonce = randomBytes(32).toString('hex');
+        const response = await fetch(url, {
+            method: 'GET',
+            redirect: 'error',
+            signal: AbortSignal.timeout(options.timeoutMs ?? 1_000),
+            headers: {
+                'x-sammo-challenge': nonce,
+                'x-sammo-signature': mac(options.key, `sammo-host-boot-request-v1\n${nonce}`),
+            },
+        });
+        if (!response.ok || !response.body) {
+            await response.body?.cancel();
+            throw new Error('Host boot response unavailable.');
+        }
+        const reader = response.body.getReader();
+        const chunks: Uint8Array[] = [];
+        let size = 0;
+        try {
+            for (;;) {
+                const { done, value } = await reader.read();
+                if (done) break;
+                size += value.length;
+                if (size > 2_048) throw new Error('Host boot response exceeds its limit.');
+                chunks.push(value);
+            }
+        } finally {
+            await reader.cancel();
+        }
+        const body: unknown = JSON.parse(Buffer.concat(chunks).toString('utf8'));
+        if (
+            !isRecord(body) ||
+            body.version !== 1 ||
+            body.challenge !== nonce ||
+            typeof body.bootId !== 'string' ||
+            !/^[a-f0-9]{64}$/.test(body.bootId) ||
+            typeof body.signature !== 'string' ||
+            !/^[a-f0-9]{64}$/.test(body.signature)
+        ) {
+            throw new Error('Host boot response is invalid.');
+        }
+        const expected = mac(options.key, `sammo-host-boot-response-v1\n${nonce}\n${body.bootId}`);
+        if (!timingSafeEqual(Buffer.from(expected, 'hex'), Buffer.from(body.signature, 'hex'))) {
+            throw new Error('Host boot response signature is invalid.');
+        }
+        if (baseline === null) {
+            baseline = body.bootId;
+            options.onVerified?.();
+        }
+        changed ||= body.bootId !== baseline;
+        return changed;
+    };
+    return {
+        hasChanged(): Promise<boolean> {
+            if (changed) return Promise.resolve(true);
+            if (inFlight) return inFlight;
+            const current = now();
+            const wall = wallNow();
+            const discontinuity = Math.abs(wall - lastWall - (current - lastAttempt)) > 1_000;
+            if (current - lastAttempt < (options.pollMs ?? 2_000) && !discontinuity) return Promise.resolve(false);
+            lastAttempt = current;
+            lastWall = wall;
+            inFlight = sample()
+                .catch(() => {
+                    if (current - lastWarning >= 60_000) {
+                        options.onUnavailable?.();
+                        lastWarning = current;
+                    }
+                    return false;
+                })
+                .finally(() => {
+                    inFlight = null;
+                });
+            return inFlight;
+        },
+    };
+};
+
+export const loadHostBootMonitor = async (env: NodeJS.ProcessEnv): Promise<HostBootMonitor | undefined> => {
+    const url = env.TURN_HOST_BOOT_URL?.trim();
+    const keyFile = env.TURN_HOST_BOOT_KEY_FILE?.trim();
+    if (!url && !keyFile) return undefined;
+    try {
+        if (!url || !keyFile) throw new Error('Incomplete optional configuration.');
+        const encoded = (await readFile(keyFile, 'utf8')).trim();
+        if (!/^[a-f0-9]{64}$/.test(encoded)) throw new Error('Invalid optional key file.');
+        return createHostBootMonitor({
+            url,
+            key: Buffer.from(encoded, 'hex'),
+            onVerified: () => console.info('[turn-daemon] host boot identity verified.'),
+            onUnavailable: () => {
+                console.warn('[turn-daemon] host boot identity unavailable; ordinary lease recovery remains active.');
+            },
+        });
+    } catch {
+        // 보조 관측기의 설정/파일 부재도 게임 기동의 필수 의존성이 되지 않는다.
+        console.warn('[turn-daemon] host boot observation disabled; check optional configuration and key file.');
+        return undefined;
+    }
+};

app/game-engine/src/turn/reservedTurnLeaseRetry.ts

@@ -0,0 +1,38 @@
+import { performance } from 'node:perf_hooks';
+import { setTimeout } from 'node:timers/promises';
+import { ReservedTurnLeaseBusyError } from './reservedTurnStore.js';
+
+/** rollback 완료 뒤에만 대기한다. 다음 loop에서 pause/lease gate와 최신 큐를 다시 읽는다. */
+export const createReservedTurnLeaseRetry = (
+    options: {
+        now?: () => number;
+        sleep?: (ms: number) => Promise<void>;
+        maxWaitMs?: number;
+        onWait?: (queueKey: string) => void;
+    } = {}
+) => {
+    const now = options.now ?? (() => performance.now());
+    const sleep = options.sleep ?? ((ms: number) => setTimeout(ms));
+    const maxWaitMs = options.maxWaitMs ?? 5 * 60_000 + 1_000;
+    let startedAt: number | null = null;
+    let lastReportedAt = -Infinity;
+    return {
+        reset(): void {
+            startedAt = null;
+            lastReportedAt = -Infinity;
+        },
+        async wait(error: unknown): Promise<boolean> {
+            if (!(error instanceof ReservedTurnLeaseBusyError)) return false;
+            const current = now();
+            startedAt ??= current;
+            const remaining = maxWaitMs - (current - startedAt);
+            if (remaining <= 0) return false;
+            if (current - lastReportedAt >= 30_000) {
+                options.onWait?.(error.queueKey);
+                lastReportedAt = current;
+            }
+            await sleep(Math.min(1_000, remaining));
+            return true;
+        },
+    };
+};

app/game-engine/src/turn/reservedTurnStore.ts

@@ -129,6 +129,14 @@ export class ReservedTurnLeaseConflictError extends Error {
     }
 }
 
+/** 실행 전 획득 충돌만 재시도한다. flush/finalize의 소유권 상실과 구분한다. */
+export class ReservedTurnLeaseBusyError extends ReservedTurnLeaseConflictError {
+    constructor(queueKey: string) {
+        super(queueKey);
+        this.name = 'ReservedTurnLeaseBusyError';
+    }
+}
+
 export class InMemoryReservedTurnStore {
     private readonly generalTurns = new Map<number, ReservedTurnEntry[]>();
     private readonly nationTurns = new Map<string, ReservedTurnEntry[]>();
@@ -249,11 +257,7 @@ export class InMemoryReservedTurnStore {
             }
         };
         pruneNationKeys(Array.from(this.nationTurns.keys()), (key) => this.nationTurns.delete(key));
-        for (const keys of [
+        for (const keys of [this.dirtyNationKeys, this.pendingNationInitializationKeys, this.leasedNationKeys]) {
-            this.dirtyNationKeys,
-            this.pendingNationInitializationKeys,
-            this.leasedNationKeys,
-        ]) {
             for (const key of Array.from(keys)) {
                 const nationId = Number(key.split(':', 1)[0]);
                 if (deletedNations.has(nationId)) {
@@ -376,7 +380,7 @@ export class InMemoryReservedTurnStore {
             }
         }
         if (claimed.count !== 1) {
-            throw new ReservedTurnLeaseConflictError(`general:${generalId}`);
+            throw new ReservedTurnLeaseBusyError(`general:${generalId}`);
         }
         this.leasedGeneralIds.add(generalId);
         return !retainedExistingLease;
@@ -427,7 +431,7 @@ export class InMemoryReservedTurnStore {
             }
         }
         if (claimed.count !== 1) {
-            throw new ReservedTurnLeaseConflictError(`nation:${nationId}:${officerLevel}`);
+            throw new ReservedTurnLeaseBusyError(`nation:${nationId}:${officerLevel}`);
         }
         this.leasedNationKeys.add(buildNationKey(nationId, officerLevel));
         return !retainedExistingLease;
@@ -904,6 +908,24 @@ export class InMemoryReservedTurnStore {
         await this.persistChanges(this.prisma, changes);
         this.acknowledgeDirtyState(changes);
     }
+
+    /** 실행 루프 종료 뒤 호출한다. rollback으로 journal에서 빠진 DB 임대도 해제한다. */
+    async releaseOwnedLeases(): Promise<void> {
+        const results = await Promise.allSettled([
+            this.prisma.generalTurnRevision?.updateMany({
+                where: { leaseOwner: this.leaseOwner },
+                data: { leaseOwner: null, leaseExpiresAt: null },
+            }),
+            this.prisma.nationTurnRevision?.updateMany({
+                where: { leaseOwner: this.leaseOwner },
+                data: { leaseOwner: null, leaseExpiresAt: null },
+            }),
+        ]);
+        const errors = results.flatMap((result) => (result.status === 'rejected' ? [result.reason] : []));
+        if (errors.length) throw new AggregateError(errors, 'Reserved turn lease cleanup failed.');
+        this.leasedGeneralIds.clear();
+        this.leasedNationKeys.clear();
+    }
 }
 
 export const createReservedTurnStore = async (options: ReservedTurnStoreOptions): Promise<ReservedTurnStoreHandle> => {
@@ -918,6 +940,12 @@ export const createReservedTurnStore = async (options: ReservedTurnStoreOptions)
     await store.loadAll();
     return {
         store,
-        close: () => connector.disconnect(),
+        close: async () => {
+            try {
+                await store.releaseOwnedLeases();
+            } finally {
+                await connector.disconnect();
+            }
+        },
     };
 };

app/game-engine/src/turn/runtimePauseGate.ts

@@ -2,7 +2,7 @@ import type { GameClockPhase } from '@sammo-ts/common';
 
 /** Gateway의 실행 gate와 durable 시계를 명령 claim 전에 맞춘다. */
 export const createRuntimePauseGate = (options: {
-    assertLease(): void;
+    assertLease(): void | Promise<void>;
     shouldPause(): Promise<boolean>;
     getPhase(): GameClockPhase;
     isExplicitlyPaused(): boolean;
@@ -11,7 +11,7 @@ export const createRuntimePauseGate = (options: {
 }): (() => Promise<boolean>) => {
     let lastPaused: boolean | null = null;
     return async () => {
-        options.assertLease();
+        await options.assertLease();
         const paused = await options.shouldPause();
         const phase = options.getPhase();
         // 오류 정지는 Gateway 상태만 PAUSED로 바꿀 수 있다. 그대로 두면 가입은

app/game-engine/src/turn/turnDaemon.ts

@@ -108,6 +108,8 @@ import { DatabaseTurnDaemonLease, TurnDaemonLeaseUnavailableError } from '../lif
 import { EngineStateManager } from './engineStateManager.js';
 import { applyRuntimeClockShift } from './runtimeClockShift.js';
 import { applyRuntimeGameSettings } from './runtimeGameSettings.js';
+import { createReservedTurnLeaseRetry } from './reservedTurnLeaseRetry.js';
+import type { HostBootMonitor } from './hostBootMonitor.js';
 
 export interface TurnDaemonRuntimeOptions {
     profile: string;
@@ -134,6 +136,7 @@ export interface TurnDaemonRuntimeOptions {
     leaseDurationMs?: number;
     leaseOwnerId?: string;
     enableLeaseHeartbeat?: boolean;
+    hostBootMonitor?: HostBootMonitor;
     /**
      * Isolated, single-process fixture acceleration only. Reserved turns are
      * loaded once and no concurrent API writer may touch this database.
@@ -985,6 +988,7 @@ const createTurnDaemonRuntimeWithLease = async (
         prepareClockRecovery = dbHooks.prepareRealtimeRecovery;
         prunePreviousAudit = dbHooks.prunePreviousAudit;
         close = async () => {
+            await adminActionConsumer?.stop();
             if (auctionBidder) {
                 await auctionBidder.close();
             }
@@ -995,12 +999,15 @@ const createTurnDaemonRuntimeWithLease = async (
                 await tournamentRewardFinalizer.close();
             }
             await dbHooks.close();
-            if (reservedTurnStoreHandle) {
+            try {
-                await reservedTurnStoreHandle.close();
+                await reservedTurnStoreHandle?.close();
+            } finally {
+                try {
+                    await gatewayGate?.close();
+                } finally {
+                    await turnDaemonLease?.close();
+                }
             }
-            await gatewayGate?.close();
-            await adminActionConsumer?.stop();
-            await turnDaemonLease?.close();
         };
     } else if (reservedTurnStoreHandle) {
         hooks = {
@@ -1009,9 +1016,12 @@ const createTurnDaemonRuntimeWithLease = async (
             },
         };
         close = async () => {
-            await reservedTurnStoreHandle.close();
-            await gatewayGate?.close();
             await adminActionConsumer?.stop();
+            try {
+                await reservedTurnStoreHandle.close();
+            } finally {
+                await gatewayGate?.close();
+            }
         };
     } else if (gatewayGate) {
         hooks = {
@@ -1070,12 +1080,17 @@ const createTurnDaemonRuntimeWithLease = async (
     const baseClose = close;
     close = async () => {
         stopClockProjectionWorker();
-        await baseClose();
+        const baseResults = await Promise.allSettled([baseClose()]);
-        await neutralAuctionRegistrar.close();
+        // queue 임대 정리 실패도 다른 연결/worker의 종료를 막지 않게 한다.
-        if (redisConnector) {
+        const results = await Promise.allSettled([
-            await redisConnector.disconnect();
+            neutralAuctionRegistrar.close(),
-        }
+            redisConnector?.disconnect(),
-        await commandConnector?.disconnect();
+            commandConnector?.disconnect(),
+        ]);
+        const failures = [...baseResults, ...results].flatMap((result) =>
+            result.status === 'rejected' ? [result.reason] : []
+        );
+        if (failures.length) throw new AggregateError(failures, 'Turn daemon connection cleanup failed.');
     };
 
     const resolvedControlQueue = options.controlQueue ?? databaseCommandQueue ?? controlQueue;
@@ -1109,7 +1124,8 @@ const createTurnDaemonRuntimeWithLease = async (
             processor,
             hooks,
             pauseGate: createRuntimePauseGate({
-                assertLease: () => {
+                assertLease: async () => {
+                    await turnDaemonLease?.checkHostBoot();
                     if (turnDaemonLease?.isLost()) throw turnDaemonLease.getLossError();
                 },
                 shouldPause: async () => (await pauseGate?.()) ?? false,
@@ -1127,6 +1143,9 @@ const createTurnDaemonRuntimeWithLease = async (
             // accumulated world before every due tick in that isolated mode;
             // production and gateway-managed runtimes keep rollback savepoints.
             stateManager: options.exclusiveFastForward ? undefined : stateManager,
+            runErrorRetry: createReservedTurnLeaseRetry({
+                onWait: (queueKey) => console.info(`[turn-daemon] waiting for reserved turn lease: ${queueKey}`),
+            }),
         },
         { profile: options.profile, defaultBudget }
     );
@@ -1179,6 +1198,7 @@ export const createTurnDaemonRuntime = async (options: TurnDaemonRuntimeOptions)
               ownerId: options.leaseOwnerId,
               leaseDurationMs: options.leaseDurationMs,
               heartbeat: options.enableLeaseHeartbeat,
+              hostBootMonitor: options.hostBootMonitor,
           })
         : null;
     if (turnDaemonLease && !(await turnDaemonLease.acquire())) {

app/game-engine/test/hostBootMonitor.test.ts

@@ -0,0 +1,134 @@
+import { createHmac, randomBytes } from 'node:crypto';
+import { once } from 'node:events';
+import { createServer } from 'node:http';
+import { afterEach, describe, expect, it, vi } from 'vitest';
+import { createHostBootMonitor, loadHostBootMonitor } from '../src/turn/hostBootMonitor.js';
+
+const cleanups: Array<() => Promise<void>> = [];
+afterEach(async () => {
+    await Promise.all(cleanups.splice(0).map((cleanup) => cleanup()));
+});
+
+const fixture = async () => {
+    const key = randomBytes(32);
+    let bootId = 'a'.repeat(64);
+    let replay: string | null = null;
+    let mode: 'valid' | 'replay' | 'forged' | 'oversized' | 'offline' | 'hang' = 'valid';
+    const requests: string[] = [];
+    const server = createServer((request, response) => {
+        const nonce = String(request.headers['x-sammo-challenge']);
+        requests.push(nonce);
+        expect(request.headers['x-sammo-signature']).toBe(
+            createHmac('sha256', key).update(`sammo-host-boot-request-v1\n${nonce}`).digest('hex')
+        );
+        if (mode === 'hang') return;
+        if (mode === 'offline') {
+            response.writeHead(503).end();
+            return;
+        }
+        if (mode === 'oversized') {
+            response.end('x'.repeat(4096));
+            return;
+        }
+        if (mode === 'replay') {
+            response.end(replay);
+            return;
+        }
+        const body = JSON.stringify({
+            version: 1,
+            challenge: nonce,
+            bootId,
+            signature:
+                mode === 'forged'
+                    ? '0'.repeat(64)
+                    : createHmac('sha256', key)
+                          .update(`sammo-host-boot-response-v1\n${nonce}\n${bootId}`)
+                          .digest('hex'),
+        });
+        replay ??= body;
+        response.end(body);
+    });
+    server.listen(0, '127.0.0.1');
+    await once(server, 'listening');
+    cleanups.push(
+        () =>
+            new Promise<void>((resolve) => {
+                server.close(() => resolve());
+                server.closeAllConnections();
+            })
+    );
+    const address = server.address();
+    if (!address || typeof address === 'string') throw new Error('Missing fixture listener');
+    return {
+        key,
+        requests,
+        url: `http://127.0.0.1:${address.port}/v1/boot-id`,
+        setBoot: (value: string) => {
+            bootId = value;
+        },
+        setMode: (value: typeof mode) => {
+            mode = value;
+        },
+    };
+};
+
+describe('Windows boot observation', () => {
+    it('uses a stable baseline, survives outages, and latches a signed boot change', async () => {
+        const f = await fixture();
+        const monitor = createHostBootMonitor({ ...f, pollMs: 0 });
+        expect(await monitor.hasChanged()).toBe(false);
+        f.setMode('offline');
+        expect(await monitor.hasChanged()).toBe(false);
+        f.setBoot('b'.repeat(64));
+        f.setMode('valid');
+        expect(await monitor.hasChanged()).toBe(true);
+        f.setMode('offline');
+        expect(await monitor.hasChanged()).toBe(true);
+    });
+
+    it.each(['replay', 'forged', 'oversized', 'hang'] as const)(
+        'rejects %s without changing the baseline',
+        async (mode) => {
+            const f = await fixture();
+            const onUnavailable = vi.fn();
+            const monitor = createHostBootMonitor({ ...f, pollMs: 0, timeoutMs: 100, onUnavailable });
+            expect(await monitor.hasChanged()).toBe(false);
+            f.setBoot('b'.repeat(64));
+            f.setMode(mode);
+            expect(await monitor.hasChanged()).toBe(false);
+            expect(onUnavailable).toHaveBeenCalledOnce();
+            f.setMode('valid');
+            expect(await monitor.hasChanged()).toBe(true);
+        }
+    );
+
+    it('samples immediately after a wall-clock jump despite a frozen monotonic cache and coalesces requests', async () => {
+        const f = await fixture();
+        let wall = 10_000;
+        const monitor = createHostBootMonitor({ ...f, now: () => 0, wallNow: () => wall });
+        expect(await monitor.hasChanged()).toBe(false);
+        expect(await monitor.hasChanged()).toBe(false);
+        expect(f.requests).toHaveLength(1);
+        wall += 300_000;
+        f.setBoot('b'.repeat(64));
+        expect(await Promise.all([monitor.hasChanged(), monitor.hasChanged()])).toEqual([true, true]);
+        expect(f.requests).toHaveLength(2);
+        expect(new Set(f.requests).size).toBe(2);
+    });
+
+    it('does not block game startup when optional configuration or its key file is absent', async () => {
+        const warning = vi.spyOn(console, 'warn').mockImplementation(() => {});
+        expect(await loadHostBootMonitor({})).toBeUndefined();
+        expect(await loadHostBootMonitor({ TURN_HOST_BOOT_URL: 'http://host/v1/boot-id' })).toBeUndefined();
+        expect(
+            await loadHostBootMonitor({
+                TURN_HOST_BOOT_URL: 'http://host/v1/boot-id',
+                TURN_HOST_BOOT_KEY_FILE: '/nonexistent-host-boot-key',
+            })
+        ).toBeUndefined();
+        expect(() =>
+            createHostBootMonitor({ url: 'http://user:password@host/v1/boot-id', key: randomBytes(32) })
+        ).toThrow('without credentials');
+        warning.mockRestore();
+    });
+});

app/game-engine/test/playAuditDecisionCost.integration.test.ts

@@ -1,7 +1,12 @@
 import { writeFile } from 'node:fs/promises';
 import { performance } from 'node:perf_hooks';
 import { afterAll, beforeAll, describe, expect, it } from 'vitest';
-import { createGamePostgresConnector, type GamePrisma, type GamePrismaClient } from '@sammo-ts/infra';
+import {
+    decodeAuditDecisionSteps,
+    createGamePostgresConnector,
+    type GamePrisma,
+    type GamePrismaClient,
+} from '@sammo-ts/infra';
 import { persistAuditDecisions } from '../src/playAudit/decisionPersistence.js';
 import { buildAuditDecisionFixture } from './fixtures/playAuditDecision.js';
 
@@ -61,11 +66,15 @@ describe.skipIf(!databaseUrl)('decision storage cost probe', () => {
         >`
             SELECT
                 (SELECT count(*) FROM play_audit_decision) AS headers,
-                (SELECT sum(jsonb_array_length(steps)) FROM play_audit_decision_chunk) AS steps,
+                (SELECT sum(step_count) FROM play_audit_decision) AS steps,
                 (SELECT sum(pg_column_size(d)) FROM play_audit_decision d) AS "headerBytes",
                 (SELECT sum(pg_column_size(c)) FROM play_audit_decision_chunk c) AS "chunkBytes"
         `;
         expect(Number(storage!.steps)).toBe(201 * 302);
+        const chunks = await db.playAuditDecisionChunk.findMany({ select: { steps: true } });
+        expect(chunks.reduce((total, chunk) => total + decodeAuditDecisionSteps(chunk.steps).length, 0)).toBe(
+            201 * 302
+        );
         const relations = await db.$queryRaw<
             { name: string; tableBytes: bigint; indexBytes: bigint; totalBytes: bigint }[]
         >`
@@ -99,7 +108,8 @@ describe.skipIf(!databaseUrl)('decision storage cost probe', () => {
         `;
         const metrics = {
             scope: 'synthetic 201 decisions, 302 steps each; isolated schema, warm local reads',
-            limitations: 'Not production p95; no gameplay baseline, WAL or retained heap measurement. SQL counts cover only the persistence transaction. Repetitive synthetic steps compress well. Relation allocation can retain space from previous runs.',
+            limitations:
+                'Not production p95; no gameplay baseline, WAL or retained heap measurement. SQL counts cover only the persistence transaction. Repetitive synthetic steps compress well. Relation allocation can retain space from previous runs.',
             decisions: 201,
             steps: Number(storage!.steps),
             chunks: 603,

app/game-engine/test/playAuditDecisionPersistence.integration.test.ts

@@ -1,5 +1,5 @@
 import { afterAll, beforeAll, describe, expect, it } from 'vitest';
-import { createGamePostgresConnector, type GamePrismaClient } from '@sammo-ts/infra';
+import { decodeAuditDecisionSteps, createGamePostgresConnector, type GamePrismaClient } from '@sammo-ts/infra';
 import { persistAuditDecisions } from '../src/playAudit/decisionPersistence.js';
 import { buildAuditExecutionFixture, buildAuditDecisionFixture as draft } from './fixtures/playAuditDecision.js';
 import { prunePreviousAuditBatch } from '../src/playAudit/retention.js';
@@ -36,7 +36,40 @@ integration('decision persistence and bounded retention', () => {
     });
     it('rolls back gameplay/header/chunks on insert failure, retries and rejects divergent replay', async () => {
         const decision = draft('decision-one');
-        if (decision.steps[0]?.kind === 'DECISION_START') decision.steps[0].effectivePolicy = {"schemaVersion":1,"general":{"priority":["징병"],"flags":{"징병":true,"출병":false}},"nation":{"priority":["천도"],"flags":{"천도":true},"values":{"reqNationGold":4321,"reqNationRice":100,"reqHumanWarUrgentGold":100,"reqHumanWarUrgentRice":100,"reqHumanWarRecommandGold":100,"reqHumanWarRecommandRice":100,"reqHumanDevelGold":100,"reqHumanDevelRice":100,"reqNpcWarGold":100,"reqNpcWarRice":100,"reqNpcDevelGold":100,"reqNpcDevelRice":100,"minimumResourceActionAmount":100,"maximumResourceActionAmount":100,"minNpcWarLeadership":100,"minWarCrew":100,"minNpcRecruitCityPopulation":100,"safeRecruitCityPopulationRatio":100,"properWarTrainAtmos":100,"cureThreshold":100},"combatForce":{"1":[2,3]},"supportForce":[4],"developForce":[5]}};
+        if (decision.steps[0]?.kind === 'DECISION_START')
+            decision.steps[0].effectivePolicy = {
+                schemaVersion: 1,
+                general: { priority: ['징병'], flags: { 징병: true, 출병: false } },
+                nation: {
+                    priority: ['천도'],
+                    flags: { 천도: true },
+                    values: {
+                        reqNationGold: 4321,
+                        reqNationRice: 100,
+                        reqHumanWarUrgentGold: 100,
+                        reqHumanWarUrgentRice: 100,
+                        reqHumanWarRecommandGold: 100,
+                        reqHumanWarRecommandRice: 100,
+                        reqHumanDevelGold: 100,
+                        reqHumanDevelRice: 100,
+                        reqNpcWarGold: 100,
+                        reqNpcWarRice: 100,
+                        reqNpcDevelGold: 100,
+                        reqNpcDevelRice: 100,
+                        minimumResourceActionAmount: 100,
+                        maximumResourceActionAmount: 100,
+                        minNpcWarLeadership: 100,
+                        minWarCrew: 100,
+                        minNpcRecruitCityPopulation: 100,
+                        safeRecruitCityPopulationRatio: 100,
+                        properWarTrainAtmos: 100,
+                        cureThreshold: 100,
+                    },
+                    combatForce: { '1': [2, 3] },
+                    supportForce: [4],
+                    developForce: [5],
+                },
+            };
         decision.summary.codeVersion = 'a'.repeat(40);
         decision.summary.executionCoverage = 'ATTEMPTS';
         decision.steps.push({ ...decision.steps[0]!, ...buildAuditExecutionFixture(), sequence: 302 });
@@ -85,7 +118,7 @@ integration('decision persistence and bounded retention', () => {
             orderBy: { ordinal: 'asc' },
         });
         expect(chunks).toHaveLength(3);
-        expect(chunks.flatMap((chunk) => chunk.steps)).toEqual(decision.steps);
+        expect(chunks.flatMap((chunk) => decodeAuditDecisionSteps(chunk.steps))).toEqual(decision.steps);
         await expect(
             db.$transaction((tx) =>
                 persistAuditDecisions(tx, [{ ...decision, summary: { ...decision.summary, completed: true } }])
@@ -109,10 +142,34 @@ integration('decision persistence and bounded retention', () => {
         expect(await db.playAuditDecisionChunk.count({ where: { decisionId: 'decision-one' } })).toBe(204);
         expect(await db.playAuditDecision.count({ where: { id: 'decision-one' } })).toBe(1);
         expect(await prunePreviousAuditBatch(db, 'decision-current')).toEqual({ status: 'progress', deleted: 200 });
-        expect(await prunePreviousAuditBatch(db, 'decision-current')).toEqual({ status: 'progress', deleted: 4 });
+        expect(await prunePreviousAuditBatch(db, 'decision-current')).toEqual({ status: 'progress', deleted: 5 });
-        expect(await prunePreviousAuditBatch(db, 'decision-current')).toEqual({ status: 'progress', deleted: 1 });
         expect(await prunePreviousAuditBatch(db, 'decision-current')).toEqual({ status: 'complete', deleted: 0 });
         expect(await db.playAuditDecision.count()).toBe(1);
         expect(await db.playAuditDecisionChunk.count()).toBe(3);
     });
+    it('removes many small decisions in one bounded batch and rolls chunks back on header failure', async () => {
+        const decisions = Array.from({ length: 100 }, (_, index) => {
+            const row = draft(`batch-old-${index}`, 'batch-old');
+            return { ...row, steps: [row.steps[0]!, row.steps.at(-1)!] };
+        });
+        await db.$transaction((tx) => persistAuditDecisions(tx, decisions));
+        await db.$executeRawUnsafe(
+            `CREATE OR REPLACE FUNCTION decision_header_failure() RETURNS trigger LANGUAGE plpgsql AS $$ BEGIN RAISE EXCEPTION 'header fixture rollback'; END $$`
+        );
+        await db.$executeRawUnsafe(
+            `CREATE TRIGGER decision_header_failure BEFORE DELETE ON play_audit_decision FOR EACH ROW EXECUTE FUNCTION decision_header_failure()`
+        );
+        try {
+            await expect(prunePreviousAuditBatch(db, 'decision-current')).rejects.toThrow();
+            expect(
+                await db.playAuditDecisionChunk.count({ where: { decisionId: { in: decisions.map(({ id }) => id) } } })
+            ).toBe(100);
+        } finally {
+            await db.$executeRawUnsafe('DROP TRIGGER decision_header_failure ON play_audit_decision');
+            await db.$executeRawUnsafe('DROP FUNCTION decision_header_failure()');
+        }
+        expect(await prunePreviousAuditBatch(db, 'decision-current')).toEqual({ status: 'progress', deleted: 200 });
+        expect(await db.playAuditDecision.count()).toBe(1);
+        expect(await prunePreviousAuditBatch(db, 'decision-current')).toEqual({ status: 'complete', deleted: 0 });
+    });
 });

app/game-engine/test/playAuditRetentionWorker.test.ts

@@ -20,7 +20,7 @@ describe('audit retention scheduling', () => {
         expect(prune).toHaveBeenCalledTimes(1);
         await vi.advanceTimersByTimeAsync(30_001);
         expect(prune).toHaveBeenCalledTimes(3);
-        await vi.advanceTimersByTimeAsync(1000);
+        await vi.advanceTimersByTimeAsync(100);
         expect(prune).toHaveBeenCalledTimes(4);
         await vi.advanceTimersByTimeAsync(600_000);
         expect(prune).toHaveBeenCalledTimes(4);

app/game-engine/test/reservedTurnLease.integration.test.ts

@@ -4,7 +4,9 @@ import { createGamePostgresConnector } from '@sammo-ts/infra';
 import {
     InMemoryReservedTurnStore,
     ReservedTurnLeaseConflictError,
+    createReservedTurnStore,
 } from '../src/turn/reservedTurnStore.js';
+import { createReservedTurnLeaseRetry } from '../src/turn/reservedTurnLeaseRetry.js';
 
 const databaseUrl = process.env.RESERVED_TURN_DATABASE_URL;
 const describeIntegration = databaseUrl ? describe : describe.skip;
@@ -132,4 +134,98 @@ describeIntegration('reserved turn daemon/API lease integration', () => {
         });
         expect(after.map((turn) => turn.actionCode)).toEqual(before.map((turn) => turn.actionCode));
     });
+
+    it('cleans persisted ownership after rollback loses the journal, preserves other writers, and reloads once', async () => {
+        if (!connector || !databaseUrl) throw new Error('integration connector is unavailable');
+        await connector.prisma.generalTurnRevision.update({
+            where: { generalId: GENERAL_ID },
+            data: { leaseOwner: null, leaseExpiresAt: null },
+        });
+        const before = await connector.prisma.generalTurn.findMany({
+            where: { generalId: GENERAL_ID },
+            orderBy: { turnIdx: 'asc' },
+        });
+        const revision = await connector.prisma.generalTurnRevision.findUniqueOrThrow({
+            where: { generalId: GENERAL_ID },
+        });
+        const old = await createReservedTurnStore({ databaseUrl, leaseOwner: 'interrupted-daemon' });
+        const savepoint = old.store.captureTransactionState();
+        await old.store.prepareTurnsForExecution(GENERAL_ID);
+        old.store.shiftGeneralTurns(GENERAL_ID, -1);
+        old.store.restoreState(savepoint);
+        expect(old.store.getQueueCounts().leasedGeneralQueues).toBe(0);
+        const next = await createReservedTurnStore({ databaseUrl, leaseOwner: 'new-daemon' });
+        try {
+            await expect(next.store.prepareTurnsForExecution(GENERAL_ID)).rejects.toBeInstanceOf(
+                ReservedTurnLeaseConflictError
+            );
+            await old.close();
+            const released = await connector.prisma.generalTurnRevision.findUniqueOrThrow({
+                where: { generalId: GENERAL_ID },
+            });
+            expect(released).toMatchObject({ leaseOwner: null, leaseExpiresAt: null, revision: revision.revision });
+            await next.store.prepareTurnsForExecution(GENERAL_ID);
+            // A delayed close of an old owner cannot clear the new owner's lease.
+            const delayedCleaner = new InMemoryReservedTurnStore(connector.prisma, {
+                maxGeneralTurns: 2,
+                maxNationTurns: 1,
+                leaseOwner: 'interrupted-daemon',
+            });
+            await delayedCleaner.releaseOwnedLeases();
+            expect(
+                await connector.prisma.generalTurnRevision.findUniqueOrThrow({ where: { generalId: GENERAL_ID } })
+            ).toMatchObject({ leaseOwner: 'new-daemon', revision: revision.revision });
+            await next.store.releaseOwnedLeases();
+            const actual = await connector.prisma.generalTurn.findMany({
+                where: { generalId: GENERAL_ID },
+                orderBy: { turnIdx: 'asc' },
+            });
+            expect(actual.map((row) => row.actionCode)).toEqual(before.map((row) => row.actionCode));
+        } finally {
+            await next.close();
+        }
+    });
+
+    it('waits out an orphaned lease without deleting it and reloads an intervening API edit', async () => {
+        if (!connector) throw new Error('integration connector is unavailable');
+        await connector.prisma.generalTurnRevision.update({
+            where: { generalId: GENERAL_ID },
+            data: { leaseOwner: 'killed-daemon', leaseExpiresAt: new Date(Date.now() + 300_000) },
+        });
+        const store = new InMemoryReservedTurnStore(connector.prisma, {
+            maxGeneralTurns: 2,
+            maxNationTurns: 1,
+            leaseOwner: 'recovery-daemon',
+        });
+        await store.loadAll();
+        const retry = createReservedTurnLeaseRetry({
+            sleep: async () => {
+                // Simulate the recorded expiry, then an authorized API edit before the next claim.
+                expect(
+                    await connector.prisma.generalTurnRevision.findUniqueOrThrow({ where: { generalId: GENERAL_ID } })
+                ).toMatchObject({ leaseOwner: 'killed-daemon' });
+                await connector.prisma.generalTurnRevision.update({
+                    where: { generalId: GENERAL_ID },
+                    data: { leaseExpiresAt: new Date(0) },
+                });
+                await connector.prisma.generalTurn.updateMany({
+                    where: { generalId: GENERAL_ID, turnIdx: 0 },
+                    data: { actionCode: 'che_수련' },
+                });
+            },
+        });
+        try {
+            let failure: unknown;
+            try {
+                await store.prepareTurnsForExecution(GENERAL_ID);
+            } catch (error) {
+                failure = error;
+            }
+            expect(await retry.wait(failure)).toBe(true);
+            await store.prepareTurnsForExecution(GENERAL_ID);
+            expect(store.getGeneralTurns(GENERAL_ID)[0]?.action).toBe('che_수련');
+        } finally {
+            await store.releaseOwnedLeases();
+        }
+    });
 });

app/game-engine/test/reservedTurnLease.test.ts

@@ -169,6 +169,18 @@ const buildHarness = (initialRevision: RevisionRow | null = null) => {
 };
 
 describe('reserved turn daemon lease', () => {
+    it('cleans both persisted lease kinds even after their journal is rolled back', async () => {
+        const harness = buildHarness();
+        const snapshot = harness.store.captureTransactionState();
+        await harness.store.prepareTurnsForExecution(7, { nationId: 3, officerLevel: 12 });
+        harness.store.restoreState(snapshot);
+        expect(harness.store.getQueueCounts()).toMatchObject({ leasedGeneralQueues: 0, leasedNationQueues: 0 });
+        await harness.store.releaseOwnedLeases();
+        expect(harness.getRevision()).toMatchObject({ leaseOwner: null, leaseExpiresAt: null, revision: 0 });
+        expect(harness.getNationRevision()).toMatchObject({ leaseOwner: null, leaseExpiresAt: null, revision: 0 });
+        expect(harness.generalDeleteMany).not.toHaveBeenCalled();
+    });
+
     it('prunes deleted general and nation queues together with their journal state', () => {
         const harness = buildHarness();
         harness.store.ensureGeneralTurns(7);

app/game-engine/test/reservedTurnLeaseRetry.test.ts

@@ -0,0 +1,38 @@
+import { describe, expect, it, vi } from 'vitest';
+import { createReservedTurnLeaseRetry } from '../src/turn/reservedTurnLeaseRetry.js';
+import { ReservedTurnLeaseBusyError, ReservedTurnLeaseConflictError } from '../src/turn/reservedTurnStore.js';
+
+describe('reserved turn acquisition retry', () => {
+    it('bounds the entire contention episode across different queues and resets after progress', async () => {
+        let now = 0;
+        const onWait = vi.fn();
+        const retry = createReservedTurnLeaseRetry({
+            now: () => now,
+            sleep: async (ms) => {
+                now += ms;
+            },
+            maxWaitMs: 2_500,
+            onWait,
+        });
+        expect(await retry.wait(new ReservedTurnLeaseBusyError('general:1'))).toBe(true);
+        expect(await retry.wait(new ReservedTurnLeaseBusyError('nation:2:12'))).toBe(true);
+        expect(await retry.wait(new ReservedTurnLeaseBusyError('general:3'))).toBe(true);
+        expect(now).toBe(2_500);
+        expect(await retry.wait(new ReservedTurnLeaseBusyError('general:1'))).toBe(false);
+        expect(onWait).toHaveBeenCalledTimes(1);
+        retry.reset();
+        expect(await retry.wait(new ReservedTurnLeaseBusyError('general:1'))).toBe(true);
+    });
+
+    it('never retries flush ownership loss, calculation errors, or failed savepoint restoration', async () => {
+        const sleep = vi.fn();
+        const retry = createReservedTurnLeaseRetry({ sleep });
+        for (const error of [
+            new ReservedTurnLeaseConflictError('general:1'),
+            new Error('calculation failed'),
+            new AggregateError([new ReservedTurnLeaseBusyError('general:1')], 'restore failed'),
+        ])
+            expect(await retry.wait(error)).toBe(false);
+        expect(sleep).not.toHaveBeenCalled();
+    });
+});

app/game-engine/test/runtimePauseGate.test.ts

@@ -3,12 +3,36 @@ import type { GameClockPhase } from '@sammo-ts/common';
 import { createRuntimePauseGate } from '../src/turn/runtimePauseGate.js';
 
 describe('runtime pause clock boundary', () => {
+    it('awaits a host authority check before reading profile state or changing the clock', async () => {
+        const calls: string[] = [];
+        const gate = createRuntimePauseGate({
+            assertLease: async () => {
+                await Promise.resolve();
+                calls.push('host');
+                throw new Error('host changed');
+            },
+            shouldPause: async () => {
+                calls.push('profile');
+                return false;
+            },
+            isExplicitlyPaused: () => false,
+            getPhase: () => 'RUNNING',
+            prepareRecovery: async () => {
+                calls.push('recovery');
+            },
+            synchronize: async () => {},
+        });
+        await expect(gate()).rejects.toThrow('host changed');
+        expect(calls).toEqual(['host']);
+    });
     it('freezes a live error pause before commands, resumes without restart, and does not freeze twice', async () => {
         let phase: GameClockPhase = 'RUNNING';
         let paused = true;
         const calls: string[] = [];
         const gate = createRuntimePauseGate({
-            assertLease: () => calls.push('lease'),
+            assertLease: () => {
+                calls.push('lease');
+            },
             shouldPause: async () => paused,
             isExplicitlyPaused: () => paused,
             getPhase: () => phase,

app/game-engine/test/turnDaemonLease.integration.test.ts

@@ -207,6 +207,39 @@ integration('database turn daemon lease and fencing', () => {
         expect(row.ownerId).toBe(firstToken ? 'owner-a' : 'owner-b');
     });
 
+    it('fences a host boot change inside a transaction before mutation, without expiring the DB lease', async () => {
+        const profile = `${profilePrefix}host-boot`;
+        let changed = false;
+        const lease = await DatabaseTurnDaemonLease.connect(databaseUrl!, {
+            profile,
+            ownerId: 'saved-vm-owner',
+            heartbeat: false,
+            leaseDurationMs: 60_000,
+            hostBootMonitor: { hasChanged: async () => changed },
+        });
+        leases.push(lease);
+        await lease.acquire();
+        await lease.markClockReady();
+        changed = true;
+        let committed = false;
+        await expect(
+            db.$transaction(async (transaction) => {
+                await lease.assertActive(transaction);
+                committed = true;
+            })
+        ).rejects.toBeInstanceOf(TurnDaemonLeaseLostError);
+        expect(committed).toBe(false);
+        expect(lease.isLost()).toBe(true);
+        expect(await lease.renew()).toBe(false);
+        // The monitor does not force-clear ownership or bypass the successor's fencing.
+        expect((await db.turnDaemonLease.findUniqueOrThrow({ where: { profile } })).ownerId).toBe('saved-vm-owner');
+        const next = await createLease(profile, 'fresh-vm-owner');
+        expect(await next.acquire()).toBeNull();
+        await lease.release();
+        expect((await db.turnDaemonLease.findUniqueOrThrow({ where: { profile } })).clockReady).toBe(false);
+        expect((await next.acquire())?.fencingEpoch).toBe(2n);
+    });
+
     it('persists and evaluates lease wall time as UTC under a non-UTC database session', async () => {
         const profile = `${profilePrefix}utc-wall`;
         const zonedUrl = new URL(databaseUrl!);

app/game-engine/test/turnDaemonLifecycle.test.ts

@@ -1,6 +1,8 @@
 import { describe, expect, it, vi } from 'vitest';
 import { ManualClock } from '@sammo-ts/common';
 import { TurnDaemonLeaseLostError } from '../src/lifecycle/databaseTurnDaemonLease.js';
+import { ReservedTurnLeaseBusyError } from '../src/turn/reservedTurnStore.js';
+import { createReservedTurnLeaseRetry } from '../src/turn/reservedTurnLeaseRetry.js';
 
 import {
     InMemoryControlQueue,
@@ -15,6 +17,68 @@ import {
 const addMinutes = (time: Date, minutes: number): Date => new Date(time.getTime() + minutes * 60_000);
 
 describe('TurnDaemonLifecycle', () => {
+    it('rolls back a busy queue, checks the operator gate, and retries an explicit run exactly once', async () => {
+        const now = new Date('2026-11-12T00:00:00Z');
+        const queue = new InMemoryControlQueue();
+        queue.enqueue({ type: 'run', reason: 'manual', targetTime: now.toISOString() });
+        let resource = 10;
+        let runs = 0;
+        const stateManager = new EngineStateManager();
+        stateManager.register('resource', {
+            capture: () => resource,
+            restore: (value) => {
+                resource = value;
+            },
+        });
+        const onRunError = vi.fn();
+        const flushChanges = vi.fn();
+        const gate = vi.fn(async () => false);
+        const retry = createReservedTurnLeaseRetry({
+            sleep: async () => {
+                expect(resource).toBe(10);
+            },
+        });
+        const lifecycle = new TurnDaemonLifecycle(
+            {
+                clock: new ManualClock(now.getTime()),
+                controlQueue: queue,
+                stateManager,
+                runErrorRetry: retry,
+                getNextTickTime: (value) => addMinutes(value, 5),
+                pauseGate: gate,
+                stateStore: {
+                    loadLastTurnTime: async () => now,
+                    loadNextGeneralTurnTime: async () => null,
+                    saveLastTurnTime: async () => {},
+                    loadCheckpoint: async () => undefined,
+                    saveCheckpoint: async () => {},
+                },
+                processor: {
+                    run: async () => {
+                        resource -= 2;
+                        if (++runs === 1) throw new ReservedTurnLeaseBusyError('general:7');
+                        queue.enqueue({ type: 'shutdown' });
+                        return {
+                            lastTurnTime: now.toISOString(),
+                            processedGenerals: 1,
+                            processedTurns: 0,
+                            durationMs: 0,
+                            partial: false,
+                        };
+                    },
+                },
+                hooks: { onRunError, flushChanges },
+            },
+            { profile: 'che', defaultBudget: { budgetMs: 100, maxGenerals: 1, catchUpCap: 1 } }
+        );
+        await lifecycle.start();
+        expect(runs).toBe(2);
+        expect(resource).toBe(8);
+        expect(gate.mock.calls.length).toBeGreaterThanOrEqual(2);
+        expect(flushChanges).toHaveBeenCalledOnce();
+        expect(onRunError).not.toHaveBeenCalled();
+    });
+
     it.each([false, true])('reports a fatal lease gate and exits even if reporting fails (%s)', async (reportFails) => {
         const now = new Date('2026-09-09T17:30:00Z');
         const error = new TurnDaemonLeaseLostError('che:default');

docs/design/play-audit.md

@@ -2,6 +2,9 @@
 
 ## 문서 상태와 사용법
 
+2026-11-09 저장 최적화: [물리 압축·정리 처리량·ZFS 운영 경계](../play-audit-storage.md).
+판단 원문과 현재 기수 보존 계약을 유지하며 저장 표현과 이전 기수 정리 구현을 변경한다.
+
 이 문서는 플레이 감사의 구조·현재 구현 경계·남은 요구사항을 관리하는 기준 문서다.
 사용 방법과 DB 적용은 [운영 안내](../play-audit-operations.md)를 따른다.
 날짜별 구현·측정 기록은 Git 이력과 상위 작업공간의 `report/`에서 확인한다.

docs/developer/host-boot-recovery.md

@@ -0,0 +1,149 @@
+# Hyper-V 저장·복원 이후 게임 복구
+
+Windows 재부팅 때 Linux VM이 저장·복원되면 게스트의 shutdown/resume hook이
+실행되지 않을 수 있다. 게임은 종료 이벤트 없이도 기존 lease 만료와 DB snapshot
+재적재로 복구한다. Windows 식별자 확인은 이 경로에 추가하는 선택 기능이다.
+
+## 예약턴 임대 복구
+
+- 실행 루프가 종료되면 DB에서 자기 owner의 장수·국가 예약턴 임대를 해제한다.
+  In-memory rollback으로 leased set에서 빠진 row도 정리한다. 큐 내용과 revision,
+  다른 실행자/API의 임대는 변경하지 않는다.
+- 계산 전 임대 획득 충돌은 전체 메모리 savepoint를 복원한 뒤 1초 단위로 재시도한다.
+  한 충돌 구간의 상한은 301초다(기본 queue lease 5분 + 여유 1초). 그동안 다음
+  loop마다 운영자 pause/stop과 daemon 권위를 확인한다. 성공한 batch에서 상한을
+  초기화하며 큐가 바뀌어도 같은 실패 구간의 상한을 연장하지 않는다.
+- 저장·finalize 단계의 소유권 상실, 메모리 복원 실패와 다른 게임 오류는 자동
+  재시도하지 않는다. 상한을 넘은 획득 충돌도 기존 PAUSED/장애 이력 경로로 간다.
+- daemon의 30초 lease와 fencing epoch, 기존 10분 이하 catch-up 및 장시간 복구
+  정책은 유지한다. 운영자가 정지한 profile을 자동 START하지 않는다.
+
+## Windows Node.js 에이전트
+
+필요 파일은 `windows-boot-agent.mjs` 하나이며 외부 npm package가 필요 없다.
+Node.js 24.22.1 이상을 사용한다. 설정 예시는
+`tools/host-recovery/windows-boot-agent.example.json`이다.
+
+Node.js가 Windows 내장 PowerShell의 CIM 조회로 `Win32_OperatingSystem.LastBootUpTime`
+UTC ticks를 읽고 hostname과 함께 hash한다. 별도의 PowerShell 스크립트 설치나
+Linux에서 Windows 관리자 명령 실행은 필요 없다. Node 에이전트를 재시작해도
+같은 Windows boot의 식별자는 같고, Windows를 재부팅하면 달라진다.
+실제 Windows에서 CIM 조회가 실패하면 가짜 식별자를 만들지 않고 시작을 거부한다.
+
+1. 파일들을 `C:\ProgramData\Sammo` 같은 고정 경로에 복사한다.
+2. 디렉터리를 만든 뒤 Windows 터미널에서 private key를 생성한다. 기존 파일을
+   덮어쓰지 않으며 key 내용은 출력하지 않는다.
+
+    ```text
+    node C:\ProgramData\Sammo\windows-boot-agent.mjs --create-key C:\ProgramData\Sammo\host-boot.key
+    ```
+
+3. 설정 예시를 `C:\ProgramData\Sammo\host-boot.json`으로 복사한다. `keyFile`은
+   위 경로, `bindAddress`는 **Linux VM에서 접근할 수 있는 Windows의 사설 주소**로
+   설정한다. 예시의 `127.0.0.1`은 안전한 기본값이며 원격 Linux에서 접근할 수 없다.
+   `port` 기본값은 17781이다. Windows 방화벽은 해당 사설 listener에 Linux VM
+   주소에서 오는 TCP 연결만 허용하도록 설정한다. 기존 공개 Caddy는 변경하지 않는다.
+4. key 파일은 실행 계정만 읽도록 NTFS 권한을 제한하고, 같은 내용을 Linux의
+   `/run/secrets/host_boot_key` 같은 private file로 비공개 전달한다. Git·메신저·
+   CLI 인자에 key 내용을 넣지 않는다. 에이전트는 key를 시작할 때 읽으므로 key
+   교체 시 양쪽 프로세스를 함께 재시작한다.
+5. 실행한다.
+
+    ```text
+    node C:\ProgramData\Sammo\windows-boot-agent.mjs --config C:\ProgramData\Sammo\host-boot.json
+    ```
+
+6. Windows 작업 스케줄러에 **시스템 시작 시** task를 등록한다. 프로그램은 Node의
+   절대 `node.exe` 경로, 인자는 위 `.mjs --config ...` 부분을 사용한다. 로그인 없이
+   실행되고 실패 시 재시작하도록 구성한다. 설정 파일을 사용하므로 터미널 세션의
+   임시 환경변수에 의존하지 않는다. 실제 Windows 설치·방화벽 변경·task 등록은
+   대상 host에서 별도 적용하고 확인해야 한다.
+
+원한다면 설정 파일 대신 `HOST_BOOT_KEY_FILE`, `HOST_BOOT_BIND_ADDRESS`,
+`HOST_BOOT_PORT` 환경변수로도 실행할 수 있다.
+
+## Linux 연결
+
+기존 game runtime/PM2가 상속하는 서버 환경에 아래 두 값을 함께 설정한다.
+설정이 없으면 기능을 사용하지 않는다. 설정이 불완전하거나 key 파일이 없거나 잘못돼도
+고정 경고와 함께 보조 관측만 비활성화하며 게임 기동은 계속한다.
+
+```dotenv
+TURN_HOST_BOOT_URL=http://windows-private-address:17781/v1/boot-id
+TURN_HOST_BOOT_KEY_FILE=/run/secrets/host_boot_key
+```
+
+Docker에서는 private key가 들어 있는 전용 디렉터리를 runtime에 read-only mount하고 두 환경변수를
+**runtime container environment**에 전달해야 한다. 호스트의 `.env`에만 추가하면
+자동 전달되지 않는다. 기존 Compose와 host overlay를 함께 적용하는 전용 wrapper를
+사용한다. 실제 secret 파일의 host 경로와 Windows 사설 주소를 확인하기 전에는
+운영 stack 설정을 추측해 바꾸지 않는다.
+
+예를 들어 host overlay의 runtime 설정에는 다음 형태로 연결할 수 있다. host 경로는
+placeholder이며 실제 운영 경로로 교체해야 한다.
+
+```yaml
+services:
+    runtime:
+        environment:
+      TURN_HOST_BOOT_URL: ${TURN_HOST_BOOT_URL:-}
+      TURN_HOST_BOOT_KEY_FILE: /run/sammo-host-recovery/host_boot.key
+        volumes:
+            - type: bind
+        source: /absolute/private/path/host-recovery
+        target: /run/sammo-host-recovery
+              read_only: true
+              bind:
+          create_host_path: true
+```
+
+위 디렉터리는 보조 기능 전용이므로 없으면 빈 디렉터리를 만들 수 있다. 파일 부재는
+관측만 비활성화한다. DB·backup 같은 필수 저장소에는 이 fallback을 적용하지 않는다.
+운영에는 이 전용 디렉터리의 `host_boot.key` 파일을 전달한다.
+
+처음 인증에 성공한 식별자를 프로세스의 기준으로 삼고 고정된 성공 로그를 남긴다. 이후 다른 식별자를
+확인하면 `TurnDaemonLeaseLostError`로 기존 실행자를 중단하고 DB를 다시 읽는다.
+종료 시 owner/epoch가 여전히 자기 것인 경우에만 DB lease와 `clock_ready`를
+해제하므로 다른 실행자를 중지시키지 않는다. 재시작은 기존 profile 상태를
+보존하며 PAUSED/STOPPED를 임의로 RUNNING으로 바꾸지 않는다.
+
+보통 최대 2초 간격으로 확인한다. 벽시계와 monotonic 경과가 1초 이상 어긋나면
+cache를 무시하고 다시 조회한다. 조회 timeout은 1초다. 네트워크·인증·응답 오류는
+분당 한 번의 고정 경고만 남기고 기존 DB lease 복구를 유지한다. 처음 인증받기 전
+발생한 reboot는 이 신호만으로 확정할 수 없으며 기존 복구 경로가 담당한다.
+
+## 인증 프로토콜과 검증
+
+key 파일은 32바이트 난수를 lowercase hex 64자로 저장한다. 매 요청마다 새 32바이트
+challenge를 만들며, key 자체를 네트워크에 보내지 않는다.
+
+- GET `/v1/boot-id`, query string 없음.
+- `x-sammo-challenge`: lowercase hex 64자.
+- `x-sammo-signature`: `HMAC-SHA256(key, "sammo-host-boot-request-v1\n" + challenge)`.
+- JSON: `version=1`, `challenge`, `bootId`(hex 64자), `signature`.
+- 응답 signature:
+  `HMAC-SHA256(key, "sammo-host-boot-response-v1\n" + challenge + "\n" + bootId)`.
+
+위 `\n`은 실제 LF다. Linux는 challenge와 응답 signature를 검증하고, redirect와
+2048바이트 초과 응답을 거부한다. 서명된 옛 응답도 새 challenge에는 유효하지 않다.
+시각 정확도나 요청 시각 만료에 의존하지 않는다. 이 프로토콜은 인증과 무결성을
+제공하며 암호화는 하지 않는다. 사설 망에서만 사용하며 공개 무인증 재개 API를
+만들지 않는다.
+
+```sh
+node --test tools/host-recovery/windows-boot-agent.test.mjs
+pnpm --filter @sammo-ts/game-engine test hostBootMonitor.test.ts runtimePauseGate.test.ts reservedTurnLeaseRetry.test.ts
+```
+
+CI에서는 Windows boot를 모사한 HTTP 응답, 위조·재사용·timeout, 실제 PostgreSQL
+lease fencing과 queue cleanup을 검증한다. 실제 Windows CIM·Task Scheduler와
+Hyper-V 재부팅 검증을 대체하지 않는다. 배포 후 기존 장애 진단 화면의 lease/
+epoch와 public lobby `turnEngineRunning`, 실제 tick/queue revision 진행을 확인한다.
+
+기능을 해제하려면 Linux의 두 환경변수를 함께 제거하고 해당 runtime을 정상
+재기동한다. Windows task도 중지할 수 있다. DB migration이나 게임 초기화는 없다.
+
+근거: [Microsoft LastBootUpTime](https://learn.microsoft.com/en-us/windows/win32/cimwin32prov/win32-operatingsystem),
+[Node.js HMAC와 timingSafeEqual](https://nodejs.org/api/crypto.html),
+[게임 시계 복구 계약](../architecture/game-clock-reconciliation.md),
+[릴리스 운영](../release-operations.md).

docs/play-audit-operations.md

@@ -5,6 +5,9 @@
 관리한다. 아래 적용 절차는 각 프로필의 DB 보존 업데이트를 위한 것이며,
 문서에 기능이 기재되어 있다고 운영 DB에 적용된 것은 아니다.
 
+저장량 최적화, 중지된 프로필 정리, ZFS 이전과 rollback은
+[저장량과 운영 정리](play-audit-storage.md)를 따른다.
+
 ## 지금 사용할 수 있는 기능
 
 | 화면     | 사용할 수 있는 정보                                                                                                               | 읽을 때 주의할 점                                                            |
@@ -55,7 +58,7 @@
 ## DB migration과 적용 순서
 
 정식 game migration에 감사 테이블과 인덱스가 포함되어 있다. `prisma db push`나
-수동 CREATE TABLE로 대신 적용하지 않는다. 현재 game chain은58개이며 다음 감사
+수동 CREATE TABLE로 대신 적용하지 않는다. 현재 game chain은59개이며 다음 감사
 migration들을 포함한다. 기존 기록을 삭제하거나 지난달 상세를 역산하지 않는다.
 
 | migration                                             | 준비되는 저장소/제약                                                  |
@@ -69,6 +72,7 @@ migration들을 포함한다. 기존 기록을 삭제하거나 지난달 상세
 | `20260916060000_widen_play_audit_ticks`               | 월 표본/정책 tick을 BIGINT로 확장하여 약60개월 이후 INTEGER 초과 방지 |
 | `20260916070000_add_play_audit_decision`              | NPC/자동턴 결정 요약과 순서별 상세 chunk, bounded 정리용 FK/index     |
 | `20260916080000_index_play_audit_decision_month`      | 기존 장수 인덱스를 월 조건 포함 인덱스로 교체하여 기수 전체 조회 방지 |
+| `20261109000000_compact_play_audit_storage`           | 무손실 압축 chunk 형식과 작은 JSON inline 압축 설정                   |
 
 운영은 [릴리스 절차](release-operations.md)의 **DB 보존 버전 업데이트**로 해당 고정
 commit을 적용한다. 이 기능을 켜기 위해 시나리오를 초기화할 필요는 없다. 수동 환경의

docs/play-audit-storage.md

@@ -0,0 +1,76 @@
+# 플레이 감사 저장량과 운영 정리
+
+2026-11-09 저장 최적화는 Core 전용 감사의 물리 표현과 정리 구현을 바꾼다.
+현재 기수 전체 기록·판단 순서·정책 참조·hash·API 응답·권한은 유지한다.
+원문 샘플링이나 보존기간 축소를 하지 않는다.
+
+## 물리 표현
+
+- 결정 chunk는 기존 `steps` JSONB 배열 또는 `{encoding:"gzip-json-v1",data:"base64"}`로 저장한다.
+  128 step 단위는 그대로이며 gzip level 3을 사용한다. 압축 envelope가 더 크면 원래 배열을 쓴다.
+- replay hash는 압축 전 원문으로 계산한다. API는 공용 decoder 뒤 기존 step schema를 검증한다.
+  손상·알 수 없는 버전은 조회 오류이며 빈 자료로 바꾸지 않는다.
+- 압축 해제는 chunk당 16MiB로 제한한다. 더 큰 원문은 기존 배열로 저장해 잘라내지 않는다.
+- 장수 snapshot `data`와 결정 `summary`는 PostgreSQL `STORAGE MAIN`,
+  `toast_tuple_target=512`로 작은 반복 JSON도 inline 압축한다. JSON 값은 바뀌지 않는다.
+- migration `20261109000000_compact_play_audit_storage`는 기존 테이블을 재작성하지 않는다.
+  chunk CHECK는 새 쓰기에 즉시 적용하고 대형 과거 자료 검사는 `NOT VALID`로 미룬다.
+  정리 후 `ALTER TABLE ... VALIDATE CONSTRAINT play_audit_decision_chunk_steps_check`를 실행한다.
+  기존 자료의 압축 변경은 별도 bounded backfill 또는 검증된 export/restore가 필요하다.
+
+## 정리 처리량과 실행 수명주기
+
+- 이전 기수 결정 부모를 기존 `(server_id,nation_id,tick,id)` 인덱스로 최대 200개 잠근다.
+  한 transaction의 chunk+header 삭제 합도 200행 이하이며 큰 결정 하나가 있어도 상한을 지킨다.
+- 다음 성공 batch까지 100ms를 기다린다. RESET과 동일 advisory lock, identity 재검사,
+  2초 statement/250ms lock 제한, rollback을 유지한다. 현재 기수의 결정과 표본은 삭제하지 않는다.
+- API outbox는 전달 후 24시간을 보존한다. 기존 인덱스 순서로 최대 500행씩 정리하고,
+  backlog가 있으면 다음 dispatch에서도 이어간다. 한 호출은 최대 10 batch/50ms이며
+  미전달 자료는 삭제하지 않는다. 빈 상태에서는 60초 주기로 돌아간다.
+- game API/daemon이 중지된 프로필도 host scheduler가 다음 CLI로 정리한다.
+  CLI는 접속 URL이나 SQL 원문을 출력하지 않는다. 기본은 dry-run이다.
+
+```sh
+# 먼저 common/logic/infra/game-engine을 빌드한다. 접속 정보는 기존 process 환경으로 전달한다.
+node tools/play-audit-maintenance.mjs --schemas=che,kwe,pwe,twe,nya,pya,hwe
+node tools/play-audit-maintenance.mjs --schemas=che,kwe,pwe,twe,nya,pya,hwe --budget-ms=3000 --apply
+```
+
+schema마다 최신 identity를 읽고 각 삭제 transaction에서 다시 확인한다. 예상 identity가 바뀌면
+그 schema의 처리를 중단한다. host timer는 다음 실행에서 새 identity를 읽는다.
+장수/국가/명령/input_event 같은 gameplay 테이블은 이 CLI의 삭제 대상이 아니다.
+
+## ZFS 운영 경계
+
+호스트는 감사 전용 tablespace와 backup 전용 dataset을 별도로 제공한다.
+hot gameplay, WAL, legacy archive를 임의로 이동하지 않는다. DB와 tablespace는 한 cluster이므로
+tablespace 파일만 백업하는 것으로 복구를 보장하지 않는다.
+
+1. 전체 DB의 일관된 압축 backup을 ZFS에 만들고 checksum, archive 읽기, 실제 복원을 확인한다.
+2. DB 보존 DEPLOY로 writer/reader/migration을 함께 적용한다.
+3. profile 상태를 기록하고 공식 STOP으로 쓰기를 중단한다. 원래 STOPPED인 profile은 유지한다.
+   가오픈까지 1시간 이상 남은 RESERVED는 runtimeExpected=false이므로 예약을 유지할 수 있다.
+   실제 process 비기동과 identity/hash 안정성을 확인하고 export의 예상 상태를 RESERVED로 고정한다.
+   유지보수 중 RESET 등으로 identity가 바뀌면 해당 profile을 새로 export/검증하며 이전 자료를 덮어쓰지 않는다.
+4. 각 schema에서 현재 serverId의 8개 감사 테이블을 일관되게 export한다.
+   월 child는 현재 월 header와 join하고 결정 chunk는 현재 결정 header와 join한다.
+5. 격리 DB에서 복원·건수·내용 hash를 검증한다. 이전 기수는 전체 ZFS backup에 남는다.
+6. 하나의 transaction에서 감사 테이블만 TRUNCATE하고 ZFS tablespace로 옮긴 뒤
+   현재 기수 export를 다시 넣는다. 부모를 먼저 넣고 모든 건수/hash와 identity를 확인한 뒤 commit한다.
+   index도 별도로 ZFS로 옮긴다. 실패하면 전체 transaction을 rollback한다.
+7. CHECK 검증과 ANALYZE를 수행하고 원래 RUNNING이었던 profile만 공식 START한다.
+   현재 기록, 새 기록의 decode/hash, lease/clock, 공개 HTTPS와 루트/ZFS 실제 사용량을 확인한다.
+
+기존 데이터가 적을 때는 bounded DELETE 뒤 별도 재작성으로 공간을 회수할 수도 있다.
+수천만 건 backlog에서 export/TRUNCATE/restore는 현재 자료를 보존하며 물리 공간을 즉시
+회수하기 위한 계획된 유지보수 절차다. backup 검증이나 profile 정지가 빠진 자동 작업으로 쓰지 않는다.
+ZFS dataset이 없을 때 루트 디렉터리에 쓰지 않도록 host mount guard와 재현 가능한 Compose 설정이 필요하다.
+
+## rollback
+
+이전 reader는 압축 envelope를 이해하지 못한다. 구형 버전으로 돌아가기 전 해당 기록을
+새 decoder로 원래 JSON 배열로 변환하고 같은 hash/개수를 검증한다. 물리 압축 설정은
+이전 애플리케이션과 호환되며 원래 TOAST 설정으로 돌릴 필요가 없다.
+현재 기수 복원 중 실패하면 transaction rollback으로 기존 table/data/tablespace를 보존한다.
+기수 전체 과거 자료가 필요하면 ZFS의 변경 전 backup을 별도 DB에 복원한다.
+운영 DB 전체를 과거 dump로 덮어써 새 gameplay를 잃게 하지 않는다.

docs/release-operations.md

@@ -376,6 +376,18 @@ Hyper-V의 VM 정지/재개, 체크포인트 또는 host 시간 보정은 게스
 NTP 로그와 Windows Hyper-V/백업 작업 이력을 대조합니다. lease 기간을 늘리거나
 시간 동기화를 임의로 끄는 것으로 해결하지 않습니다.
 
+예약턴의 실행 전 임대 충돌은 메모리 rollback 뒤 1초 간격, 한 구간 최대 301초 동안
+해제·만료를 기다립니다. flush/finalize의 ownership 상실은 재시도하지 않습니다.
+종료 시에는 rollback으로 journal에서 사라진 자기 owner의 queue 임대도 DB에서
+해제합니다. 다른 owner와 queue 내용/revision은 보존합니다. 대기 중에도 매 loop의
+운영자 pause/stop과 daemon gate를 통과하며 상한을 넘으면 기존 오류 정지를 유지합니다.
+
+Windows의 저장·복원으로 guest 이벤트가 없을 때를 위한 선택적 boot identity 확인은
+[Windows Node.js 에이전트 설치와 Linux 연결](./developer/host-boot-recovery.md)을 따릅니다.
+`TURN_HOST_BOOT_URL`과 private `TURN_HOST_BOOT_KEY_FILE`을 함께 설정하면 인증된
+식별자 변경을 기존 lease 상실/재기동으로 처리합니다. 통신 실패는 기존 복구를
+막지 않으며 운영자의 PAUSED/STOPPED를 자동 해제하지 않습니다.
+
 Turn daemon의 DB persistence interactive transaction은 연결 획득 대기(`maxWait`)
 10초와 callback 실행(`timeout`) 30초를 각각 제한합니다. Prisma 기본 연결 대기
 2초는 일시적인 DB 지연에도 callback 시작 전 `P2028`로 서버를 PAUSED로 만들 수

packages/infra/prisma/migrations/20261109000000_compact_play_audit_storage/migration.sql

@@ -0,0 +1,18 @@
+-- 대형 기존 기록을 migration 중 재작성/전체 검사하지 않는다. 이전 배열과 새 압축 형식을 함께 허용한다.
+BEGIN;
+SET LOCAL lock_timeout = '2s';
+ALTER TABLE "play_audit_decision_chunk" DROP CONSTRAINT "play_audit_decision_chunk_steps_check";
+ALTER TABLE "play_audit_decision_chunk" ADD CONSTRAINT "play_audit_decision_chunk_steps_check"
+    CHECK (COALESCE(
+        jsonb_typeof("steps") = 'array' OR (
+            jsonb_typeof("steps") = 'object'
+            AND "steps"->>'encoding' = 'gzip-json-v1'
+            AND jsonb_typeof("steps"->'data') = 'string'
+        ), false
+    )) NOT VALID;
+-- 기본 약 2KiB 문턱 아래의 반복 JSON도 압축하되 작은 snapshot은 TOAST 밖으로 쪼개지 않는다.
+ALTER TABLE "play_audit_general" ALTER COLUMN "data" SET STORAGE MAIN;
+ALTER TABLE "play_audit_general" SET (toast_tuple_target = 512);
+ALTER TABLE "play_audit_decision" ALTER COLUMN "summary" SET STORAGE MAIN;
+ALTER TABLE "play_audit_decision" SET (toast_tuple_target = 512);
+COMMIT;

packages/infra/src/index.ts

@@ -15,3 +15,4 @@ export * from './playAuditDiplomacy.js';
 export * from './messageEnvelope.js';
 export * from './webPushOutbox.js';
 export * from './playAuditBestEffort.js';
+export * from './playAuditDecisionCodec.js';

packages/infra/src/playAuditDecisionCodec.ts

@@ -0,0 +1,44 @@
+import { gzipSync, gunzipSync } from 'node:zlib';
+import type { InputJsonValue } from './turnEngineDb.js';
+
+const ENCODING = 'gzip-json-v1';
+const MAX_STEPS = 128;
+const MAX_BYTES = 16 * 1024 * 1024;
+
+/** 조회/재실행 hash에 쓰는 원문은 바꾸지 않고 저장할 때만 무손실 압축한다. */
+export const encodeAuditDecisionSteps = (steps: readonly InputJsonValue[]): InputJsonValue => {
+    if (steps.length > MAX_STEPS) throw new Error('Audit decision chunk exceeds its step limit');
+    const json = JSON.stringify(steps);
+    // 드문 초대형 기존 payload도 버리거나 자르지 않는다. 압축 reader 상한 밖이면 원래 형식으로 저장한다.
+    if (Buffer.byteLength(json) > MAX_BYTES) return steps.slice();
+    const encoded = { encoding: ENCODING, data: gzipSync(json, { level: 3 }).toString('base64') };
+    return Buffer.byteLength(JSON.stringify(encoded)) < Buffer.byteLength(json) ? encoded : steps.slice();
+};
+
+/** 이전 JSONB 배열과 압축 형식을 함께 읽는다. 손상/알 수 없는 버전은 빈 기록으로 숨기지 않는다. */
+export const decodeAuditDecisionSteps = (value: unknown): unknown[] => {
+    if (Array.isArray(value)) {
+        if (value.length > MAX_STEPS) throw new Error('Audit decision chunk exceeds its step limit');
+        return value;
+    }
+    if (!value || typeof value !== 'object' || !('encoding' in value) || !('data' in value))
+        throw new Error('Invalid audit decision storage');
+    if (
+        value.encoding !== ENCODING ||
+        typeof value.data !== 'string' ||
+        value.data.length > Math.ceil(MAX_BYTES / 3) * 4 + 256 ||
+        !/^(?:[A-Za-z0-9+/]{4})*(?:[A-Za-z0-9+/]{2}==|[A-Za-z0-9+/]{3}=)?$/.test(value.data)
+    )
+        throw new Error('Invalid audit decision encoding');
+    let decoded: unknown;
+    try {
+        decoded = JSON.parse(
+            gunzipSync(Buffer.from(value.data, 'base64'), { maxOutputLength: MAX_BYTES }).toString('utf8')
+        );
+    } catch {
+        // DB 원문이나 압축 payload가 오류 응답/로그에 포함되지 않게 한다.
+        throw new Error('Corrupt audit decision storage');
+    }
+    if (!Array.isArray(decoded) || decoded.length > MAX_STEPS) throw new Error('Invalid audit decision steps');
+    return decoded;
+};

packages/infra/src/readModelOutboxDispatcher.ts

@@ -195,8 +195,9 @@ export const pruneDeliveredReadModelOutbox = async (
             SELECT "id"
             FROM "read_model_outbox"
             WHERE "delivered_at" < ${input.deliveredBefore}
-            ORDER BY "id"
+            ORDER BY "delivered_at", "available_at", "id"
             LIMIT ${limit}
+            FOR UPDATE SKIP LOCKED
         )
         RETURNING "id"
     `);

packages/infra/test/playAuditDecisionCodec.test.ts

@@ -0,0 +1,54 @@
+import { gzipSync } from 'node:zlib';
+import { describe, expect, it } from 'vitest';
+import { decodeAuditDecisionSteps, encodeAuditDecisionSteps } from '../src/playAuditDecisionCodec.js';
+
+describe('lossless audit decision storage', () => {
+    it('compresses repeated context while preserving every value and its order', () => {
+        const steps = Array.from({ length: 128 }, (_, sequence) => ({
+            sequence,
+            phase: 'general',
+            generalId: 51,
+            nationId: 3,
+            cityId: 12,
+            year: 190,
+            month: 7,
+            kind: 'RNG',
+            value: sequence / 128,
+            text: '판단·줄바꿈\n따옴표 " 그대로',
+            empty: null,
+        }));
+        const stored = encodeAuditDecisionSteps(steps);
+        expect(Array.isArray(stored)).toBe(false);
+        expect(Buffer.byteLength(JSON.stringify(stored))).toBeLessThan(Buffer.byteLength(JSON.stringify(steps)) / 3);
+        expect(decodeAuditDecisionSteps(stored)).toEqual(steps);
+    });
+    it('keeps tiny and historical uncompressed arrays readable', () => {
+        expect(encodeAuditDecisionSteps([])).toEqual([]);
+        expect(decodeAuditDecisionSteps([{ marker: 'legacy', sequence: 0 }])).toEqual([
+            { marker: 'legacy', sequence: 0 },
+        ]);
+    });
+    it('rejects unknown versions, bad base64, truncated gzip and invalid contents', () => {
+        for (const value of [
+            null,
+            {},
+            { encoding: 'gzip-json-v2', data: '' },
+            { encoding: 'gzip-json-v1', data: '%%%' },
+            { encoding: 'gzip-json-v1', data: 'YQ==' },
+            { encoding: 'gzip-json-v1', data: gzipSync('{}').toString('base64') },
+            { encoding: 'gzip-json-v1', data: gzipSync(JSON.stringify(Array(129).fill(null))).toString('base64') },
+        ])
+            expect(() => decodeAuditDecisionSteps(value)).toThrow();
+        expect(() => encodeAuditDecisionSteps(Array(129).fill(null))).toThrow();
+    });
+    it('bounds decompression and preserves oversized raw records rather than truncating them', () => {
+        const steps = [{ text: 'x'.repeat(16 * 1024 * 1024) }];
+        expect(encodeAuditDecisionSteps(steps)).toEqual(steps);
+        expect(() =>
+            decodeAuditDecisionSteps({
+                encoding: 'gzip-json-v1',
+                data: gzipSync(JSON.stringify(steps)).toString('base64'),
+            })
+        ).toThrow('Corrupt');
+    });
+});

release-manifest.json

@@ -2,6 +2,6 @@
     "formatVersion": 1,
     "controllerProtocol": 2,
     "gatewaySchemaHead": "20261107000000_kakao_identity_link_history",
-    "gameSchemaHead": "20260916080000_index_play_audit_decision_month",
+    "gameSchemaHead": "20261109000000_compact_play_audit_storage",
     "components": ["gateway-api", "gateway-frontend", "release-controller", "game-api", "game-engine", "game-frontend"]
 }

tools/host-recovery/README.md

@@ -0,0 +1,12 @@
+# Windows host boot identity agent
+
+`windows-boot-agent.mjs`는 외부 npm package 없이 실행하는 Node.js 에이전트다.
+Windows OS의 부팅 시각 기반 식별자를 인증된 Linux 게임 데몬에 제공한다.
+
+설치, key 생성, 자동 시작과 선택 기능 장애 시 동작은
+[Hyper-V 복구 운영 안내](../../docs/developer/host-boot-recovery.md)를 따른다.
+설정 예시는 `windows-boot-agent.example.json`이며 실제 key와 설정은 Git에 넣지 않는다.
+
+```sh
+node --test tools/host-recovery/windows-boot-agent.test.mjs
+```

tools/host-recovery/windows-boot-agent.example.json

@@ -0,0 +1,5 @@
+{
+    "keyFile": "C:\\ProgramData\\Sammo\\host-boot.key",
+    "bindAddress": "127.0.0.1",
+    "port": 17781
+}

tools/host-recovery/windows-boot-agent.mjs

@@ -0,0 +1,131 @@
+import { execFile } from 'node:child_process';
+import { createHash, createHmac, randomBytes, timingSafeEqual } from 'node:crypto';
+import { readFile, writeFile } from 'node:fs/promises';
+import { createServer } from 'node:http';
+import { hostname } from 'node:os';
+import { win32, resolve } from 'node:path';
+import { fileURLToPath } from 'node:url';
+import { promisify } from 'node:util';
+
+const mac = (key, value) => createHmac('sha256', key).update(value).digest('hex');
+
+export const bootIdentity = (host, ticks) => {
+    if (!/^\d{17,20}$/.test(ticks)) throw new Error('Windows returned an invalid boot time.');
+    return createHash('sha256').update(`${host.toLowerCase()}\n${ticks}`).digest('hex');
+};
+
+export const readWindowsBootIdentity = async () => {
+    if (process.platform !== 'win32') throw new Error('Run the host boot agent on Windows.');
+    const executable = win32.join(
+        process.env.SystemRoot ?? 'C:\\Windows',
+        'System32',
+        'WindowsPowerShell',
+        'v1.0',
+        'powershell.exe'
+    );
+    // Node.js가 OS 내장 CIM 조회를 실행한다. 별도 PowerShell 스크립트/모듈은 필요 없다.
+    const { stdout } = await promisify(execFile)(
+        executable,
+        [
+            '-NoProfile',
+            '-NonInteractive',
+            '-Command',
+            '(Get-CimInstance -ClassName Win32_OperatingSystem -Property LastBootUpTime).LastBootUpTime.ToUniversalTime().Ticks.ToString()',
+        ],
+        { windowsHide: true, timeout: 10_000, maxBuffer: 2_048 }
+    );
+    return bootIdentity(hostname(), stdout.trim());
+};
+
+export const createBootIdentityServer = ({ key, getBootId }) => {
+    if (!Buffer.isBuffer(key) || key.length !== 32) throw new Error('A 32-byte host boot key is required.');
+    const server = createServer((request, response) => {
+        response.setHeader('Cache-Control', 'no-store');
+        if (request.method !== 'GET' || request.url !== '/v1/boot-id') {
+            response.writeHead(404).end();
+            return;
+        }
+        const challenge = request.headers['x-sammo-challenge'];
+        const signature = request.headers['x-sammo-signature'];
+        if (
+            typeof challenge !== 'string' ||
+            !/^[a-f0-9]{64}$/.test(challenge) ||
+            typeof signature !== 'string' ||
+            !/^[a-f0-9]{64}$/.test(signature) ||
+            !timingSafeEqual(
+                Buffer.from(signature, 'hex'),
+                Buffer.from(mac(key, `sammo-host-boot-request-v1\n${challenge}`), 'hex')
+            )
+        ) {
+            response.writeHead(401).end();
+            return;
+        }
+        const bootId = getBootId();
+        if (typeof bootId !== 'string' || !/^[a-f0-9]{64}$/.test(bootId)) {
+            response.writeHead(503).end();
+            return;
+        }
+        response.writeHead(200, { 'Content-Type': 'application/json' });
+        response.end(
+            JSON.stringify({
+                version: 1,
+                challenge,
+                bootId,
+                signature: mac(key, `sammo-host-boot-response-v1\n${challenge}\n${bootId}`),
+            })
+        );
+    });
+    server.headersTimeout = 5_000;
+    server.requestTimeout = 5_000;
+    server.keepAliveTimeout = 1_000;
+    server.maxHeadersCount = 20;
+    return server;
+};
+
+export const main = async (args = process.argv.slice(2), env = process.env) => {
+    if (args.length === 2 && args[0] === '--create-key') {
+        await writeFile(args[1], `${randomBytes(32).toString('hex')}\n`, { flag: 'wx', mode: 0o600 });
+        console.info('Host boot key created. Copy it privately to the Linux secret file.');
+        return;
+    }
+    if (args.length === 2 && args[0] === '--config') {
+        const config = JSON.parse(await readFile(args[1], 'utf8'));
+        if (
+            !config ||
+            typeof config.keyFile !== 'string' ||
+            typeof config.bindAddress !== 'string' ||
+            !Number.isInteger(config.port)
+        ) {
+            throw new Error('Host boot agent configuration is invalid.');
+        }
+        env = {
+            ...env,
+            HOST_BOOT_KEY_FILE: config.keyFile,
+            HOST_BOOT_BIND_ADDRESS: config.bindAddress,
+            HOST_BOOT_PORT: String(config.port),
+        };
+    } else if (args.length) throw new Error('Usage: node windows-boot-agent.mjs [--create-key path | --config path]');
+    if (!env.HOST_BOOT_KEY_FILE) throw new Error('HOST_BOOT_KEY_FILE is required.');
+    const encoded = (await readFile(env.HOST_BOOT_KEY_FILE, 'utf8')).trim();
+    if (!/^[a-f0-9]{64}$/.test(encoded)) throw new Error('Host boot key file is invalid.');
+    const port = Number(env.HOST_BOOT_PORT ?? '17781');
+    if (!Number.isInteger(port) || port < 1 || port > 65535) throw new Error('HOST_BOOT_PORT is invalid.');
+    // OS 부팅 시각은 agent 재시작에도 같으며, 시각 차이 계산으로 재부팅을 추정하지 않는다.
+    const bootId = await readWindowsBootIdentity();
+    const server = createBootIdentityServer({ key: Buffer.from(encoded, 'hex'), getBootId: () => bootId });
+    server.on('error', () => {
+        console.error('Host boot agent listener failed.');
+        process.exitCode = 1;
+    });
+    server.listen(port, env.HOST_BOOT_BIND_ADDRESS ?? '127.0.0.1', () => {
+        console.info('Host boot identity agent is listening.');
+    });
+    for (const signal of ['SIGINT', 'SIGTERM']) process.once(signal, () => server.close());
+};
+
+if (process.argv[1] && resolve(process.argv[1]) === fileURLToPath(import.meta.url)) {
+    main().catch(() => {
+        console.error('Host boot agent startup failed; check OS, private key file, and listener configuration.');
+        process.exitCode = 1;
+    });
+}

tools/host-recovery/windows-boot-agent.test.mjs

@@ -0,0 +1,42 @@
+import { after, test } from 'node:test';
+import assert from 'node:assert/strict';
+import { createHmac, randomBytes, timingSafeEqual } from 'node:crypto';
+import { once } from 'node:events';
+import { bootIdentity, createBootIdentityServer } from './windows-boot-agent.mjs';
+
+test('agent restart preserves OS boot identity; a new OS boot changes it', () => {
+    assert.equal(bootIdentity('HOST', '638985120000000000'), bootIdentity('host', '638985120000000000'));
+    assert.notEqual(bootIdentity('HOST', '638985120000000000'), bootIdentity('HOST', '638985120000000001'));
+    assert.throws(() => bootIdentity('HOST', 'invalid'));
+});
+
+test('HTTP transport rejects anonymous requests and signs the current challenge and boot identity', async () => {
+    const key = randomBytes(32);
+    let bootId = 'a'.repeat(64);
+    const server = createBootIdentityServer({ key, getBootId: () => bootId });
+    server.listen(0, '127.0.0.1');
+    await once(server, 'listening');
+    const url = `http://127.0.0.1:${server.address().port}/v1/boot-id`;
+    after(
+        () =>
+            new Promise((resolve) => {
+                server.close(resolve);
+                server.closeAllConnections();
+            })
+    );
+    assert.equal((await fetch(url)).status, 401);
+    assert.equal((await fetch(`${url}?bypass=1`)).status, 404);
+    const challenge = randomBytes(32).toString('hex');
+    const signature = createHmac('sha256', key).update(`sammo-host-boot-request-v1\n${challenge}`).digest('hex');
+    const headers = { 'x-sammo-challenge': challenge, 'x-sammo-signature': signature };
+    const response = await fetch(url, { headers });
+    assert.equal(response.status, 200);
+    assert.equal(response.headers.get('cache-control'), 'no-store');
+    const body = await response.json();
+    assert.equal(body.challenge, challenge);
+    const expected = createHmac('sha256', key).update(`sammo-host-boot-response-v1\n${challenge}\n${bootId}`).digest();
+    assert.ok(timingSafeEqual(expected, Buffer.from(body.signature, 'hex')));
+    bootId = 'b'.repeat(64);
+    assert.equal((await (await fetch(url, { headers })).json()).bootId, bootId);
+    assert.equal((await fetch(url, { headers: { ...headers, 'x-sammo-signature': '0'.repeat(64) } })).status, 401);
+});

tools/integration-tests/scripts/play-audit-npc-lifecycle.ts

@@ -3,7 +3,7 @@ import assert from 'node:assert/strict';
 import { mkdir, readFile, writeFile } from 'node:fs/promises';
 import { resolve } from 'node:path';
 import { asRecord, type TurnCheckpoint } from '@sammo-ts/common';
-import { createGamePostgresConnector } from '@sammo-ts/infra';
+import { createGamePostgresConnector, decodeAuditDecisionSteps } from '@sammo-ts/infra';
 import { createTurnDaemonRuntime, seedScenarioToDatabase } from '@sammo-ts/game-engine';
 
 const databaseUrl = process.env.DATABASE_URL;
@@ -39,7 +39,7 @@ const verify = async (progress: unknown[]) => {
             include: { chunks: { orderBy: { ordinal: 'asc' } } },
         });
         assert(row && row.chunks.length, `Missing actual NPC ${JSON.stringify(selection)} decision`);
-        const steps = row.chunks.flatMap((chunk) => (Array.isArray(chunk.steps) ? chunk.steps : []));
+        const steps = row.chunks.flatMap((chunk) => decodeAuditDecisionSteps(chunk.steps));
         assert.equal(steps.length, row.stepCount);
         assert.equal(asRecord(steps[0]).kind, 'DECISION_START');
         assert(steps.some((step) => asRecord(step).kind === 'DECISION_END'));

tools/play-audit-maintenance.mjs

@@ -0,0 +1,95 @@
+#!/usr/bin/env node
+// stopped/reserved profile도 같은 fenced batch로 정리한다. 기본은 관찰만 하며 외부 scheduler가 실행한다.
+import { performance } from 'node:perf_hooks';
+import { setTimeout } from 'node:timers/promises';
+import { createGamePostgresConnector, pruneDeliveredReadModelOutbox } from '../packages/infra/dist/index.js';
+
+// PM2/호출자가 넘긴 turn-daemon 역할로 engine root의 자동 기동 경로에 들어가지 않는다.
+process.env.GAME_ENGINE_ROLE = 'audit-maintenance';
+const { prunePreviousAuditBatch } = await import('../app/game-engine/dist/index.js');
+
+const args = process.argv.slice(2);
+const apply = args.includes('--apply');
+const schemas =
+    args
+        .find((arg) => arg.startsWith('--schemas='))
+        ?.slice(10)
+        .split(',') ?? [];
+const budgetMs = Number(args.find((arg) => arg.startsWith('--budget-ms='))?.slice(12) ?? 3000);
+if (
+    !schemas.length ||
+    schemas.some((schema) => !/^[a-z][a-z0-9_]{0,62}$/.test(schema)) ||
+    !Number.isSafeInteger(budgetMs) ||
+    budgetMs < 1 ||
+    budgetMs > 60000 ||
+    args.some((arg) => arg !== '--apply' && !arg.startsWith('--schemas=') && !arg.startsWith('--budget-ms='))
+) {
+    throw new Error('Usage: --schemas=che,kwe,... [--budget-ms=3000] [--apply]');
+}
+let failed = false;
+for (const schema of schemas) {
+    const url = process.env.DATABASE_URL ? new URL(process.env.DATABASE_URL) : new URL('postgresql://localhost');
+    if (!process.env.DATABASE_URL) {
+        url.hostname = process.env.POSTGRES_HOST ?? 'localhost';
+        url.port = process.env.POSTGRES_PORT ?? '5432';
+        url.username = process.env.POSTGRES_USER ?? 'sammo';
+        url.password = process.env.POSTGRES_PASSWORD ?? '';
+        url.pathname = `/${process.env.POSTGRES_DB ?? 'sammo'}`;
+    }
+    url.searchParams.set('schema', schema);
+    const connector = createGamePostgresConnector({ url: url.href, maxConnections: 1 });
+    try {
+        await connector.connect();
+        const db = connector.prisma;
+        const world = await db.worldState.findFirst({ select: { meta: true } });
+        const serverId =
+            world?.meta && typeof world.meta === 'object' && !Array.isArray(world.meta) ? world.meta.serverId : null;
+        if (typeof serverId !== 'string' || !serverId.trim()) {
+            console.log(JSON.stringify({ schema, status: 'identityMissing', deleted: 0 }));
+            continue;
+        }
+        if (!apply) {
+            console.log(JSON.stringify({ schema, status: 'dryRun', apply: false }));
+            continue;
+        }
+        const deadline = performance.now() + budgetMs;
+        let auditDeleted = 0;
+        let outboxDeleted = 0;
+        let status = 'progress';
+        let auditComplete = false;
+        let outboxComplete = false;
+        do {
+            if (!auditComplete) {
+                const result = await prunePreviousAuditBatch(db, serverId);
+                auditDeleted += result.deleted;
+                status = result.status;
+                auditComplete = result.status === 'complete';
+                if (result.status === 'identityChanged' || result.status === 'busy') break;
+            }
+            if (!outboxComplete) {
+                outboxDeleted += await db.$transaction(async (tx) => {
+                    await tx.$executeRaw`SET LOCAL statement_timeout='2000ms'`;
+                    await tx.$executeRaw`SET LOCAL lock_timeout='250ms'`;
+                    const deleted = await pruneDeliveredReadModelOutbox(tx, {
+                        deliveredBefore: new Date(Date.now() - 86400000),
+                        limit: 500,
+                    });
+                    outboxComplete = deleted < 500;
+                    return deleted;
+                });
+            }
+            if (auditComplete && outboxComplete) break;
+            await setTimeout(25);
+        } while (performance.now() < deadline);
+        console.log(
+            JSON.stringify({ schema, status, auditDeleted, outboxDeleted, complete: auditComplete && outboxComplete })
+        );
+    } catch {
+        // 연결 URL과 원문 SQL을 출력하지 않고 scheduler가 감지할 수 있는 종료 코드를 남긴다.
+        failed = true;
+        console.error(JSON.stringify({ schema, status: 'failed' }));
+    } finally {
+        await connector.disconnect();
+    }
+}
+if (failed) process.exitCode = 1;