Files
clean-architecture-backend-…/docs/reviews/2026-09-24-src-module-code-review.md
T
DongHyeonkaandClaude Opus 5.5 6d27356e2b refactor: 부가 기능을 capability 모듈로 포장하고 기본 조립을 web-only로 전환
설계: docs/superpowers/specs/2026-09-25-capability-module-packaging-design.md
계획/진행 기록: docs/superpowers/plans/2026-09-25-capability-module-packaging.md

- Wave 0: 규칙 고정(AGENTS.md capability 완료 기준·테스트 규칙) + 차단 결함 7건
  (compose overlay 검사, redis-sdk CI job, 적용된 V9/V13 원복 + V16, @EnableScheduling,
  messaging bridge 순서, prod 에러 상세 validator, sample loggers read-only)
- Wave 1: 데이터·보안 Critical 10건 (durable operation 25P02, operation 유스케이스 이동,
  outbox 재인코딩, Kafka TLS/SASL, 알림 트랜잭션, SNS 서명, tus checksum, GraphQL fragment,
  Redis lane, gRPC FAILED_TERMINAL)
- Wave 2: 리팩터링 잔재·실험용 자리 삭제, outbox v1 fencing(V17), 멱등성 V2 정본화
- Wave 3: 라이브러리형 코드 정리(동작하지 않는 설정·인터페이스·gate 제거, web-advanced 분리)
- Wave 4: capability 11개를 "있으면 켜짐" 모듈로 포장, 끄는 장치 제거, 기본 조립 web-only,
  qualificationBootJar 분리
- Wave 5: 빌드 로직 축소(17.2k -> 10.8k줄), sample 테스트 컨벤션 적용

삭제한 코드는 docs/capability-status.md "제거된 코드" 표에 복구 커밋(208077fd)과 함께 기록.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
2026-09-25 12:49:07 +09:00

1131 lines
164 KiB
Markdown
Raw Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
# `src/` 모듈별 코드 리뷰 — 2026-09-24
## 개요
| 항목 | 내용 |
|---|---|
| 대상 | `src/` 아래 모든 등록 leaf(`src/config/architecture/modules.json`) + `src/build-logic`, `src/build-tools`, root Gradle, `.github/workflows` |
| 기준 | 현재 작업 트리 (`main` @ `208077fd`). `944a1e34`(2026-09-17) 이후 변경분을 중점 검토 |
| 규모 | main 코드 약 32만 줄 |
| 방법 | 기능 단위 15개 그룹으로 나눠 리뷰 에이전트 15개를 병렬 실행(읽기 전용, Gradle 미실행). 이후 Critical 15건은 오케스트레이터가 코드에서 직접 재확인 |
| 축 | correctness, readability/simplicity, architecture(HARD-STOP 1-5), security, performance, tests, 문서 drift |
| 이전 리뷰 | `docs/reviews/2026-09-18-techlead-review/*`, `2026-09-19-review-remediation-master.md` 기준으로 closed 항목의 회귀 여부를 함께 판정 |
## 판정
**Verdict: BLOCKED**
- Critical 15건, Required 76건(중복 1건 제외), Consider 약 55건.
- 게이트 결과: `./gradlew architectureCheck --console=plain` → **BUILD SUCCESSFUL** (1m 43s). 의존 방향·application-core 순수성·runtime membership·Clean Architecture 의존 검사 전부 통과.
- 모듈 경계(HARD-STOP 1, 2, 3, 5)는 모든 그룹에서 위반 0건. 막히는 이유는 경계가 아니라 **동작 결함, 보안 설정의 실제 미적용, 최근 리팩터 커밋의 회귀**다.
- HARD-STOP 4 위반 1건: web 어댑터가 저장소 포트·트랜잭션·객체 인가를 직접 수행(I-1 = A-4).
## 그룹별 요약
| 그룹 | 대상 leaf | 판정 | Critical | Required |
|---|---|---|---|---|
| A Core | domain-core, shared-contract, application-core, support, identifier | NEEDS_FIX | 0 | 6 (A-4는 I-1과 같은 결함) |
| B Notification | application-notification, adapter-outbound-notification | NEEDS_FIX | 2 | 5 |
| C File server | application-fileserver, persistence-fileserver, outbound fileserver, web-fileserver, composition | BLOCKED | 1 | 9 |
| D Object storage | application-objectstorage, adapter-outbound-objectstorage | NEEDS_FIX | 0 | 4 |
| E JPA | adapter-outbound-persistence-jpa | BLOCKED | 2 | 4 |
| F Mongo | adapter-outbound-persistence-mongo | NEEDS_FIX | 0 | 6 |
| G Redis | redis-runtime, redis-sdk, cache-redis | NEEDS_FIX | 1 | 5 |
| H HTTP client / messaging bridge | httpclient, outbound messaging, messaging-kafka | NEEDS_FIX | 1 | 5 |
| I Web | adapter-inbound-web | BLOCKED | 1 | 6 |
| J GraphQL | adapter-inbound-graphql | NEEDS_FIX | 1 | 5 |
| K Realtime / gRPC adapter | adapter-inbound-websocket, adapter-inbound-grpc | NEEDS_FIX | 0 | 7 |
| L Bootstrap / sample | app-bootstrap, sample-portfolio | NEEDS_FIX | 1 | 3 |
| M Messaging platform | `messaging:*` | NEEDS_FIX | 2 | 5 |
| N gRPC platform | `grpc:*`, `grpc-advanced:*` | NEEDS_FIX | 1 | 5 |
| O Build / CI | build-logic, build-tools, root Gradle, workflows | BLOCKED | 2 | 2 |
## Critical 15건 (오케스트레이터 재확인 결과 포함)
"코드 확인"은 인용된 줄을 직접 열어 주장과 일치함을 확인했다는 뜻이다. 실행으로 재현한 것은 없다.
| # | 그룹 | 결함 | 위치 | 재확인 |
|---|---|---|---|---|
| 1 | O | root plugin `apply()`가 `../docker-compose.<profile>.yml` 부재 시 예외 → Docker 빌드 컨텍스트(`src`)에서 `verifyDependencyLocks` 불가, release 이미지 빌드 차단 | `src/build-logic/.../bootstrap/DeveloperBootstrapPlugin.java:40-50` | 코드 확인 (`repositoryDirectory = projectDir.getParentFile()`, Dockerfile `WORKDIR /build/src`) |
| 2 | O | CI `redis-sdk` job의 `--tests '*…'` 따옴표가 word-split 후에도 남아 매칭 0 → 항상 실패, `release-gate`가 success 요구 | `.github/workflows/ci-quality-gates.yml:120`, `_reusable-gradle.yml:58-63` | 코드 확인 (CI 미실행) |
| 3 | E | 이미 적용된 Flyway 마이그레이션 3개를 제자리 수정 → 이전 버전이 적용된 DB 전부 checksum mismatch로 기동 실패 | `persistence-jpa/.../postgresql/V9__…sql`, `V13__…sql`, `jpa/outbox-storage/V1__…sql` | 코드 확인 (`git log --follow`: V9는 08-15 생성 후 9da00eb7에서 주석 변경, V13은 09-18 생성 후 09-23 변경, outbox V1은 07-31 생성 후 09-23 변경. `validate-on-migrate` 끄는 설정 없음). 템플릿에서 파생된 서비스 DB도 영향 |
| 4 | E | durable operation 재제출이 같은 트랜잭션에서 unique 위반 후 재조회 → PostgreSQL 25P02로 실패 | `persistence-jpa/.../operation/DurableOperationStoreAdapter.java:47-78` | 코드 확인 (PostgreSQL 실행 미검증) |
| 5 | L | outbox relay의 `@Scheduled`가 `@EnableScheduling` 없이는 동작 안 함. 앱 전체에서 유일한 `@EnableScheduling`이 `idempotency.provider=jdbc` 조건부 | `app-bootstrap/.../outbox/OutboxRelayScheduler.java:36`, `.../idempotency/IdempotencyConfig.java:17-19` | 코드 확인 (main의 `@EnableScheduling`은 sample·composition·IdempotencyConfig 3곳뿐) |
| 6 | H | canonical bridge의 `@ConditionalOnBean(EncodedMessagePublisher)`가 platform auto-config보다 먼저 평가됨(after 지정 없음) + 빈 `producer-id`가 조건을 통과해 `new ProducerId("")` 예외 | `adapter/outbound/messaging/.../MessagingBridgeRootAutoConfiguration.java:25,39-44` | 코드 확인 (컨텍스트 조립 테스트 미실행) |
| 7 | M | outbox relay·DLQ·retry-topic이 이미 인코딩된 `EncodedMessage`를 `publish()`로 넘김 → 재인코딩 실패 → 모든 row REJECTED/FAILED | `messaging-outbox-jdbc-postgresql/.../OutboxRelay.java:223-229`, `DefaultMessagePublisher.java:177` | 코드 확인 |
| 8 | M | 검증은 통과하는 Kafka TLS/SASL 설정이 producer에 적용되지 않음 (`KafkaSecurityConfigurer`는 bean으로만 생성, 호출 0) | `messaging-spring-boot-starter-kafka/.../KafkaMessagingAutoConfiguration.java:109-154` | 코드 확인 |
| 9 | B | 알림 recovery/replay 워커가 `@Modifying` 쿼리를 트랜잭션 없이 실행 → 매 pass 실패, 예외는 WARN으로 삼켜짐 | `adapter/outbound/notification/.../dispatch/LeaseRecoveryService.java:106` 외 3곳 | 코드 확인 (파일 내 `inWrite/TransactionPort/@Transactional` 0건. PG 실행 미검증) |
| 10 | B | SNS 인증서 URL 검증이 `*.amazonaws.com` 전체 허용(공격자 S3 버킷 통과), TopicArn allowlist·Timestamp 검사 없음, 콜백은 `permitAll` | `.../ses/SnsSignatureVerifier.java:99-109`, `SesProviderCallbackAssembler.java:19` | 코드 확인 (suffix `"amazonaws.com"`). SES callback 활성 시에만 노출(기본 off) |
| 11 | C | tus `Upload-Checksum`을 chunk가 아니라 파일 전체 누적 digest와, 그것도 offset 커밋 후에 비교 | `adapter/inbound/web-fileserver/.../tus/TusController.java:167-174` | 코드 확인 |
| 12 | I | durable-operation 조회/취소를 web 어댑터가 저장소 포트·트랜잭션·인가로 직접 수행(HARD-STOP 4), 운영 조립처 없음 | `adapter/inbound/web/.../operationasync/OperationQueryService.java:54-58,80` | 코드 확인 (`new OperationQueryService(` main 0건) |
| 13 | J | 인가 단계가 fragment를 visit budget 없이 재귀 전개 → 1.6KB 문서로 2^40 방문, cost 단계보다 먼저 실행 | `adapter/inbound/graphql/.../runtime/GraphQlDocumentAuthorizationHandler.java:109-137` | 코드 확인 (chain 순서 `selection → authorization → cost`) |
| 14 | G | SCRIPT lane 한도(기본 8)가 연결이 아니라 동시 lease 수 → Redis 정상이어도 동시 9번째 호출부터 거부, rate-limit fail-closed로 전면 429/503 | `redis-runtime/.../RedisRuntimeOwner.java:134-140`, `RedisSdkSettings.java:687` | 코드 확인. closed H8의 회귀 |
| 15 | N | gRPC idempotency가 FAILED_TERMINAL 행에서 claim 없이 PROCEED → 커밋 시 예외 또는 중복 실행, fake는 이 전이를 허용해 차이를 숨김 | `grpc/grpc-policy/.../GrpcIdempotencyInterceptor.java:109-114` | 코드 확인 (production caller 0, 잠복) |
## 공통 패턴
1. **closed로 기록됐지만 코드가 따라오지 않은 항목**: Redis H8(#14)·M4(rate-limit hash), Kafka H6(timeout 취소 무효), object storage L3(파일 unchanged), websocket BOOT-002(prod 판정 재구현), messaging MSG-006(cleanup job 미전환), gRPC M1(admission race 일부만), JPA H10(tenant 규칙 이름 기반), GraphQL M1(DataLoader 에러 코드). closure 문서가 코드 상태를 보장하지 못한다.
2. **보안·안전 설정이 바인딩·검증만 되고 적용되지 않음**: Redis 안전 설정 약 20개, Kafka TLS/SASL(#8), gRPC `environment` 기본 LOCAL, GraphQL `persistedOperationOnly`·`idempotencyKey`, httpclient dynamic `trace-propagation`, `StartupSafetyValidator`의 prod 에러 상세 차단 주장, `wire-type-manifest.yaml`, Mongo tenancy/CSFLE, shared-contract cardinality 보호. 운영자는 보호받는다고 믿는다.
3. **조립되지 않은 코드가 대량으로 남아 있음**: GraphQL 약 57%(14.7k LOC), redis-sdk 약 22k LOC, notification R1 약 4.5k LOC, web library-only 약 5.7k LOC, object storage direct/multipart 약 3.2k LOC, messaging 미사용 API 타입. 일부 문서는 이 코드를 "제공되는 보호"로 서술한다.
4. **fake가 실제 adapter와 계약이 달라 결함을 숨김**: gRPC ledger fake(#15), tus fake digest(#11), Redis fake HDEL/PEXPIRE, notification recorder fake, durable operation in-memory store(#4), messaging fake publisher(#7).
5. **최근 리팩터 커밋의 회귀**: `208077fd`(#1), `9da00eb7`(#2, #3, BOOT-002, `StartupSafetyValidator` 축소), `60b6a319`(#3 V13), `ace8aaae`(sample이 중앙 ArchUnit에서 빠짐).
## 권고 순서
1. **Wave 0 — 결정 없이 바로 고칠 수 있는 차단 항목**: #1, #2, #3(파일 3개 원복 + 새 버전 추가 + 버전 파일 해시 고정 테스트), #5, #6, L-2(prod 에러 상세 validator), L-3(sample `loggers` read-only).
2. **Wave 1 — 데이터·보안 Critical**: #4, #7, #8, #9, #10, #11, #13, #14, #15, #12(유스케이스를 application-core로 옮기거나 기능 삭제).
3. **Wave 2 — 그룹별 Required**: 아래 각 그룹 섹션 순서대로. closed 항목 회귀분은 remediation master에 재오픈 기록.
4. **Wave 3 — 조립되지 않은 코드 정리**: 삭제 vs 유지는 소유자 결정이 필요(아래).
### 결정 필요
- 조립되지 않은 대형 코드(redis-sdk, GraphQL `advanced/**` 등, notification R1, web library-only, object storage direct/multipart)를 삭제할지, 별도 artifact로 옮길지, 유지할지. 이전 closure(C3, NOTIF-ADR-005, M1 LIBRARY_ONLY)가 "유지"로 결정했으나, 이번 리뷰는 그 전제(도입 가능성, 테스트된 부품)가 일부 성립하지 않음을 확인했다. — 승인자: 저장소 소유자
## 검증 기록
| 항목 | 결과 |
|---|---|
| `cd src && ./gradlew architectureCheck --console=plain` | BUILD SUCCESSFUL, 172 tasks (`verifyApplicationCoreDependencyPurity`, `verifyNoIgnoredSourcePackages`, `verifyRuntimeModuleRegistry`, `verifyRuntimeModuleMembership`, `verifyCleanArchitectureDependencies`, `:app-bootstrap:architectureTest` 포함) |
| Critical 15건 코드 재확인 | 15/15 인용 위치의 코드가 주장과 일치 |
| 테스트·`check`·`ci` | 미실행 (전체 실행은 별도 승인 범위. 리뷰 에이전트는 병렬 Gradle 충돌 방지를 위해 읽기 전용) |
| 실행으로만 확정 가능한 항목 | #4(PostgreSQL 25P02), #6(auto-config 조립 순서), #9(TransactionRequiredException), #2(CI), #13(DoS 시간) — 각 그룹 섹션에 확인용 테스트를 제안해 둠 |
> **보정 (2차 리뷰 반영):** B-2의 "위조 callback이 `SesSuppressionUpdater`로 suppression을 만든다"는 과장이다. 이 클래스는 참조가 0이다. 서명 검증 우회 자체는 유효하다. A-2는 bridge 없이 `RuntimeEnvironment.isProductionActive(...)` 정적 호출 한 줄로 고칠 수 있다. 상세는 `2026-09-24-src-maintainability-review.md`의 "1차 리뷰 보정" 절.
## 그룹별 상세
아래는 그룹별 리뷰 결과 원문이다. 경로는 저장소 루트 기준이며, 일부 그룹은 약어(`B/`, `BR/`, `S/`, `.../`)를 쓴다. 각 발견의 Status는 이전 리뷰 ID 기준(new / regression / known-open)이다.
### Group A — Core
#### Leaf별 판정
| leaf | 판정 | 이유 |
|---|---|---|
| `src/domain-core` | PASS | 순수 Java. main 7개 파일이 마커와 포트뿐; README가 말한 ArchUnit 규칙이 app-bootstrap architectureTest에 실제로 있음 |
| `src/shared-contract` | NEEDS_FIX | 코드 결함 없음. 쓰는 곳 없는 "계약" 타입 여럿, README가 없는 런타임 보호를 약속 |
| `src/application-core` | NEEDS_FIX | 테스트 0·호출자 0인 V2 발행 코드 약 466줄; operation 조회/취소 유스케이스 부재로 web 어댑터가 대신 보유; 문서 불일치 |
| `src/adapter/outbound/support` | NEEDS_FIX | 예외 메시지를 그대로 WARN 로그 → PII·비밀값 누출 가능; `RuntimeProfileSnapshot` "유일한 reader" 주장 틀림 |
| `src/adapter/outbound/identifier` | NEEDS_FIX | 문서만 문제(의존성·훅·테스트 이름 불일치). `UuidCodec`은 죽은 코드 |
#### 발견 (심각도 순)
1. **[Required][security] `FailOpenDependencyLogger`가 `cause.getMessage()`를 그대로 WARN 로그에 남김** — `src/adapter/outbound/support/src/main/java/dev/caskeleton/adapter/outbound/support/FailOpenDependencyLogger.java:37-47`
- 근거: `error="{}: {}"`에 `getSimpleName()`과 `getMessage()`. Javadoc `:10-11`은 "PII cannot reach the log". 호출자 `FailOpenNotificationProvider.java:42`로 fork의 `GoogleEmailClient`/`SlackClient` 예외가 올라옴. 테스트 `FailOpenDependencyLoggerTest.java:64-73`은 메시지 `"connection refused"`로 `@gmail.com` 부재를 확인 → 아무것도 증명 못 함.
- 왜 문제인가: 메일 예외 메시지엔 수신자 주소, RestClient I/O 예외엔 Slack webhook URL 전체(비밀값)가 흔히 포함. 따옴표·줄바꿈이 key=value 로그 형식도 깸.
- 수정: 예외 클래스 이름만 기록(필요 시 길이 제한·정제된 reason code). 테스트는 `alice@example.com`, `https://hooks.slack.com/...`를 메시지에 넣어 재작성.
- 상태: new (application-core README "notification 값은 로그에 안 남긴다"를 우회하는 경로)
2. **[Required][security] BOOT-002 회귀: prod 판정을 손으로 재구현, 가드는 bootstrap만 검사** — `src/adapter/inbound/websocket/src/main/java/dev/caskeleton/adapter/inbound/websocket/config/WebSocketPlatformSafetyConfiguration.java:78-84`
- 근거: `"prod".equals(profile) || "production".equals(profile)` (trim/소문자 없음) vs `RuntimeEnvironment.fromProfile`(trim+소문자). 결과가 `:63` `WebSocketPlatformStartupValidator(production(...))`로 전달. 9da00eb7(09-23)에서 생긴 파일. `RuntimeProfileSnapshot.java:12-13`의 "the only production type ... that reads getActiveProfiles"는 거짓(fileserver, httpclient, objectstorage, websocket도 직접 읽음). 가드 `RuntimeProfileAccessArchitectureTest.java:13,21`은 `dev.caskeleton.bootstrap..`만 검사. 근본 원인: 브리지가 outbound leaf에 있어 `INBOUND_ADAPTERS_DO_NOT_DEPEND_ON_OUTBOUND_ADAPTERS` 때문에 websocket이 못 씀.
- 왜 문제인가: `SPRING_PROFILES_ACTIVE=Prod` → bootstrap 검증기는 PROD로 보지만 websocket prod 안전 검증은 건너뜀(fail-open 분기).
- 수정: websocket은 `RuntimeEnvironment.isProductionActive(List.of(env.getActiveProfiles()))`(shared-contract edge 허용); ArchUnit 범위를 `dev.caskeleton..`로(messaging, sample 제외); Javadoc(`RuntimeEnvironment.java:39-41` 포함) 정정.
- 상태: BOOT-002 회귀 (websocket 그룹 코드)
3. **[Required][tests] V2 단건 발행 코드가 테스트 0, 호출자 0** — `src/application-core/src/main/java/dev/caskeleton/application/outbox/v2/PublishNextOutboxDeliveryUseCase.java:26-242`, `OutboxV2RetryPolicy.java:33-151`
- 근거: `git grep 'PublishNextOutboxDelivery\|OutboxV2RetryPolicy\|OutboxPublicationAdmissionPort'` → 자기 파일과 docs뿐; 테스트 삭제 이력 없음; 계획서 `docs/superpowers/plans/2026-09-18-outbox-v2-one-record-publication-policy.md`의 "Prove …" 8개 항목 모두 미체크. 위험: `:111-115` `publisher.publish()`가 future 반환 전 동기로 오래 걸리면 `publishTimeout` 미적용 → lease 만료 후 `requireTransition`이 `IllegalStateException`; `:84-87` 2건 이상 claim 후에야 예외 → 행들이 lease 동안 묶임.
- 수정: `PublishNextOutboxDeliveryUseCaseTest`, `OutboxV2RetryPolicyTest` 추가, 또는 어댑터/bulkhead가 당분간 없으면 삭제(`OutboxBackoffPolicy`와 역할 중복).
- 가정/반론: 곧 wiring하지 않음 / Javadoc이 "Unwired" 명시한 의도적 후보.
- 상태: new (H7과 같은 종류; H7 종결은 당시 16개 클래스만)
4. **[Required][architecture] operation 조회·취소 유스케이스 부재 → web이 store 포트·트랜잭션·객체 권한 규칙 보유** — `src/application-core/src/main/java/dev/caskeleton/application/operation/`
- 근거: `adapter/inbound/web/.../operationasync/OperationQueryService.java:54-58,71-80`이 `DurableOperationStorePort.find/cancel`, `transactions.inWrite` 직접 호출; `OperationAccessPolicy.java:28`(owner+tenant)가 web에 있음 → `application-core/CLAUDE.md:211`(D3 "no thin web→read-port path"), `ObjectAccessPolicy.java:16`("Owned here rather than by the transport")에 어긋남. 반대로 `SubmitDurableOperationUseCase`는 `CommandUseCase`가 아니고 production 호출자 0, `LegacyOutboxAppendPort`(Javadoc상 legacy/sample 전용) 사용.
- 수정: `GetDurableOperationUseCase`(QueryUseCase), `CancelDurableOperationUseCase`(CommandUseCase + `@UseCaseCapability`) 추가; owner·tenant 판정은 `ObjectAccessPolicy` 뒤로; web은 `WebRequestContext`→actor·tenant 변환만.
- 가정/반론: `DurableOperationStorePort`가 D3의 read-port / `AdapterBoundaryArchitectureTest.java:354`가 inbound→outbound 포트 접근을 허용하고 기능이 아직 조립 전. HARD-STOP 4 해당 여부는 web 그룹 판단.
- 상태: new
5. **[Required][docs] shared-contract가 존재하지 않는 런타임 보호를 약속** — `src/shared-contract/README.md:350`, `src/shared-contract/CLAUDE.md:22`
- 근거: README "1001번째 tenant 부터는 bucket folding 자동 적용" — `CardinalityBounds`/`MetricNaming` production 참조 0, 실제 필터 `MetricsCardinalityMeterFilter.java`는 `ForbiddenMetricTags`만. `RedisHealthSnapshotProvider` 구현·소비자 0인데 CLAUDE.md는 "used by adapters and bootstrap health composition"(M5 잔재). `AlertSeverity`, `BaggageAllowlist`는 테스트에서만.
- 수정: tenant 상한을 `MeterFilter.maximumAllowableTags`로 실제 연결하거나 README 주장 삭제; 소비자 없는 타입 삭제 또는 testFixtures로.
- 상태: new
6. **[Required][docs] leaf CLAUDE.md/README가 현재 코드와 다름**
- `application-core/CLAUDE.md:76-106`, `README.md:143-165,545-575`: notification 패키지는 application-notification으로 이동했는데 규칙 잔존; cache, lease, realtime, operation, filepublication, inbox 절 없음.
- `identifier/CLAUDE.md:30-33`: 의존성 "`:domain-core` + uuid-creator" (실제 `:application-core`뿐); `:38` `.claude/hooks/ca_import_gate.py` 없음; `:47` `UuidCodecSpec` → `UuidCodecTest`.
- `identifier/README.md`: 패키지명 오류(`dev.caskeleton.adapter.identifier`), 폐기된 Spock 절, SSOT를 `allowedProjectDependencies`로 기재.
- `support/README.md:6-8,24`: "CLAUDE.md 없음"(있음), SSOT를 `src/build.gradle`로(실제 modules.json), 없는 `OutboundHttpDependencyLogger` 언급.
- `shared-contract/README.md:436,499`: `RateLimiter`, `RateLimiterFactory` 없음.
- 수정: modules.json·코드 기준 동기화, 사라진 타입/파일 언급 삭제.
7. **[Consider][correctness] `ValidatedIntegrationEvent`가 `envelopeSha256`–`envelopeBytes` 일치를 확인 안 함** — `src/application-core/.../messaging/event/ValidatedIntegrationEvent.java:40-88` (DB 복원 경로 `OutboxStoreAdapter.java:197`, `CanonicalOutboxEventRowMapper.java:71`가 저장된 해시를 그대로 신뢰; `CORRELATION_GRAMMAR`가 `*`라 길이 무제한인데 메시지는 "bounded"). 수정: 생성자에서 해시 재계산·비교, correlationId `{0,127}`. 가정/반론: DB 행 손상·변조 가능 / 인코더가 올바른 해시 생성.
8. **[Consider][correctness] v1 `IdempotencyExecutor.runAndComplete`가 완료 기록 실패를 action 실패로 처리** — `src/application-core/.../idempotency/IdempotencyExecutor.java:103-110` (action 성공 후 `codec.serialize`/`store.complete` 실패 → `discard` → 재시도 때 action 재실행; 전체를 한 `inWrite`로 감쌀 때만 안전 — sample `PublishPosterImageUseCase.java:119`는 그렇게 함 — 전제 미기재·미강제. V2 `IdempotencyExecutorV2.java:207-211`은 `preserveUnknown` 예외가 원래 예외를 삼킴). 수정: try 범위를 `action.get()`으로 축소; V2는 `addSuppressed`.
9. **[Consider][structure] 죽은 코드** — `identifier/.../UuidCodec.java` (외부 참조 0; "canonical" 계약도 틀림 — jshell로 `UUID.fromString("1-1-1-1-1")` 수용 확인), `application-core/src/test/.../CompiledPackageScanner.java`(참조 0), `application-core/.../fileexport/FileExportPort.java`(`@Deprecated(forRemoval)`, `exportCsv` 호출자 0). 수정: 삭제.
10. **[Nit]** identifier가 `ca.spring-library`를 써서 순수 단위 테스트에 webmvc-test·tomcat 유입 → `ca.java-library`; `MasterSwitch.java:57-66` 주석 줄 깨짐; `DomainContextStrategy`/`Factory`는 case 하나뿐인 확장 지점.
#### 확인 후 기각한 lead
- `RateLimiterFactory`, support README lead: 사실 → #6. notification PII: `Notification` 객체 자체는 logger로 안 감(예외 메시지 경로로 샘 → #1).
- 5개 leaf build.gradle edge 모두 허용 범위; application-core의 비-`dev.caskeleton`/`java`/`javax` import 0.
- `DB_LOCK_TIMEOUT` retryable=false는 `docs/registries/error-codes.yaml:873`과 일치; 새 `MasterSwitch` 상수는 `values()`로 모두 소비; `StrictSettingsBinder`는 httpclient env manifest 테스트가 간접 검증; `CacheAsideExecutor`/`LeaseWatchdog`는 H5 결정대로 유지, 회귀 없음.
#### 커버리지
정독: domain-core, support, identifier 전체; shared-contract 변경분 + activation/runtime/tracing/ratelimit/concurrency/metrics; application-core 944a1e34 이후 변경분 전체 + idempotency(v1·V2), outbox relay, cache executor, operation. 훑음: realtime, lease, inbox, lock, filepublication, OutboxPayloadPolicy. 동작 확인은 jshell UUID 파싱 1건.
### Group B — Notification (vertical slice)
#### Per-leaf verdict
| leaf | verdict | one-line reason |
|---|---|---|
| `application-notification` | NEEDS_FIX | Template/variable failure in dispatch repeats forever; recorder overwrites the recipient row with stale data; admin reconcile has no tenant check; ~4.5k lines of R1 code with no production implementation or caller. |
| `adapter-outbound-notification` | NEEDS_FIX | Forged SNS callbacks can be accepted; recovery and replay passes likely fail on every run; the HTTP response size cap does not limit what is read. |
No HARD-STOP violations; every import within `modules.json` edges.
#### Findings (most severe first)
1. **[Critical][correctness] Background recovery and replay passes run `@Modifying` queries with no transaction, so every pass fails** — `src/adapter/outbound/notification/src/main/java/.../platform/dispatch/LeaseRecoveryService.java:106`
- Evidence: first statement of `recoverOnce()` is `leases.expireOverdue(batchSize)` → `RecipientDeliveryJpaRepository.java:58-60` `@Modifying ... int expireOverdue`. `grep 'TransactionPort\|inWrite'` over the adapter main → 0. `NotificationDispatchService.java:334-338` already records this exact failure for a query in the same repository ("…so the release threw TransactionRequiredException"). PG integration tests wrap modifying calls in `fixture.inTransaction(...)` (`PostgreSqlRecipientLeaseFencingIntegrationTest.java:60,152`). Same defect in: `ProviderEventProjectionService.java:58,120` (`markFailed`, `bindAttempt` outside `inWrite`), `ProviderEventReplayWorker.java:72` (`markFailed` in catch, `@Modifying` at `ProviderEventJpaRepository.java:144-149`), `ReconciliationService.java:82` (`ledger.append` without transaction).
- Why it matters: `NotificationBackgroundWorkers` swallows the exception with a WARN → crash recovery, expiry retirement and callback replay silently never happen. The dispatch error path relies on recovery (`NotificationSchedulerWorker.java:96-98`) → failed deliveries stay DISPATCHING forever.
- Fix: wrap each row's work in `TransactionPort.inWrite` (e.g. move recovery into an application-notification service with the dispatch transaction boundary); move `resolve`/`markFailed` inside `inWrite`; add one real-DB PG-lane test running `recoverOnce`/`replayOnce`.
- Status: new. Not execution-confirmed (Gradle not run); the PG test would confirm.
2. **[Critical][security] SNS signature checks can be bypassed (bucket host, no topic allowlist, no freshness)** — `.../SnsSignatureVerifier.java:101-109`, `.../SesProviderCallbackAssembler.java:19`
- Evidence: host check accepts `amazonaws.com` or any `*.amazonaws.com` + path `/SimpleNotificationService-*.pem` → an attacker's S3 bucket (`https://<bucket>.s3.amazonaws.com/SimpleNotificationService-x.pem`) passes and their cert verifies any envelope (`SnsCertificateUrlTest` has no S3 case). `SesCallbackAdapter.java:50-53` checks only `isValid` — no TopicArn allowlist, so a genuine AWS-signed message from the attacker's own topic is accepted. `Timestamp` never checked. Endpoint is `permitAll` (`CallbackMvcSecurityConfiguration.java:35`). Certificate cache (`NotificationCallbackProviderConfig.java:36-46`) is unbounded, keyed by attacker-supplied URL, network fetch inside `computeIfAbsent`.
- Why it matters: forged Bounce/Complaint/Delivery → suppressions (`SesSuppressionUpdater`) and false delivery states; with #6 one request can OOM.
- Fix: host `^sns\.[a-z0-9-]+\.amazonaws\.com(\.cn)?$`; required per-profile `allowed-topic-arns` when SES callbacks are on; reject stale Timestamp; bounded cache + cert subject check.
- Status: new. Exposure only when an SES profile + callbacks are enabled (default off).
3. **[Required][correctness] Template/variable failure in dispatch repeats forever (poison pill)** — `NotificationDispatchService.java:146`
- Evidence: `render()`, `gateway.profile()`, `URI::create` run before the attempt row is inserted, uncaught. Submit only checks `templates.get(plan.template())` (`NotificationSubmissionService.java:118`); variable schema, SUBJECT control chars, URI scheme checked only at dispatch (`CanonicalNotificationRenderer.java:70`, `TemplateSlotPolicy`). No attempt row → recovery re-queues (`LeaseRecoveryService.java:111-120`). No render-failure test.
- Why it matters: a missing variable or newline in a subject → "accepted", then fails and re-queues every lease period forever; never FAILED.
- Fix: catch pre-dispatch `NotificationException`s before `insertDispatching` → FAILED with `TEMPLATE_FAILURE`; run the variable-schema check in `prepare()` too.
- Status: new
4. **[Required][correctness] Recovery leaves some rows DISPATCHING forever, then gets stuck on them** — `LeaseRecoveryService.java:122-140`
- Evidence: when every attempt already has a completion (crash between `recorder.record` and `applyNextAction`, or failed retry dispatch) the loop does nothing; `reconcile` → `Unsupported` also leaves the row (SMTP and SES have no reconciliation → every mid-send SMTP crash). `selectExpiredDispatching` (`ORDER BY lease_until LIMIT batchSize`) returns the same rows each pass. No "all attempts completed" test.
- Why it matters: once `claimBatchSize` such rows exist no other abandoned delivery is ever recovered.
- Fix: all attempts completed → re-derive recipient state from the last attempt via retry policy; `Unsupported` → RECONCILIATION_REQUIRED (not re-picked, operator-visible).
- Status: new
5. **[Required][correctness] Recorder writes a stale recipient snapshot without the lease check, undoing the late-match callback it just applied** — `DispatchOutcomeRecorder.java:105,123-124,135`
- Evidence: `bindUnmatched(...).forEach(projection::project)` rolls the callback result (e.g. DELIVERED) onto the recipient row (`JpaDeliveryAttemptStore.java:221-239`); right after, `recipients.save(updated)` writes back `deliveryOutcome()`/`evidenceLevel` from the dispatch-start snapshot. `save` is unfenced, breaking `RecipientLeaseStorePort.java:14` ("Every operation after the claim is fenced"). Only test uses a no-op `projections.save` fake (`DispatchOutcomeRecorderTest.java:176`).
- Why it matters: attempt row DELIVERED, recipient row back to UNKNOWN; the event is already APPLIED so replay won't fix it.
- Fix: re-read the recipient after projection; write only recorder-owned fields through `saveHeldBy(lease)`.
- Status: new
6. **[Required][security/perf] HTTP response "cap" reads the whole body first** — `JdkNotificationHttpGateway.java:136-141`
- Evidence: `BodySubscribers.mapping(ofByteArray(), body -> copyOf(...))` buffers everything then truncates; comment `:73-75` claims never-ending chunked responses are prevented — false. No test.
- Fix: `BodySubscribers.ofInputStream()` + `readNBytes(max)` (or a custom subscriber); oversized-body test.
- Status: new
7. **[Required][security] Admin reconcile has no tenant check and calls providers inside a write transaction** — `NotificationAdminApplicationService.java:156-158`, `:287-289`
- Evidence: `redrive`/`suppress` call `requireTenant`, `reconcile` does not → tenant-scoped operator acts on another tenant's attempts. `setProviderState` changes a global runtime but a tenant-scoped actor may call it. `AdminOperationExecutor` wraps the batch in `transactions.inWrite` → up to 200 provider round-trips in one DB transaction (contradicts `NotificationDispatchService.java:36-39`). `AdminAuthorizationGuard.java:43` always reports `SUPPRESS` as the denied authority.
- Fix: look up each attempt's tenant + `requireTenant`; `PROVIDER_CONTROL` only for unscoped actors; executor only claims/saves the operation, provider calls outside the transaction.
- Status: new
8. **[Consider][structure] Large body of code with no caller** — ~95 files / 4.5k lines of R1 types in root package `application.notification` (per-file `grep -rlw` = 0 except `NotificationPort`, `Notification`, `NotificationChannel`); `platform/reactor/*` no bean/test/consumer yet `reactor-core`/`reactor-test` declared; `spring-web` declared as "Slack RestClient" but `grep org.springframework.web` over the adapter = 0; `EmailNotifier`, `SmsNotifier`, `MobilePushNotifier`, `WebPushNotifier`, `ApplicationReceiptServiceImpl` no main references. Fix: delete Reactor facade + deps + `spring-web` now; re-check the R1 "retain" decision. Counter: NOTIF-ADR-005 retains R1 without a removal date.
9. **[Consider][correctness] SMTP reports pre-send failures as "ambiguous"; one dispatch branch clears next-dispatch time** — `SmtpNotificationProviderAdapter.java` throws `IllegalArgumentException` for non-email contact points and `reveal` failures surface as plain exceptions → not `ProviderCallNotStartedException` → `NotificationDispatchService.java:278-287` records AMBIGUOUS, duplicateRisk, reconciliation for a never-sent message. `DispatchPipeline.java:57-58` returns `READY_TO_DISPATCH` when `notBefore` not reached → caller `transition(..., Optional.empty())` sets `next_dispatch_at` NULL → row never claimed again (hard to hit). Fix: `ProviderCallNotStartedException` for these; pass `notBefore` as next-dispatch time.
10. **[Consider][docs]** Adapter `README.md:3-5` / `CLAUDE.md:18-31` call the whole implementation "R0 legacy" (R2 platform is most of the leaf; SMTP/SES shipped); `application-notification` has no CLAUDE.md (rules in `application-core/CLAUDE.md:76-111`); `CLAUDE.md:8` Boot 4.0.0 vs 4.0.8; "no inline FQN" rule violated on 65 adapter main lines (e.g. `LeaseRecoveryService.java:34-36`) — rest of known-open L6. Fix: rewrite around R2; new `src/application-notification/CLAUDE.md`.
11. **[Nit][structure] `contract-v1` test templates orphaned** — `src/adapter/outbound/notification/src/test/resources/notification/templates/**` (`grep -rn 'contract-v1\|notification/templates' src docs` → only historical plan `docs/superpowers/plans/2026-07-28-...md:927-930`; no `getResource`/`ClassPathResource`; filters `|html_text`, `|mrkdwn` match neither `{name}` placeholders nor `TemplateSlotMode`; consumer `NotificationTemplateRendererTest` deleted in 9da00eb7). Fix: delete the 4 files. Status: new (lead confirmed)
#### Leads checked and dismissed
- HARD-STOP/edges: application-notification imports only `java.*` + `dev.caskeleton`; adapter only `application.notification`, own packages, `support`, `shared.error`.
- H3 (Thymeleaf OGNL) not regressed (default `PlaceholderTemplateEngine`, `NotificationPlatformCoreConfig.java:70`); L7 keep; L9 effectively closed (`NotificationPlatformReachabilityTest.java:67-72`).
- AES-GCM random 12-byte nonce + header as AAD; Twilio signature constant-time; HTML slot `href` https-only + allowlisted tags/attrs; SSRF resolves every address (DNS rebinding window documented).
#### Coverage note
Deep: dispatch, scheduler, recovery, background, recorder, reconciliation, projection, submission, pipeline, retry policy, admin; SNS/SES callback path; HTTP gateway/endpoints; template engine/policy/renderer; provider runtime/gateway; SMTP adapter; persistence-jpa stores/SQL to confirm behavior. Skimmed: non-default adapters (FCM, APNs, WebPush, Twilio, Webhook), inbox/contact, settings/autoconfigure. R1: usage scan only.
### Group C — File server (vertical slice)
#### Per-leaf verdict
| leaf | verdict | one-line reason |
|---|---|---|
| `src/application-fileserver` | NEEDS_FIX | Quarantined bytes deleted; recovery path never runs; cancel leaves records/bytes behind; terminal sessions not checked; no per-file owner authz. |
| `src/adapter/outbound/persistence-fileserver` | NEEDS_FIX | 0 tests; race-critical writer/cleanup SQL untested anywhere. |
| `src/adapter/outbound/fileserver` | NEEDS_FIX | Every append re-reads the whole prefix. Path-traversal guards solid. |
| `src/adapter/inbound/web-fileserver` | BLOCKED | tus checksum checked against the wrong digest, after bytes are saved (Critical). |
| `src/composition/fileserver-platform` | PASS | Wiring only; every edge allowed by modules.json. |
No HARD-STOP violations: build.gradle edges ⊆ `allowed_dependencies`; `application-fileserver/src/main` has 0 Spring/Jakarta imports.
#### Findings (most severe first)
1. **[Critical][correctness/security] tus `Upload-Checksum` compared with the running whole-file digest, after the chunk is saved** — `src/adapter/inbound/web-fileserver/.../tus/TusController.java:167-174`
- Evidence: `uploadService.append(...)` commits the offset first; then `checksumVerifier.verify(header, appended.sha256Snapshot())`, where `sha256Snapshot` is the running digest from byte 0 (`LocalAppendEngine.java:55-58,71`).
- Why it matters: an honest client with a correct per-chunk checksum fails on every chunk after the first; a mismatching chunk stays saved with the offset advanced and finalize publishes it. Returns 422 where tus says 460.
- Test gap: `TusProtocolContractTest.java:149-167` sends a *wrong* checksum on chunk 2 and the fake's `digestOf` (`InMemoryUploadApplicationService.java:117`) is also cumulative — the test hides the bug.
- Fix: per-request digest inside append; pass expected checksum into `append`, verify before `commitOffset`, roll back on mismatch; map to 460; add a correct-checksum chunk-2 test.
- Status: new
2. **[Required][correctness/data] QUARANTINE verdict schedules staged bytes for deletion** — `src/application-fileserver/.../upload/DefaultFinalizeUploadService.java:236-253`
- Evidence: javadoc `:220-222` says only REJECT schedules removal; code queues `FAILED_VERIFICATION_CONTENT` for both. `DefaultCleanupService.cleanStaging` (`:134-146`) deletes without state check (`isReclaimable` `:219-223` covers published content only). Admin `reverify` (`DefaultFileserverAdminService.java:136-157`) then has no bytes.
- Fix: queue cleanup only for REJECT (or move to `quarantineRoot`); `cleanStaging` re-checks reclaimability.
- Status: new
3. **[Required][correctness] "Ambiguous completion → reconciliation" does not work end to end** — `DefaultFinalizeUploadService.java:317-333`
- Evidence: no caller of the reconciler (`grep -rn 'reconcileRecoverable\|\.reconcile(\|FileReconciliationService'` → bean def `FileserverMaintenanceServiceConfiguration.java:59-65` + one test); `RecoveryQueue.pending/resolve` have no production caller. Content key written only in the VERIFYING→READY commit, so a failed commit leaves bytes under a key no record names → `reconcilePrePublishWithKey` (`DefaultFileReconciliationService.java:108-155`) unreachable (tests fabricate a VERIFYING record with a key). `AmbiguousCompletionException` from `publish` (`AtomicMoveContentPublisher.java:63-80`) never reaches the queue.
- Why it matters: `docs/fileserver/operations.md:71-76`, `security.md:72-75` promise this recovery.
- Fix: choose `ContentKey` in the application and persist it on VERIFYING before the move (write-ahead); add a scheduled worker calling `reconcileRecoverable` and draining `RecoveryQueue`.
- Status: new
4. **[Required][correctness] `cancel` is a partial copy of the delete logic** — `DefaultUploadApplicationService.java:250-270`
- Evidence: `markDeleting` accepts READY (`FileTransitionRepository.java:68-69`) but cancel only queues `forStaging`; `cleanStaging` never moves DELETING→DELETED.
- Why it matters: tus DELETE on a finished upload (or single-shot rollback after `releaseLease` fails post-READY) leaves published bytes and quota forever, record stuck in DELETING; every cancelled upload stays DELETING.
- Fix: refuse cancel after UPLOADED, or reuse `DefaultFileLifecycleService.delete` retire step (`:84-104`) + DELETED transition in staging cleanup.
- Status: new
5. **[Required][correctness] Abandoned uploads never cleaned up** — `cleanup/CleanupType.java` (`EXPIRED_UPLOAD`)
- Evidence: no producer of `EXPIRED_UPLOAD` (`grep enqueue(`); `LocalOrphanScanAdapter` walks only `contentRoot` (`:64,80`); `forOrphan` items marked done without deleting (`DefaultCleanupService.java:155-158`). tus accepts `Upload-Defer-Length: 1` (`TusRequestParser.java:48-50`) but no PATCH sets the length, so such uploads can never finish.
- Why it matters: unbounded disk growth any client can trigger.
- Fix: sweep sessions past `expiresAt` → EXPIRED + `EXPIRED_UPLOAD`; orphan items actually delete; reject `Upload-Defer-Length` until implemented.
- Status: new
6. **[Required][correctness] Terminal sessions treated as live** — `DefaultUploadApplicationService.java:319-328`
- Evidence: `UploadSession.terminal()` never read (grep `\.terminal()` → 0). Append to cancelled/rejected upload → `acquireLease` refused by SQL → `JpaUploadSessionStore.java:146-150` reports "writer lease is held by another owner", `retryable=true`; tus HEAD still 204 with an offset.
- Fix: `requireActiveSession` and `status` return 410/404 when `terminal()`.
- Status: new
7. **[Required][performance] Every append rehashes the whole prefix — O(n²) IO** — `src/adapter/outbound/fileserver/.../LocalBlockingContentStore.java:365-374`
- Evidence: `reattach` builds a new `LocalUploadHandle` per request (`DefaultUploadApplicationService.append:209-211`) so `digestAt` misses and `rehashPrefix` re-reads 0..offset (`LocalAppendEngine.java:55-58`); finalize reads the file twice more (`stagingDigest:346-362`, `verifyStaged:41`) on the servlet thread outside `transferExecutor` (`TusController.java:178`).
- Why it matters: 10 GiB in 5 MiB chunks ≈ 10 TiB re-read.
- Fix: bounded per-uploadId digest cache keyed by offset; reuse finalize digest in `verifyStaged`.
- Status: new
8. **[Required][security] Authorization checks role tiers only** — `RoleBasedFileAccessPolicy.java:44-61`
- Evidence: `descriptor` argument unused; `FileDescriptor` has no owner; all controllers upload into `defaultNamespace()`. Any writer can cancel/delete/move anyone's file; any reader downloads everything. `FileOperation.FINALIZE` never authorized (grep 0) contrary to `security.md`. Admin reads use `READ_METADATA` (`DefaultFileserverAdminService.java:209-211`), mitigated by transport `RestrictedPathRule`.
- Fix: `ownerPrincipalId` on `FileRecord`/`FileDescriptor` checked by the default policy; authorize FINALIZE; admin tier for admin reads.
- Assumption: multiple mutually untrusted users. / Counterargument: may be a shared-drive template — then docs must say so, and the port still gives adopters no owner data.
- Status: new
9. **[Required][correctness] Upload concurrency effectively 4 per node; limit returns 413** — `DefaultTransferAdmissionController.java:48-56`
- Evidence: all entry points share one default namespace scope → `scope-upload-permits=4` (`fileserver.yml:89`) caps the node, `instance-upload-permits=16` never reached; limit throws `QUOTA_EXCEEDED` (→413, `FileserverErrorCode.java:37`) which clients treat as permanent. `JpaFileQuotaService.reserve` checks no limit.
- Fix: `TRANSFER_ADMISSION_REJECTED` (429) for concurrency; scope by principal.
- Status: new
10. **[Required][tests] persistence-fileserver's race-critical SQL untested** — `src/adapter/outbound/persistence-fileserver` (0 test files)
- Evidence: only coverage is `persistence-jpa/src/postgresqlIntegrationTest/.../PostgreSqlFileserver*` (`persistence-jpa/build.gradle:86-87`); `grep 'claimForCleanup\|terminate(\|reclaimExpiredClaims'` in tests → in-memory fakes only. `UploadLeaseRepository.terminate/claimForCleanup`, V4 `ACTIVE` clauses, claim-token settlement never run against real SQL.
- Fix: tests in the leaf's own strict lane: acquire-after-terminate, claim-while-leased, stale-token settlement.
- Status: new
11. **[Consider][correctness] Draft-12 create ignores the POST body** — `Draft12UploadController.java:84-94,171-174` (Content-Length taken as final length, body dropped; empty POST → 0-byte upload). Fix: read `Upload-Length` or append the body.
12. **[Nit][docs]** `adapter/outbound/fileserver/CLAUDE.md` lists 2 deps and says bootstrap owns the platform (actual: 4 deps, composition owns it); `composition/fileserver-platform/build.gradle` comments name `persistence-jpa`/`adapter:inbound:web`.
#### Leads checked and dismissed
- Path traversal: `DefaultPhysicalPathResolver` (server key, shape regex, normalize + re-check) and Nginx mapper sound.
- M4 publish TOCTOU, C2 verification fail-open, L2 no-progress spin: no regression.
- Append rollback truncates+forces, escalates to ambiguous when unprovable. Correct.
- Size limits: multipart shares `max-file-size` placeholder; batch part count capped (MVC + reactive).
- `@EntityScan`/`@EnableJpaRepositories` follow repo pattern. `bin/` gitignored. Admin plane gated by `RestrictedPathRule` (`SecurityConfig.java:73-90`).
#### Coverage note
Read in full: upload/finalize/cancel/cleanup/reconciliation/admin services, append/publish/store, tus/draft-12/upload controllers, reactive handlers, persistence session/lease/quota/cleanup adapters, migrations V3/V4. Skimmed: R1/R2 publication, problem mapping, metrics, most composition configs. Left out: always-true null checks in `FileserverStartupConfiguration.java:61-66`; reactive `contextFactory::current` reads a thread-local (inert in servlet setup).
### Group D — Object storage
#### Per-leaf verdict
| leaf | verdict | one-line reason |
|---|---|---|
| `application-objectstorage` | PASS | Registry edges OK (application-core only); strong identity parsing with tests. Fix #1 will likely add a fence field to `ObjectFinalizeRequest`. |
| `adapter-outbound-objectstorage` | NEEDS_FIX | No HARD-STOP. Staged-publication kernel + handoff service (wired, used by the sample) have a data-deletion path, non-idempotent replays, crash-recovery gaps, and no behavioral tests. |
#### Findings (most severe first)
1. **[Required][correctness] Abort can delete data behind an already-published reference; finalize not fenced by the handoff** — `src/adapter/outbound/objectstorage/.../kernel/StagedObjectPublicationKernel.java:257-269`
- Evidence: `abort` calls `provider.retire(...)` before `terminate(operation, ABORTED)`; `PUBLISHED` is terminal so `ObjectOperationRecord.java:143-148` throws — after the delete. Local-dev really deletes (`LocalDevObjectDataStore.java:147-155`). `ObjectPublicationHandoffService.authorizeAbort` (`:100-113`) reads only the stage record; `ObjectPublicationHandoffRecord.java:136-139` allows abort when released or lease expired. `finalizePublication` (`:183-237`) never reads the handoff; `ObjectFinalizeRequest` has no fence.
- Result: claim → finalize → release → authorizeAbort → abort deletes the file while manifest/pointer/lifecycle stay PUBLISHED; same with a late finalize racing an abort after lease expiry.
- Severity note: only `filesystem-local-dev` is deployable and S3 `retire` is UNSUPPORTED (`S3ManagedObjectProvider.java:121-125`); becomes Critical once a production provider implements retire.
- Fix: load the operation first and reject `REFERENCE_PUBLISH_IN_PROGRESS`/`PUBLISHED`; CAS an ABORT_IN_PROGRESS fence before the destructive call (like `DirectMultipartCoordinator`); add fence+claimant to `ObjectFinalizeRequest` and require an active matching lease; `authorizeAbort` rejects once in publish phase.
- Status: new
2. **[Required][correctness] Handoff release, authorizeAbort and staged abort replays are not idempotent** — `.../control/ObjectPublicationHandoffRecord.java:119-120,165-171`
- release on an already-released record → `requireOwner` throws "owner/fence is stale"; the sample calls `releaseClaim` on every reconcile of a terminal READY intent (`ReconcilePosterImageUploadUseCase.java:96-97`) → second reconcile always throws.
- authorizeAbort `:137` throws if released with a future lease, otherwise writes a new revision and never reports `REPLAYED`.
- second `abort`: `deleteIfExists` false → `StagedObjectPublicationKernel.java:264-265` throws "could not be retired"; `AbortPosterImageUploadUseCase` (marked `IDEMPOTENT`, `:28`) re-calls abort (`:91-92`) while ABORT_AUTHORIZED → stuck permanently. Contrast `DirectMultipartCoordinator.abortMultipart:322-324` returns `REPLAYED`.
- Fix: same terminal state + same fence/claimant → return `REPLAYED` without writing; abort returns `REPLAYED` if operation already `ABORTED`, before touching the provider.
- Status: new
3. **[Required][correctness] Four crash windows leave staged operations stuck** — `.../kernel/StagedObjectPublicationKernel.java`
- stage: operation advanced to DATA_UPLOADED (`:123-127`) before stage record created (`:139`) → replay `loadStage` null → `:91-92` "not uploadable"; data orphaned and un-abortable.
- verifyIntegrity: stage CAS (`:161`) before operation advance (`:166`) → replay REPLAYED at `:156-157`, later finalize fails at `:206`.
- finalize: once REFERENCE_PUBLISH_IN_PROGRESS (`:210`), `:206-207` throws "not finalizable" on every retry.
- scan verdict: same ordering in `RoutingObjectScanMaintenanceAdapter.java:84-110`.
- The managed path (`RoutingObjectMutationAdapter.publish:140-154`) recovers from these states.
- Fix: operation record as single source of truth; secondary writes create-or-require-same; replay path completes the missing advance; finalize accepts REFERENCE_PUBLISH_IN_PROGRESS and re-runs idempotent `references.publish`.
- Status: new
4. **[Required][tests] Staged kernel and handoff service have no behavioral tests** — `.../test/.../kernel/StagedObjectPublicationKernelTest.java:11-23` (tests only `ObjectScanVerdictPolicy`); grep for `ObjectPublicationHandoffService`/`finalizePublication`/`authorizeAbort` in adapter tests → 0; `ObjectPublicationHandoffRaceTest` one takeover case; bootstrap `ObjectStorageCompositionIntegrationTest:73-83` legacy put/get only; sample `PosterImagePublicationWorkflowTest` mocks ports. Fix: local-dev full-flow + replay tests, failure injection between the two writes (wrap `ObjectControlStore`). Status: new
5. **[Consider][correctness] Local control-store `create` not atomic** — `.../filesystem/LocalDevObjectControlStore.java:51-65` (`CREATE_NEW` then write; concurrent reserve reads empty/partial file → corrupt decode `CanonicalJsonObjectControlRecordCodec.java:72-73`; crash mid-write leaves key corrupt forever). Fix: temp file + force + `Files.createLink` as in `LocalDevObjectDataStore.create:55-60`. Assumption/Counter: few concurrent retries in local-dev / README claims restart+corruption characterization.
6. **[Consider][security] Legacy filesystem adapter: unbounded read + absolute server path in HTTP response** — `.../FilesystemObjectStorageAdapter.java:61,71`
- `:71` still `Files.readAllBytes` — remediation master records L3 closed, but `git log` shows the file unchanged since `821fe00c`/`5f10b791`.
- `:61` returns `target.toUri()` (`file:///<absolute>/...`); `PosterWebMapper.java:35` puts it in the HTTP `location` field; sample `application.yml:324-327` enables this backend by default.
- Fix: logical URI (`objectstorage:` + key); check `Files.size` against a configured max before reading.
- Status: regression of L3 (closure record does not match code); path leak new.
7. **[Consider][structure] ~3.2k lines of direct/multipart code unreachable, plus empty placeholders**
- 0 main callers: `DirectTransferCoordinator`, `DirectMultipartCoordinator`, `ManagedMultipartCoordinator`, `MultipartOperationResolver`, `S3ObjectOperationResolver`, `ObjectOperationResolutionService`, `ObjectOperationEpochRecord`; `SelectedObjectStorageProviderFactory.{managedMultipart,directTransfer,directMultipart}Provider(..)` only from `S3ProviderCompositionTest`. `RoutingObjectDirectGrantAdapter` 0 references anywhere; `ObjectStorageMaintenanceCapabilityConfig` empty `@Configuration`. Ports `DirectObjectUploadPort`, `DirectObjectDownloadGrantPort`, `DirectMultipartUploadPort` have no consumers; `ObjectPurgeMaintenancePort` no implementation.
- S3 contribution resolves credentials with `System::getenv` (`S3ObjectStorageProviderContribution.java:41`) instead of `SecretSourcePort` (same pattern fixed for httpclient H6).
- Fix: delete the two empty types; stop constructing the unused multipart provider/presigner (`:130-158`); move direct/multipart code+ports out of main until a consumer exists; `SecretSourcePort` before S3 promotion.
- Assumption/Counter: no fork needs direct/multipart before S3 promotion / code captures worked-out S3 behavior.
- Status: new
8. **[Consider][readability] Publication/identity logic duplicated** — `.../config/RoutingObjectMutationAdapter.java:354-421,473-505` (`persistPublishedReference`/`createOrRequireSame` ≈ `ObjectReferencePublicationService.java:20-92`; `objectId`/`sha256Hex` ≈ `StagedObjectPublicationKernel.java:389-415`; `split("\\.")[1]` at `:504` and `RoutingObjectReadAdapter.java:103`; holds a `clock` but calls `Instant.now()` 10×). Fix: route through `ObjectReferencePublicationService`, one identity helper in `codec`, `ObjectReferenceCodec.routeToken(..)`, inject clock.
9. **[Consider][correctness] `inspect` always reports scan state NOT_REQUIRED** — `.../config/RoutingObjectReadAdapter.java:53` (SCAN_CLEAN destinations require CLEAN at finalize, `StagedObjectPublicationKernel.java:191`). Fix: derive from destination `publicationRequirement`.
10. **[Nit][correctness] Close order not actually reversed** — `.../config/ObjectStorageCapabilityAssembler.java:74,113-115` (`Map.copyOf` unordered; one throwing close skips the rest). Fix: ordered list + try/catch + `addSuppressed`.
11. **[Nit][docs]** `CLAUDE.md:30` allowed deps omit `application-objectstorage`; `CLAUDE.md:17,40` say `application-core` for types now in `application-objectstorage`; `CLAUDE.md:8` Boot 4.0.0 vs 4.0.8; `README.md:40` says local-dev rejected under "prod/production" (code: `prod` only); `.tmp` "report-first reconciliation" comments (`LocalDevObjectDataStore.java:81`, `LocalDevObjectControlStore.java:100`) point to non-existent code.
#### Leads checked and dismissed
- HARD-STOP 5: all edges allowed.
- Key traversal/namespace escape: fixed regex (`ObjectStorageProviderOperation.java:36-45`, `LocalObjectPathGuard.java:119-136`), root confinement, symlink rejection; references have fixed format + checksum.
- Presigned grants redact `toString`; `DirectGrantReplayCache` bounded (M4 holds). S3 blocked by readiness gate (H7 holds); L8 holds.
- `resolve()` objectId formula matches staged kernel; `LocalObjectStreamTransfer.write` checks digest before linking; direct multipart abort idempotent; Ed25519 approval checks two distinct keys + validity window.
#### Coverage note
Read in full: kernel (staged/handoff/operation), routing adapters, local-dev provider, capability config/assembler, legacy filesystem adapter, identity/request contracts, diff since baseline. Skimmed: S3 providers, codecs, direct/multipart coordinators, legacy adoption.
### Group E — JPA / PostgreSQL persistence
#### Leaf별 판정
| leaf | verdict | 한 줄 사유 |
|---|---|---|
| `src/adapter/outbound/persistence-jpa` | **BLOCKED** | Critical 2건: 이미 적용된 Flyway 마이그레이션을 제자리 수정, durable operation 재제출이 PostgreSQL에서 트랜잭션을 깨뜨림. HARD-STOP 위반 없음(엔티티·리포지토리가 inbound/bootstrap으로 새지 않음). |
| `src/config/jpa/readiness-cards.yaml` | PASS | 외부 prerequisite 하나가 없는 카드를 가리킴(FYI). |
#### 발견 (심각도 순)
1. **[Critical][correctness/migration] 이미 적용된 versioned 마이그레이션 세 개를 제자리에서 수정** — `src/adapter/outbound/persistence-jpa/src/main/resources/db/migration/postgresql/V9__widen_capability_schema_stream.sql:3-8`, `.../postgresql/V13__outbox_canonical_transport_compatibility.sql:18-26,71`, `.../jpa/outbox-storage/V1__initialize_or_adopt.sql:140,218,296`
- 근거(`git log --follow`): V9는 2026-08-15 생성, 9da00eb7에서 주석 변경. V13은 60b6a319(09-18) 생성, 9da00eb7에서 `char(64)→varchar(64)`, `NOT VALID` 추가. outbox-storage V1은 07-31 생성 후 트리거 `FOR EACH STATEMENT`, 파티션 제거, digest sha256으로 변경. V9 새 주석은 "renumbering an applied production migration would invalidate Flyway history"라면서 그 수정 자체가 checksum을 바꿈.
- 왜 문제인가: Flyway SQL checksum은 주석 포함 파일 전체 CRC. `validate-on-migrate`는 Boot 기본값(true)이고 끄는 설정 없음 → 이전 파일이 적용된 모든 DB(local 프로필도 PostgreSQL+Flyway → 개발자 DB, 파생 서비스 DB)가 기동 시 checksum mismatch로 실패. 09-18판 V13이 적용된 DB는 컬럼이 `bpchar`로 남아 `ddl-auto: validate`에서도 실패.
- 수정: 세 파일을 적용 당시 바이트로 복원(V9 `944a1e34`, V13 `60b6a319`). H3·M5 수정은 새 버전(`V16__…` ALTER TYPE + 재검증, outbox-storage `V3__…`)으로. 적용된 버전 파일 해시를 고정하는 단위 테스트 추가.
- Status: new (H3/M5 수정 중 생김)
2. **[Critical][correctness] 같은 요청 재제출이 PostgreSQL에서 트랜잭션을 깨뜨림** — `src/adapter/outbound/persistence-jpa/.../operation/DurableOperationStoreAdapter.java:47-78`
- 근거: `saveAndFlush`에서 `uq_durable_operation_scope`(23505) → `catch (DataIntegrityViolationException)` 안에서 같은 트랜잭션으로 `findByTenantAndPrincipal…` 재조회. `SubmitDurableOperationUseCase.submit`(application-core:87-99)이 미리 조회하지 않으므로 경쟁뿐 아니라 **모든** 동일 재제출이 이 경로. PostgreSQL은 에러 난 트랜잭션의 후속 문장을 25P02로 거부(`autosave` 설정 없음, grep 0). `SimpleJpaRepository.saveAndFlush`의 `@Transactional`이 caller 트랜잭션을 rollback-only로 표시하는 문제도 겹침. 같은 leaf `JpaProviderEventLedger.java:72-76`은 이 함정을 설명하고 `ON CONFLICT DO NOTHING`으로 회피.
- 왜 문제인가: "resubmission being safe" 계약이 실제 DB에서 500. in-memory fake만 통과; PostgreSQL 계약 테스트(`DurableOperationStoreContractTest:183-188`)는 adapter를 거치지 않고 raw SQL로 충돌만 확인.
- 수정: native `INSERT … ON CONFLICT (tenant, principal, operation_name, request_hash) DO NOTHING`; 0건이면 기존 row 조회. 실제 PostgreSQL에서 adapter 경유 이중 submit 테스트.
- Status: new (원래 있던 결함, M6 수정 때 누락)
3. **[Required][correctness] `CanonicalOutboxAppendAdapter.append`에 caller 트랜잭션 확인 없음** — `.../outbox/CanonicalOutboxAppendAdapter.java:36-77`
- 근거: 형제 `OutboxStoreAdapter.append`(:55-56)에는 `requireTransaction` 있음. 포트 Javadoc은 "participates in the caller's existing write transaction" 요구. `repository.save`는 트랜잭션이 없으면 스스로 열고 커밋. `PersistenceTransactionOwnershipTest`, `CanonicalOutboxAppendAdapterTest` 모두 미커버.
- 왜 문제인가: 잘못 호출되면 outbox row가 비즈니스 쓰기와 따로 커밋 → 조용한 dual-write.
- 수정: 같은 guard 추가; 세 벌(outbox/operation/liveevent)로 복사된 guard를 `transaction` 패키지 공용 helper로; ownership 테스트에 이 adapter 추가.
- Status: new (H7 수정이 형제 클래스에만 적용)
4. **[Required][correctness+tests] 전체/단건 읽음 처리가 ARCHIVED inbox 항목을 되살림** — `.../notification/platform/inbox/InboxItemJpaRepository.java:83-99`, `InboxItemEntity.java:107-127`
- 근거: `markAllRead` 서브쿼리가 `read_at IS NULL`만 보고 state로 거르지 않는데 `archive()`는 `readAt`을 설정하지 않음 → 읽지 않고 보관한 항목이 `state='READ'`로 바뀌어 기본 목록에 재등장. `unreadCount`는 ARCHIVED 제외(배지 0)인데 markAllRead는 그 항목을 갱신. 엔티티 `markRead`/`markSeen`도 ARCHIVED 미차단. `JpaNotificationInbox`/`InboxItemJpaRepository` 실행 테스트 없음; `PostgreSqlInboxCutoffIntegrationTest:86-105`는 운영 쿼리를 **복사한 SQL 문자열**을 돌리고 enum에 없는 `'UNREAD'` 사용.
- 수정: 서브쿼리에 `AND state NOT IN ('ARCHIVED','EXPIRED','DELETED')`; 엔티티 전이에서 terminal/ARCHIVED 차단; 통합 테스트는 실제 repository로.
- Status: new
5. **[Required][correctness] canonical 이벤트가 legacy claim을 타면서 aggregate 순서와 tenant 범위를 잃음** — `.../postgresql/PostgreSqlOutboxClaimRepository.java:15-29`, `README.md:177`
- 근거: FIFO gate가 `aggregate_id`와 `occurred_at <`만 봄. `CanonicalOutboxAppendAdapter`가 저장하는 `tenant_scope`/`aggregate_type`/`aggregate_sequence`/`event_index`를 claim이 안 씀(`AggregateOrder`는 "Total order within one aggregate identity"). `occurred_at`이 같은 두 이벤트는 함께 claim, 순서 미정(`...RowLifecycleIntegrationTest:311`이 "LegacyFifoLimitation"으로 기록; README는 "aggregate 당 최대 한 row"). 다른 tenant/type의 같은 `aggregate_id` DEAD row가 서로 영구 차단.
- 수정: canonical row(`contract_id IS NOT NULL`) gate를 `(tenant_scope, aggregate_type, aggregate_id)` + `(aggregate_sequence, event_index)` 순서로; 부분 유니크 인덱스; README 수정.
- 가정/반론: 한 command가 같은 aggregate에 같은 `occurredAt` 이벤트를 2개 이상 냄 / 이 경로 플래그 기본 off(`outbox.yml:22`), V2 polling 경로는 이미 해결.
- Status: new
6. **[Required][tests/security] H10 tenant 규칙이 필드 이름 `tenantId`로만 판정** — `src/adapter/outbound/persistence-jpa/src/testFixtures/.../testkit/arch/TenantScopedRepositoryPredicate.java:19,48-50`
- 근거: Javadoc "Structural rather than name-based"인데 `TENANT_FIELD = "tenantId"` 이름 비교. `DurableOperationEntity.tenant`, `IdempotencyRecordEntity.tenant`, `OutboxEventEntity.tenantScope` 누락; 세 리포지토리 모두 `extends JpaRepository`(findAll/deleteAll 상속).
- 수정: `tenant*` 필드/컬럼 모두 또는 marker annotation; 세 리포지토리를 `Repository` marker로 축소.
- 가정/반론: 템플릿 사용자가 상속 CRUD를 호출 / 현재 adapter는 범위 지정 쿼리만 사용.
- Status: new (H10 종결 범위의 빈틈)
7. **[Consider][correctness] 잘못된 canonical row 하나가 relay 전체를 멈춤** — `.../outbox/OutboxStoreAdapter.java:73-85,161-170` (배치 중 `IllegalStateException` → claim 트랜잭션 전체 롤백 → 같은 row가 매 주기 재선택). 수정: 해당 row DEAD 전이 후 계속. 가정/반론: DB CHECK로 못 막는 불일치(payload/envelope, 수동 UPDATE) / V13·V14 CHECK가 대부분 차단.
8. **[Consider][performance] outbox 지표 쿼리가 5초마다 테이블 전체를 셈** — `.../outbox/OutboxEventJpaRepository.java:21-27` (`countGroupedByStatus` 무필터, 7일치 PUBLISHED 포함, `OutboxRelayScheduler` 매 poll(PT5S) 호출; `OutboxReaper.deletePublishedBefore` 무제한 단일 트랜잭션 삭제). 수정: `status <> 'PUBLISHED'`만 집계; 삭제는 `LIMIT` 루프.
9. **[Consider][structure] `JpaObservabilityAutoConfiguration` 미사용 코드** — `.../autoconfigure/JpaObservabilityAutoConfiguration.java:26` (자기 자신 외 참조 0, package-private; Javadoc이 bootstrap 타입 링크). 수정: 삭제. 반론: M10에서 고급 capability 경로로 남겼을 수 있음.
10. **[Consider][security] experimental tenant guard의 "audited" admin scope가 기록을 남기지 않고 게이트 없이 생성됨** — `.../experimental/tenant/TenantAwareRepositoryGuard.java:36-54` (reason을 ThreadLocal에만 둠; public 생성자라 `ExperimentalEntryConsentTest:34-36` 게이트 목록에서 빠짐; `SchemaMultiTenantConnectionProvider.releaseConnection:50-57`은 reset 실패 시에도 tenant `search_path` 커넥션을 pool에 반환). 수정: audit sink + `enabledBy(gate, …)` factory; reset 실패 시 커넥션 `abort`. 반론: 운영 호출처 없는 experimental 코드.
11. **[FYI][docs] readiness card 외부 prerequisite가 없는 카드를 가리킴** — `src/config/jpa/readiness-cards.yaml:434-439` (`messaging-cdc-dispatch.v1`이 messaging 카드 11개에 없음; `minimum-readiness: R2` 어휘가 messaging `maturity:`와 불일치; `JpaReadinessRegistryValidator:169-179`는 형식만 검사; `postgresqlOutboxCdcCleanupIntegrationTest` 태스크 부재). 수정: `not-implemented` 동안 prerequisite 제거 또는 validator에 대상 파일 조회 추가.
12. **[Nit][docs]** `CLAUDE.md:237-243` 벤더 SQLState 표에 `55P03 → DB_LOCK_TIMEOUT` 누락(`PostgreSqlSqlStateErrorMapping.java:43`에는 있음).
#### 점검 후 기각한 lead
- SQL 문자열 조립: `set_config` 값은 bind parameter; `NamedStatementInspector`가 `*/` 차단; identifier는 validate된 registry 경유.
- `PersistenceExceptionTranslator` cause chain 탐색·unknown→empty 정상; `TransactionRetryClassifier` 40001/40P01만 재시도 정상.
- HARD-STOP 2/3: inbound의 persistence import 0; bootstrap은 SPI/config 타입만.
- live-event position: counter row 락이 커밋까지 유지 → 커밋 순서 = position 순서, gap 없음.
- 이전 `JpaModuleBoundaryTest` `auditing` 잔여물 해소. `JpaProviderEventLedger.bindUnmatched` stale 엔티티는 해시 매칭으로 자기 복구.
#### 커버리지
자세히: outbox, operation, liveevent, inbox, transaction, failure/h2/postgresql 에러 매핑, autoconfigure, experimental tenant/rls, 변경된 마이그레이션(V9, V13, V14, V15, outbox-storage V1/V2), 관련 테스트, readiness cards. 훑음: notification platform 저장소 나머지, `PostgreSqlOwnerSafeIdempotencyStore`(claim 경로), springdata/hibernate 패키지. Flyway checksum 실패·25P02는 코드와 벤더 동작 근거 판단(실제 DB 재현 없음).
### Group F — MongoDB persistence
#### Per-leaf verdict
| leaf | verdict | one-line reason |
|---|---|---|
| `adapter-outbound-persistence-mongo` | NEEDS_FIX | No HARD-STOP violation; zero-dependency edge confirmed. Operator injection, multi-profile, bulk mapping, CSFLE and change-stream paths have correctness/security holes; tenancy docs promise enforcement the code does not have. |
#### Findings (most severe first)
1. **[Required][security] Operator injection through an `Object` value on the equality paths** — `src/adapter/outbound/persistence-mongo/.../query/PolicyAwareMongoQueryBuilder.java:43-47`, `.../imperative/atomic/AtomicFilter.java:66-74`
- Evidence: `whereEquals(String path, Object value)` → `Criteria.where(path).is(value)`; `AtomicFilter.toQuery()` passes values through `.is(entry.getValue())` with no type check. Javadoc (`:16-19`) claims NoSQL operator injection is "a compile-time impossibility". A `Map`/`Document` such as `{"$ne": null}` or `{"$regex": ".*"}` is sent as an embedded document the server reads as an operator — bypasses the field operator allowlist and `MongoRegexPolicy`; `AtomicFilter.id(Map.of("$gt",""))` defeats the single-document guarantee. No test covers it; the security lane (`build.gradle:189`) claims "injection" coverage but `MongoSecurityIntegrationLaneTest` has none.
- Fix: one shared `requireScalar(value)` guard used by the builder and `AtomicFilter` (reject `Map`/`Bson`/`Document`, and collections for EQ) + regression test with `Map.of("$ne", …)`.
- Status: new
2. **[Required][correctness] With more than one profile, startup and readiness ignore `active-profile`** — `.../autoconfigure/MongoStartupValidator.java:69-92`, `.../autoconfigure/MongoPlatformHealthIndicator.java:86-89`
- Evidence: `active-profile` is read only by the client customizer via raw `environment.getProperty` (`client/MongoProfileClientSettingsAutoConfiguration.java:52-53`). Validator iterates every profile and throws on topology mismatch; health uses `anyMatch(profile.topology() != observed)`. Multi-profile is a tested setup (`MongoClientSettingsFactoryTest.java:117`); validator/health tests use one profile only. `primary` (REPLICA_SET) + `local` (STANDALONE) → startup always fails, or health permanently DOWN.
- Fix: bind `activeProfile` once into typed settings; customizer, validator and health use only that profile; add a two-profile test.
- Status: new (side effect of the H1 fix)
3. **[Required][correctness/data loss] Change-stream pipeline and consumer lose events** — `.../changestream/MongoChangeStreamPipeline.java:48-64`, `.../changestream/consumer/ReactiveMongoChangeStreamConsumer.java:174-179`
- (a) `advancesPosition` passes only events strictly after the high-water mark (`isAfter`); events of one multi-document transaction share a `clusterTime`, so the 2nd+ are dropped. `MongoChangeStreamPipelineTest.twoEventsInOneTransactionAreDistinct` checks identity only, never the pipeline.
- (b) `rawOf` passes only `fullDocument`; a delete arrives as `{}` without `operationType`/`documentKey`, contrary to `MongoChangeProjector` Javadoc.
- (c) `BUSY`/`PARKED` do not advance the checkpoint (`projector/MongoChangeStreamRunner.java:293-297`) but `concatMap` continues and the next event's checkpoint moves past the skipped one.
- (d) `MongoResumeCheckpoint.belongsTo()` has no production caller; factories fill `clusterIdentity` with `""` (`:186-194`), so the cross-cluster token check never runs.
- Why Required not Critical: unreachable in the shipped bootJar (reactive excluded, `change-streams=true` refused) — but it is the consumer the README tells forks to adopt.
- Fix: accept equal `clusterTime` and dedupe by identity; pass the full change envelope; error on a result that cannot advance the checkpoint; enforce `belongsTo` in `resumeFrom`.
- Status: new
4. **[Required][correctness] Bulk path skips Spring mapping** — `.../imperative/bulk/MongoBulkExecutor.java:75-86`, `.../imperative/bulk/MongoBulkWritePlan.java:65-69`
- Evidence: bulk builds `UpdateOneModel` from raw `getQueryObject()`/`getUpdateObject()` and calls the driver; atomic path uses `operations.updateFirst(query, update, documentType, collection)` (`MongoAtomicOperationsTemplate.java:137-139`) which applies `@Field` names and custom conversions. `DomainDocumentId`/`BigInteger` values have no codec (estimator `toJson()` fails); `@Field` aliases are written under the wrong name. Tests use String values only.
- Fix: take `documentType` and map via `QueryMapper`/`UpdateMapper`, or reuse `MongoOperations.bulkOps(...)`; estimate on the mapped result.
- Status: new (MNG-026 fixed the policy bypass; mapping bypass remains)
5. **[Required][security] `MongoCsfleClientFactory` ignores the profile's encryption policy** — `.../advanced/encryption/csfle/MongoCsfleClientFactory.java:40-47`
- Evidence: `schemaMap` keyed by bare `profile.collection()` while the driver matches `db.coll` → nothing auto-encrypted; `profile.fields()` never read; `keyVaultCredential` not passed to `keyVaultMongoClientSettings` (contradicts `MongoCsfleProfile` Javadoc). No test, no caller.
- Fix: key by namespace, generate schema from `fields()`, set key-vault settings from the separate credential, unit-test the generated settings.
- Status: new
6. **[Required][security/docs] Shared-collection tenancy enforcement point does not exist** — `.../advanced/tenancy/shared/TenantScopedMongoOperations.java:19`, `docs/mongodb/advanced/multi-tenancy.md:13-14,79`
- Evidence: no implementation or caller of `TenantScopedMongoOperations`; `MongoTenantPredicateInjector` called only from a test; docs say it is the entry point so callers "cannot construct an unscoped operation by forgetting". The injector itself is sound (fails closed on duplicate criteria, overwrites caller `tenantId`).
- Fix: ship an implementation wrapping atomic ops + executor as the only route, or delete the interface and rewrite the doc.
- Assumption: adopters trust the doc. / Counterargument: Advanced, flag-gated; fork expected to implement.
- Status: new
7. **[Consider][performance] Reactive cursor re-serializes every row twice** — `.../reactive/cursor/MongoReactiveCursorPublisher.java:78-80` (`converter.write` + `toBsonDocument` in `doOnNext` just to measure bytes; fails for projection proxies). Fix: `RawBsonDocument` or decode-time hook. Status: new (side effect of M4 fix)
8. **[Consider][correctness] Migration lock swallows all write errors as "lost the race"** — `.../migration/MongoCollectionMigrationLock.java:112`. Fix: return empty only for code 11000, rethrow otherwise. Status: new
9. **[Nit][readability/docs]** `MongoPlatformAutoConfiguration.java` has 41 lines of inline FQNs (forbidden by leaf CLAUDE.md) and an orphaned Javadoc at `:192-198`; `CLAUDE.md:46` says `implementation(project(...))` but it is `runtimeOnly` (`app-bootstrap/build.gradle:129`); `CLAUDE.md:53` broken sentence.
#### Leads checked and dismissed
- Zero allowed dependencies: main imports resolve only to `…adapter.outbound.mongo`; no `project(` in build.gradle; no re-declared application types.
- Regex: `Pattern.quote` on starts-with/contains; `whereMatches` validates; every query has `maxTimeMS`.
- Keyset cursor: HMAC + `MessageDigest.isEqual` + length cap. OK.
- Resume token: `toString` shows length only; invalidate → `startAfter`. OK.
- CSFLE key material / credential URI: not stored or logged.
- Prior H2 no regression; prior H6 now closed (16 files tracked).
- `DataAccessException` unwrap loses cause but classifies UNCLASSIFIED/non-retryable — safe.
- Tenant field updates blocked unless explicitly registered updatable.
#### Coverage note
Read in full: tenancy, CSFLE, changestream core/consumer/runner/pipeline, query builder/policy/regex, atomic, bulk, imperative executor, failure classifier, client settings/profile autoconfig, platform autoconfig, startup validator, health, topology probe, migration lock, diff since `944a1e34`. Skimmed: QE, search/vector, sharding, timeseries, gridfs, transaction, reactive executor, schema, observation, admin.
### Group G — Redis
#### Leaf별 판정
| leaf | 판정 | 한 줄 사유 |
|---|---|---|
| `redis-runtime` | NEEDS_FIX | lease 하나가 연결 하나를 독점하고 SCRIPT 한도 8 → Redis가 정상이어도 동시 9번째 요청부터 거부. 공개 안전 설정 약 20개가 무효. 자기 테스트 1개 |
| `redis-sdk` | NEEDS_FIX | 약 21.7k LOC인데 facade 구현·pub/sub 전송·조립 코드가 없어 도입 불가. 자기 테스트 0개 |
| `cache-redis` | NEEDS_FIX | rate-limit hash 무한 증가(M4 회귀), lease replay가 유효 시간 부풀림, 남의 테스트 64개 보유, realtime adapter는 죽은 코드 |
#### Findings (심각도 순)
1. **[Critical][performance/correctness] SCRIPT lane이 호출자마다 연결을 독점 → Redis 정상이어도 동시 9번째 호출부터 거부** — `src/adapter/outbound/redis-runtime/.../RedisRuntimeOwner.java:134-140`
- Evidence: `if (current >= limits.get(kind)) throw reject(...)`의 `current`는 연결 수가 아니라 대여 중 lease 수. 한도 `RedisSdkAutoConfiguration.java:279` `limits.put(SCRIPT, getMaximumScriptConnections())`, 기본값 `RedisSdkSettings.java:687` `= 8`. 호출부(`RedisEdgeRateLimitAdapter.java:109`, `RedisIdempotencyStoreAdapter.java:109`, `RedisDistributedLeaseAdapter.java:145`)가 RTT 전체 동안 lease 보유. `RedisScriptLaneCapacityTest`가 세 번째 동시 borrow 거부를 단언(동작 고정).
- Why it matters: rate-limit은 fail-closed → lane 포화 시 모든 요청 503/429; Redis 지연이 조금만 튀어도 전면 거부. 요청이 Redis에 전송되지 않았는데 idempotency·lease는 `Indeterminate` 보고(`RedisIdempotencyStoreAdapter.java:128-131`, `RedisDistributedLeaseAdapter.java:166-172`). `RedisHealthContributor.java:51`도 REGULAR lane에서 lease를 빌려 포화 시 `redisRequired` DOWN.
- Fix: REGULAR/SCRIPT lane은 공유 멀티플렉싱 연결 N개 round-robin + in-flight `Semaphore`/`requestQueueSize`로 동시성 제한; 독점 lease는 TRANSACTION/BLOCKING/PUBSUB만; borrow 거부는 `Unavailable`로 분류.
- Status: H8 회귀 (H8은 "한도 단위가 틀렸다"였는데 단위는 그대로 두고 64→8로 낮춤)
2. **[Required][correctness/performance] fixed-window/sliding-counter hash 무한 증가** — `src/adapter/outbound/cache-redis/.../ratelimit/RateLimitScripts.java:58-59,89-90`
- Evidence: `HDEL`은 정확히 `windowStart - 2w` 필드 하나만 지우고, 허용 요청마다 `PEXPIRE`로 TTL 갱신. 기본값(sliding, w=1s, TTL 7s)에서 1.5s 간격 요청 → 버킷 0, 3000, 6000…이 삭제 대상에 안 걸리고 key도 만료 안 됨. `:39-40` 주석("refreshed per hit이 아니다")은 코드와 반대. fake(`InMemoryRedisCommandGateway.java:1717-1753`)가 HDEL/PEXPIRE를 흉내내지 않아 미테스트.
- Why it matters: 활동 subject마다 메모리 누적; 같은 Redis의 idempotency/lease가 eviction·OOM 영향.
- Fix: fixed window는 `start`/`count` 두 필드 덮어쓰기; sliding은 `start`/`cur`/`prev` 회전; live lane HLEN 상한 테스트.
- Status: M4 회귀(수정 불완전)
3. **[Required][correctness] lease replay(status 2)가 서버 잔여 TTL을 무시하고 전체 TTL 재부여** — `.../lease/RedisDistributedLeaseAdapter.java:157-161`
- Evidence: `ReplayedSameOperation(handle(request, ownership, startedAt))`가 재시도 시점부터 `leaseTtl - drift` 부여; 스크립트가 돌려주는 잔여 `PTTL`을 버림. 같은 파일 `inspect()`는 M7 수정으로 `observedValidityOf(reply.remainingMillis())` 사용(`:213`).
- Why it matters: 첫 응답 유실 후 재시도 시 보유자가 서버보다 더 오래 lease를 가졌다고 믿음 → 두 보유자 겹침 가능.
- Fix: status 2에서 `observedValidityOf(reply.remainingMillis())`; `localValidityOf`의 TTL≤drift 검사를 Redis 호출 전에.
- Status: new (M7 계열)
4. **[Required][security/performance] 공개 Redis 안전 설정 약 20개가 바인딩·검증만 되고 적용 안 됨** — `src/adapter/outbound/redis-runtime/.../RedisSdkSettings.java:717,725`
- Evidence: 모든 getter를 `redis-runtime`/`redis-sdk`/`cache-redis`/`app-bootstrap` `src/main`에서 `\.<getter>\(\)` grep → 설정 클래스 밖 참조 0: `getMaximumInFlightBytes`, `getMaximumReplyBytes`, 모든 `getMax*`(`limits`), `getCollection`/`getAdvanced`/`getMaxBlock` timeout, `getAcquireTimeout`, `getTlsHandshakeTimeout`, `getOverflowPolicy`, `getOfflineQueueCommands`. SDK는 하드코딩된 `RedisOperationLimits.defaults()`(`RedisOperationLimits.java:56`). 공개 env key로 등록됨(`docs/registries/env-keys.yaml:1959-2109` `APP_REDIS_LIMITS_*` 11개, `:2409`, `:2455`).
- Why it matters: 운영자는 응답 크기·in-flight 상한으로 OOM을 막는다고 믿지만 아무것도 안 막음(closed된 H3와 같은 종류).
- Fix: 우선 설정·env key 삭제; 필요한 것만 `RedisOperationLimits.from(settings)`로 연결.
- Status: new
5. **[Required][tests/architecture] redis-runtime·redis-sdk 테스트가 cache-redis에 있어 해당 leaf focused check가 빈 초록불** — `src/adapter/outbound/cache-redis/build.gradle:22`
- Evidence: cache-redis 테스트 76개 중 64개가 `.../redis/sdk/**`(2481줄 `InMemoryRedisCommandGateway` 포함); `redis-sdk/src`에는 `main`만; `redis-runtime` 테스트는 `RedisSdkAutoConfigurationRegistrationTest` 1개; 두 leaf 모두 CLAUDE.md/README 없음.
- Why it matters: CLAUDE.md는 "소유 leaf의 `:test`를 돌려라" — redis-sdk는 NO-SOURCE, runtime은 1개(HARD-STOP 6 위험). 게이트 위반은 아님(`ArchitecturePlugin.java:22-23`은 production 선언만 검사).
- Fix: `sdk/**` 테스트를 소유 leaf로 이동, fake는 `java-test-fixtures`로 공유, `testImplementation` edge 제거, leaf별 CLAUDE.md 추가.
- Status: new (build.gradle 주석에 "temporarily")
6. **[Required][structure] "optional SDK"와 realtime adapter를 실제로 쓸 수 없음** — `src/adapter/outbound/redis-sdk/.../api/RedisOperations.java`, `src/adapter/outbound/cache-redis/.../realtime/*`
- Evidence: `implements RedisOperations|ReactiveRedisOperations` main/test 0건; `implements RedisPubSubGateway`는 test `InMemoryRedisPubSubGateway.java:18` 하나뿐, runtime에 `connectPubSub`/`StatefulRedisPubSubConnection` 0건(구현 불가); `RedisConnectionRegistry`(redis-sdk) main 호출자 0이고 `RedisRuntimeOwner` lane 한도 로직 중복; `RedisEphemeralFanoutAdapter`, `RedisConnectionRegistryAdapter` 조립처 없음(test 외 grep 0).
- Why it matters: README는 "SDK를 composition에 넣을 때만" 적용된다는데 넣을 대상이 없음. 약 22k LOC가 유지보수·리뷰 표면만 차지.
- Fix: redis-sdk와 realtime 패키지 삭제; 필요 시 facade 구현 + PUBSUB 연결과 함께 재도입.
- Assumption/반론: 6개월 내 SDK 소비자 없음 / C3 closure가 "optional library 유지" 결정 — 단, 도입 가능하다는 전제 위.
- Status: new (C3 closure 미포함)
7. **[Consider][architecture] runtime에 SDK 표면이 섞여 있음** — `redis-runtime/.../RedisCommandGateway.java`(1227줄), `LettuceRedisCommandGateway.java`(1282줄) (메서드 118개 중 cache-redis main이 쓰는 건 약 15개; raw/admin 자격 증명·raw allowlist(`RedisSdkAutoConfiguration.java:119-142,182-193`)가 runtime에; runtime·sdk가 같은 패키지 9개 공유; `RedisCorrectnessRoles`가 consumer selector 값 하드코딩). Fix: runtime gateway는 semantic adapter에 필요한 메서드만, 나머지는 sdk로(#6 삭제 시 함께 사라짐); 패키지 `...redis.runtime.*`로.
8. **[Consider][readability] EVALSHA+NOSCRIPT 재로드 헬퍼 4벌 복사, 서로 조금씩 다름** — `CacheScripts.java:55-87`, `RateLimitScripts`, `LeaseScripts`, `IdempotencyScripts` (`CacheScripts`만 `startsWith("NOSCRIPT")` 폴백). Fix: `RegisteredLuaScript` 하나로 공유.
9. **[Consider][observability] 장애 로그가 JVM 수명 동안 한 번만 WARN, 예외 타입이 wrapper 이름** — `RedisSemanticFailureLogger.java:19,24-26` (`WARNED` 미초기화; `.get()`에서 잡아 `failureType`이 거의 항상 `ExecutionException`). Fix: cause unwrap, WARN을 시간 창 단위로.
10. **[Consider][security/ops] SCRIPT ACL probe가 스크립트 내부 명령 권한을 확인 안 함** — `RedisSemanticScriptAclProbe.java:420` (`return {1}`은 명령 0개 → ADVANCED 계정에 `+hset/+pexpire/+time` 없어도 기동 통과, 첫 요청 실패; `redisRequired`는 REGULAR lane만 probe → ADVANCED 자격 증명이 깨져도 readiness UP). Fix: `TIME/HSET/PEXPIRE/DEL`을 실제 수행하는 probe 스크립트; script capability 선택 시 health가 SCRIPT lane probe.
11. **[Consider][architecture] application executor를 adapter leaf에서 조립** — `RedisIdempotencyConfig.java:251-265` (PostgreSQL 쪽 `PostgreSqlIdempotencyProviderConfig.java:36-60`은 processing-lease 2m, codec `json.v1` 하드코딩으로 따로 조립 → `APP_IDEMPOTENCY_PROCESSING_LEASE` 적용 여부가 provider마다 다름). Fix: executor는 bootstrap에서 설정 소스 하나로 한 번만 조립, leaf는 store만 제공.
12. **[Nit][docs]** `app-bootstrap/CLAUDE.md:80,89`가 `autoconfigure/redis/RedisCapabilityConfig`·`RedisKeys`를 가리킴(실제 `cache-redis/.../capability/`); `cache-redis/CLAUDE.md`의 위치·패키지 루트·"Boot 4.0.0" 틀림; `src/README.md:302-306` `APP_CACHE_REDIS_ENABLED`/`CLIENT_MODE`는 `env-keys.yaml:2786-2794`에서 orphaned; cache-redis README의 `:app-bootstrap:test --tests '*RedisSemanticScriptAclProbeTest'`는 잘못된 leaf; `app.redis.capacity.*` env 레지스트리 `owner_module`이 cache-redis.
#### 확인 후 기각한 lead
- NOSCRIPT 재로드 경쟁: NOSCRIPT = 미실행, `SCRIPT LOAD` idempotent, cluster에선 Lettuce advanced `scriptLoad`가 모든 master 전파.
- cluster 해시 태그·구분자 주입: 단일 키 스크립트; `CapabilityKeyspace.requireSegment`가 비-마지막 segment의 `:` 거부; digest 검증.
- KEYS는 policy에서 `BLOCKED`, region 무효화는 generation INCR. M9, M13/L3 closed 확인.
- 비밀 로깅: `RedisCredentials.toString` 마스킹, failure logger는 예외 클래스만. HARD-STOP 1-5 해당 없음.
#### Coverage note
깊게: cache-redis main 전체, runtime owner/topology/auto-config/health/credential/keys, gateway script 메서드, 관련 테스트·fake 일부. 훑음: redis-sdk(policy yml, pub/sub, script 연산, connection registry), Lettuce gateway 데이터 명령.
### Group H — Outbound HTTP client + messaging bridge adapters
#### Per-leaf verdict
| leaf | verdict | one-line reason |
|---|---|---|
| `adapter/outbound/httpclient` | NEEDS_FIX | Dynamic Target sends trace headers to external hosts and skips redirect safety; custom API-key header not stripped on cross-origin redirect; reactive failures reported as NOT_SENT; OAuth2 token fetch blocks inside a map lock |
| `adapter/outbound/messaging` | NEEDS_FIX | Canonical bridge bean never created in the assembled app (auto-config ordering); fixing that breaks startup on the shipped default (blank `producer-id`) |
| `adapter/outbound/messaging-kafka` | PASS | Edges OK; only an interrupt-swallow (Consider) and javadoc drift (Nit) |
HARD-STOP 1–5: none. All `project(...)` edges within `modules.json`.
#### Findings (most severe first)
1. **[Critical][correctness] Canonical bridge `@ConditionalOnBean` evaluated before the platform auto-config; blank `producer-id` breaks startup** — `src/adapter/outbound/messaging/src/main/java/dev/caskeleton/adapter/outbound/messaging/autoconfigure/MessagingBridgeRootAutoConfiguration.java:25,39-44`
- (a) `@AutoConfiguration` has no `after`/`afterName`. The `EncodedMessagePublisher` bean (`DefaultMessagePublisher`) is registered only in `MessagingCoreAutoConfiguration:461`, imported by `dev.caskeleton.messaging.autoconfigure.MessagingPlatformRootAutoConfiguration`. Boot's `AutoConfigurationSorter` sorts alphabetically first → `dev.caskeleton.adapter…` precedes `dev.caskeleton.messaging…` → `@ConditionalOnBean(EncodedMessagePublisher.class)` sees no definition yet. Neither class has `AutoConfigureAfter`/`Order` (grep).
- (b) `config/messaging.yml:15` `producer-id: ${APP_MESSAGING_PRODUCER_ID:}` always exists as an empty string; `@ConditionalOnProperty(name="producer-id")` treats `""` as "not false" and matches → `new ProducerId("")` → IAE at `ProducerId.java:16`.
- No test assembles both: `MessagingPlatformBridgeAutoConfigurationTest` uses `withBean(EncodedMessagePublisher)` (user bean, always first) and "property absent".
- Why it matters: docs (MSG-015) say the canonical bridge exists; in the real assembly `IntegrationEventPublishPort` is never created, and `canonical-transport-enabled=true` is refused by the validator (fail-closed, no data loss). Fixing (a) alone then breaks startup on defaults via (b).
- Fix: `@AutoConfiguration(afterName = "dev.caskeleton.messaging.autoconfigure.MessagingPlatformRootAutoConfiguration")`; producer-id condition → `@ConditionalOnExpression("!'${app.messaging.producer-id:}'.isBlank()")`; one `ApplicationContextRunner` test with starter + bridge (blank / set).
- Needs confirmation: Gradle not run; the assembly test confirms or refutes.
- Status: new
2. **[Required][security] Dynamic Target sends `traceparent` and `X-Correlation-Id` to user-chosen external hosts even with `trace-propagation=false`** — `src/adapter/outbound/httpclient/.../dynamic/DefaultDynamicTargetGateway.java:154`
- Evidence: `prepare()` always calls `PlatformHeaderPropagation.apply(headers)`; no main code reads `DynamicTargetPolicy.tracePropagation` (grep: record, settings, test asserts only). `docs/httpclient/configuration-reference.md:178`: "Off by default for dynamic targets".
- Fix: `if (policy.tracePropagation()) headers = PlatformHeaderPropagation.apply(headers);` + absence test.
- Status: new (H4 fix added propagation unconditionally to the dynamic path)
3. **[Required][security] Dynamic redirect loop bypasses `RedirectEvaluator`/`SensitiveHeaderStripper`** — `DefaultDynamicTargetGateway.java:121-136`
- Evidence: on 303 only `method = GET` changes and `operation.withMethod(...)` keeps the body (`HttpOperation.java:91-101`) — the bug `BlockingRedirectCoordinator.java:97-104` already fixed; no `BODY_NOT_REPLAYABLE` check on 307/308; caller headers (e.g. `X-Api-Key`, allowed by `HeaderPolicy`) forwarded cross-origin; no test calls `DefaultDynamicTargetGateway.exchange` (grep 0).
- Why it matters: `max-redirect-hops` is configurable (`HttpClientPlatformSettings.java:156`); enabling it sends POST payloads to an upstream-chosen host.
- Fix: reuse `RedirectEvaluator.evaluate` + `stripForCrossOrigin` per hop, `EmptyBody` on 303; redirect contract tests (303 body, 307 one-shot, cross-origin header).
- Status: new
4. **[Required][security] Allowed API-key header `X-Client-Key` not stripped on cross-origin redirect** — `.../security/SensitiveHeaderStripper.java:23,47`, `.../restclient/BlockingExecutionSupport.java:90`
- Evidence: `ALLOWED_API_KEY_HEADERS = Set.of("X-Api-Key","Api-Key","X-Client-Key")` (`HttpClientAuthenticationAutoConfiguration.java:34-35`); the stripper in use is `standard()` which strips only `x-api-key`/`api-key`; `withAdditional` has no main caller; empty `allowedHosts` → `TrustedTargetPolicy.java:168` allows every host.
- Why it matters: a profile with `redirect.enabled` + `allow-cross-origin` leaks the secret to the redirect target (both default off).
- Fix: strip the actually resolved `RequestCredentials.headers().keySet()` on cross-origin hops; delete `withAdditional`.
- Status: new
5. **[Required][correctness] Reactive path reports post-send failures as NOT_SENT / `HttpConnectException`** — `.../webclient/ReactiveAttemptExecutor.java:167`, `.../reactor/ReactorFailureClassifier.java:106-108`, `.../restclient/StableBlockingExceptionMapper.java:37`
- Evidence: `classify(failure, baseMetadata.stage())` always gets `VALIDATION` (`provesNotSent=true`); no `AttemptProgressTracker` on the reactive path (grep 0); unknown exceptions → `fallback` `notSent(UNKNOWN)` → mapper `default` → `HttpConnectException`. Examples: `DataBufferLimitException` from `DataBufferUtils.join` over `maxWireBytes` (`WebClientResponseMapper.java:33`; handled nowhere), mid-response `PrematureCloseException`; no reactive too-large / mid-body tests.
- Why it matters: NOT_SENT promises the server never received the request; callers re-send an executed POST → duplicate side effects. Breaks parity with blocking (H9 fixed blocking only).
- Fix: record stage inside `exchangeToMono` (`REQUEST_BODY` after send, `RESPONSE_HEADERS`/`RESPONSE_BODY` on receipt); map `DataBufferLimitException` → `HttpResponseTooLargeException`; mapper `default` → dedicated unknown type.
- Status: new
6. **[Required][performance] OAuth2 token network fetch runs inside `ConcurrentHashMap.compute`** — `.../auth/OAuth2CredentialProvider.java:62-66`, `.../auth/SingleFlightTokenLoader.java:79-83`
- Evidence: `cache.compute(key, … tokenLoader.load(key))` holds the bin monitor up to 10 s; same-key/same-bin callers queue on the monitor instead of joining the single-flight future; after a timeout the next caller starts another 10 s load → Nth caller ≈ N×10 s. `SingleFlightTokenLoaderTest` tests the loader alone. Also: a timed-out caller calls `cancel(true)` on the shared future → late joiners get raw `CancellationException`; `CompletableFuture.cancel` does not interrupt → the 2-thread pool can stay blocked.
- Fix: `get` + expiry check, `tokenLoader.load` outside the lock, then `cache.put` (loader already dedupes); translate `CancellationException` → `HttpAuthenticationException`.
- Status: new
7. **[Consider][structure] `categoryOf` duplicated as `getSimpleName()` string switches, already diverged** — `BlockingAttemptExecutor.java:292-312`, `ReactiveAttemptExecutor.java:183-199` (reactive lacks `HttpCircuitOpenException`/`RateLimitRejected`/`BulkheadRejected` though `ReactiveAttemptResiliencePipeline.java:57-70` throws them → telemetry `UNKNOWN`). Fix: `FailureCategory category()` on each `HttpClientException` subclass; delete both switches.
8. **[Consider][error-handling] Bridge `prepare` turns every `RuntimeException` into `REJECTED_BEFORE_SEND` without a cause** — `.../platformbridge/PlatformIntegrationEventPublishAdapter.java:59-63`. Fix: narrow to `IllegalArgumentException`; carry cause via `OutboxRelayFailureReport` or `OutboxPublishRefusedException(outcome, cause)`.
9. **[Consider][correctness] Kafka seam swallows `InterruptedException`** — `src/adapter/outbound/messaging-kafka/.../KafkaSenderAutoConfiguration.java:110` (`OutboxMessagePublishAdapter.java:94` → `AMBIGUOUS`, `OutboundMessagePublisher.java:36` logs and ignores → relay keeps processing during shutdown, up to 30 s per item). Fix: catch in the lambda, `Thread.currentThread().interrupt()`, rethrow.
10. **[Nit][docs]** `src/README.md:320-345` describes activation via `ca-skeleton.capabilities.http-client` / `OutboundHttpSettings.bindLegacy` — actual `app.httpclient.enabled` (`HttpClientPlatformSettings.PREFIX`, `config/httpclient.yml:12`, `.env:236`); `OutboundHttpSettings` not in main. Kafka javadoc (`KafkaSender.java:7`, `KafkaAdapterConfig.java:14`) says the project must supply `KafkaSender` — the leaf supplies both; `KafkaAdapterSettings.addresses()` unused.
#### Leads checked and dismissed
- README vs CLAUDE activation key: confirmed → #10 (code + CLAUDE correct).
- SSRF first line: userinfo rejected, IDN → punycode before allowlist, suffix at label boundary, all DNS answers validated, Apache socket pinned; DYNAMIC + JDK/Reactor/proxy refused at startup (`ClientProfileValidator:173-207`).
- Engine-level redirects disabled on all 4 transports; Apache auto-retry off; TLS hostname verification `BOTH` with no off switch.
- Retry idempotency: 408/425/429/5xx, write failure, timeout all gated by `safelyIdempotent`; negative `Retry-After` ignored.
- Blocking response size enforced while reading (`CountingBoundedInputStream`); no header/credential logging (0 loggers in main), redacted `toString`.
- C1-style component-scan bypass: exclusion regex in `CaSkeletonApplication.java:106,108` covers both packages.
- `OutboxEnvelopeJson` payload pre-validated by `OutboxPayloadPolicy.requireValidPayload`; `OutboxMessagePublishAdapter.java:78` `join()` bounded by platform `orTimeout`.
- Known-open: L5, L6 appear wired in code (not run-verified); M9 — no spring-kafka in `messaging-kafka` lock.
#### Coverage note
Deep: messaging main autoconfigure/config/outbox/platformbridge/core/realtime; messaging-kafka main; httpclient `dynamic/*`, blocking/reactive executors, redirect, reader/limiter, header, stripper, OAuth2/single-flight, retry engine/backoff, Apache/Jetty factory + cited tests. Skimmed: messaging R1 compiler/envelope/schema; httpclient `service`/`spring7`/`migration`/`observation`/`profile`. #1 and #5 need an assembly test / reactive contract test to confirm.
### Group I — Inbound HTTP (web)
#### Leaf별 판정
| leaf | 판정 | 한 줄 이유 |
|---|---|---|
| `adapter-inbound-web` (`src/adapter/inbound/web`) | **BLOCKED** | 인바운드 어댑터가 저장소 포트 직접 호출·트랜잭션 개시·객체 단위 인가 결정(HARD-STOP 4). 인증 없는 콜백은 본문 크기 제한 없이 수신. 에러 envelope 계약에 빈틈. |
#### 발견 (심각한 순)
1. **[Critical][architecture] durable-operation 조회/취소 유스케이스가 web 어댑터 안에 있고 운영에서 조립도 안 됨** — `src/adapter/inbound/web/.../operationasync/OperationQueryService.java:54-58,80`, `.../operationasync/OperationAccessPolicy.java:28-40`
- 근거: `OperationQueryService`가 `DurableOperationStorePort.find/cancel` 직접 호출, `transactions.inWrite(...)`로 트랜잭션 개시, `OperationAccessPolicy.mayAccess`로 principal/tenant 인가 판단 → AGENTS.md "repository 직접 호출과 비즈니스 정책 금지", CLAUDE.md 표는 authorization을 `application-core`에 둠.
- 조립 불가: `new OperationQueryService(`는 `testFixtures/.../OperationFixtureSupport.java:74`에만(운영 grep 0) → `app.web-platform.durable-operations.enabled=true`면 기동 실패. MVC에서 `WebMvcRequestContextHolder.store` 호출 0 → `WebMvcRequestContextArgumentResolver.java:40` `require`가 항상 예외. `OperationHttpController.java:75,100`의 `WebResourceNotVisibleException`은 opt-in problem 핸들러만 처리 → 기본 envelope 계약에선 catch-all **500**. 실서버 IT는 `OperationFixtureController.java:53`이 합성 context를 넣어 못 잡음(닫힌 H7과 같은 모양).
- 수정: find/cancel + 접근 정책을 `application-core` 유스케이스(`QueryDurableOperationUseCase`/`CancelDurableOperationUseCase`)로; 컨트롤러는 `Authentication` → command 매핑만; `GlobalExceptionHandler`에 `WebResourceNotVisibleException` → 404. 당장 쓸 계획이 없으면 MVC/Reactive 컨트롤러와 `operationasync` 삭제가 가장 작은 변경.
- 가정/반론: 가시성 판단 = 인가 정책 / `docs/web/repository-adaptation.md` §7이 HTTP projection을 web에 둠 — 단 저장소 호출·트랜잭션 조율까지는 아님.
- 상태: new (Group A #4와 같은 결함을 반대편에서 본 것)
2. **[Required][security] 운영 servlet 앱에 요청 본문 상한 없음; 인증 없는 콜백이 본문 전체를 메모리에 읽은 뒤 크기 검사** — `.../notification/platform/callback/NotificationCallbackMvcController.java:85-87`
- 근거: `@RequestBody byte[] body` 후 `if (body.length > maxBodyBytes)` — Spring 컨버터(`readNBytes(contentLength)`/`readAllBytes()`)가 이미 전체 버퍼링. 경로는 `CallbackMvcSecurityConfiguration.java:31-35` `permitAll`. Reactive는 `BoundedCallbackBodyReader.java:30` `DataBufferUtils.join(..., maxBytes)`로 막음 — servlet만 누락.
- 근본 원인: `WebMvcBudgetFilter`는 fixture 앱에서만 생성(main grep 0); `backend.web.budgets.enabled=true`면 `WebMvcBudgetExceptionHandler`가 운영에 없는 `BudgetProblemMapper` 빈을 찾아 기동 실패; 그런데 `docs/web/repository-adaptation.md:158,160`은 본문 크기·JSON 깊이를 "Application, always"가 막는다고 기재. `NotificationCallbackMvcControllerTest.java:117-124`는 배열을 직접 넘겨 컨버터 경로 미검증.
- 영향: `APP_NOTIFICATION_PLATFORM_CALLBACKS_ENABLED=true`(기본 false, `notification.yml:44`)면 누구나 수 GB 본문으로 힙 고갈; 가상 스레드라 스레드 수 제한도 없음.
- 수정: 컨트롤러가 `HttpServletRequest`를 받아 `getContentLengthLong() > maxBodyBytes`면 413, 본문은 `getInputStream().readNBytes(maxBodyBytes + 1)`; 예산 필터를 유지하려면 `WebMvcPlatformAutoConfiguration`에서 등록, 아니면 문서 §8 정정.
- 상태: new
3. **[Required][correctness] Envelope 에러 계약에서 `ProblemDetail`이 새어 나감** — `.../error/GlobalExceptionHandler.java:54,290-293`
- 근거: `ResponseEntityExceptionHandler`가 처리하는 예외 20종 중 8개 hook만 override → 나머지(`MissingServletRequestParameterException`, `MissingRequestHeaderException`, `MissingServletRequestPartException`, `HandlerMethodValidationException`, `ResponseStatusException`, `AsyncRequestTimeoutException`, `TypeMismatchException` 등)는 Spring 7.0.9 `handleExceptionInternal`(630-631)에서 `ProblemDetail` 본문; `EnvelopeBodyAdvice.java:45`는 `application/json`이 아니면 감싸지 않음 → `application/problem+json` 그대로 → README.md:168(D5 "ProblemDetail 거부")과 모순. 실제 경로: sample `PosterImagePublicationController.java:58-61` 필수 헤더 `Idempotency-Key`·필수 파라미터 누락. 이 예외들의 테스트 0건. 추가로 `handleMethodArgumentNotValid`는 `getFieldErrors()`만 → 클래스 레벨 제약 위반은 `details: []`.
- 수정: `handleExceptionInternal` 하나 override(status → `OperationalError`, `ErrorResponseFactory.body` 반환); `getGlobalErrors()`도 details에; 필수 파라미터/헤더 누락 slice 테스트.
- 상태: new (C2 종결 범위는 opt-in 핸들러까지)
4. **[Required][correctness] `RequestLoggingFilter`가 Spring Security 뒤에서 돌고 trace id를 따로 생성** — `.../filter/RequestLoggingFilter.java:37,65-71,69`
- 근거: `@Component`에 `@Order` 없음 → LOWEST_PRECEDENCE, Security 체인(-100) 뒤. MDC가 이 필터 안에서만 채워짐 → `AuthErrorResponseWriter.java:46-70`의 401/403 응답 `meta` id 전부 null, 경고 로그에 request_id 없음, 해당 요청은 `http_request` 접근 로그도 없음. `:69`는 traceparent 없으면 새 값 생성 → `WebMvcRequestIdFilter.java:78-80`의 `WebTraceId`와 불일치(H1과 같은 모양, trace id 쪽).
- 수정: MDC 채우기를 security보다 앞의 `WebMvcRequestIdFilter`로 이동; 로깅 필터는 `WebTraceId`도 읽기; `OperationalContractRuntimeTest`에 "request-id 필터가 `springSecurityFilterChain`보다 앞" 단언.
- 상태: new (H1/H7 형제)
5. **[Required][security] `FORWARDED_HEADERS_TRUSTED` 레이트리밋 키가 XFF 맨 왼쪽 값 사용, peer 검증 없음; forwarded 헤더 신뢰 판단이 네 벌** — `.../ratelimit/ForwardedHeaderClientIpResolver.java:15`
- 근거: `split(",", 2)[0]`, `getRemoteAddr()` 신뢰 프록시 확인 없음. 저장소 자체 `infra/fileserver/nginx/nginx.conf:44`가 `$proxy_add_x_forwarded_for`(뒤에 덧붙임) → 맨 왼쪽은 클라이언트가 정함 → 요청마다 새 버킷, 비인증 레이트리밋 우회. `RateLimitKeyResolverTest.java:79-86`이 이 동작을 정답으로 고정. 판단이 네 곳: `proxy/TrustedProxyPolicy`(CIDR, 운영 호출 0), `ExternalRequestUrlResolver`(정확 문자열 집합), 이 resolver(검증 없음), reactive 콜백 `NotificationCallbackWebFluxHandler.java:105`(`request.uri()` 그대로 → servlet과 서명 대상 URL 불일치).
- 수정: `TrustedProxyPolicy` 하나로 통일(XFF 오른쪽부터 첫 비신뢰 주소; peer 비신뢰면 remoteAddr). 더 간단: 이 모드 삭제 + `server.forward-headers-strategy=native`(Tomcat RemoteIpValve `internal-proxies`) + `REMOTE_ADDR_ONLY`.
- 가정/반론: 뒤에 덧붙이는 LB(ALB, nginx 기본) 뒤에서 켬 / 기본값 안전, 문서에 "덮어쓰는 ingress에서만".
- 상태: new (H3 종결은 서버 기본값만)
6. **[Required][architecture] 알림 제출 정책 기본값을 컨트롤러가 결정(HARD-STOP 4 후보)** — `.../notification/platform/submission/NotificationSubmissionController.java:48-51,81-89,120-123`
- 근거: category 생략 → `"transactional"`, 템플릿 버전 `1`, locale `Locale.getDefault()`(호스트 JVM). category는 suppression 범위(`SuppressionEntry.java:53-54`, `CompositeNotificationEligibilityPolicy.java:40`)와 중복 제거 식별자에 들어감 → 생략 시 marketing suppression 미적용 = transport 계층의 비즈니스 결정. `NotificationTemplateController.java:74` `1L`도 같음.
- 수정: `AcceptNotificationCommand`에서 세 값을 `Optional`로, 기본값은 유스케이스가 설정값으로 적용.
- 가정/반론: category가 동의/suppression 기준(코드상 그러함) / 호출자가 명시적으로 "transactional"을 넘길 수 있음. 소유자가 정책으로 보면 Critical.
- 상태: new
7. **[Required][docs/contract] `wire-type-manifest.yaml`이 런타임에 없는 wire 계약을 공개** — `.../main/resources/META-INF/web/wire-type-manifest.yaml:25-29` (`LONG: INTEGER_STRING`, `BIG_DECIMAL: DECIMAL_STRING`, ":4 client author가 읽는 계약"; 실제로는 숫자 — `JacksonSerializationPolicyTest.java:83-93`이 BigDecimal plain 숫자 단언, sample `StoredObjectResponse.java:4` `long size` 숫자; `WebWireTypeManifest` 빈은 생성만(`WebMvcPlatformAutoConfiguration.java:123-127`); 테스트는 상수=상수). 수정: yaml·빈·클래스 삭제(최소), 유지하려면 Jackson 모듈로 적용 + 실제 JsonMapper 테스트. 상태: new (lead 확인)
8. **[Consider][structure] 운영 소비자 없는 라이브러리 코드가 많고 운영 경로와 중복** — main 67개 파일(약 5.7k LOC) 운영 참조 0(클래스명별 grep); 외부 importer 0 패키지: `security/`, `cache/`, `openapi/`, `idempotency/`; `security/WebCorsPolicyValidator`가 실제 `CorsSettings`보다 엄격한 규칙인데 미호출; `observability/WebAuditPublisher` 참조 0; `RateLimitInterceptor.disabled(...)` M2 이후 dead; `cursor/CursorCodec.withDevKey()` 키 하드코딩. 수정: 쓸 것은 조립, 나머지 삭제; CORS 규칙은 `CorsSettings`로 합치고 validator 삭제. 반론: M1 종결 때 LIBRARY_ONLY 수용. 일부 known-open L4/L5/L6.
9. **[Consider][docs] 같은 코드 이름이 스위치에 따라 다른 status** — `problem-catalog.yaml:19-22` `VALIDATION_FAILED` 422 vs `OperationalError.java:15` 400; README D5에 opt-in problem 경로(`spring.mvc.problemdetails.enabled`) 미기재; reactive 핸들러도 `spring.mvc.*`로 켜짐(`WebFluxProblemExceptionHandler.java:44-45`); reactive 스택엔 envelope 핸들러가 없어 스위치 off면 Boot 기본 에러 JSON. 수정: README에 opt-in 경로 기재, 코드/status 정합, reactive는 `spring.webflux.problemdetails.enabled`.
#### 확인 후 기각한 리드
- problem-catalog 422 vs README 400: 의도된 opt-in 경로(C2 종결), 기본 계약은 400 → 남는 문서 문제는 #9. `application-web-contract.yaml` `problemdetails: true`: testFixtures 전용.
- CORS wildcard + credentials: `CorsSettings.java:32-37`에서 거부, `setAllowedOrigins` 정확 일치.
- HARD-STOP 2/3: web main에 JPA/Spring Data/outbound import 0; `ProviderCallbackHttpRequest`는 bootstrap bridge에서 매핑. 문제는 #1뿐.
- 의존 edge: `application-core`, `application-notification`, `shared-contract` 모두 허용; `web-fileserver`는 test 범위.
- 인증 쪽 필터 순서 정상(request-id HIGHEST+10 < security; 콜백 체인 `securityMatcher` 분리; `RestrictedPathRule`이 `anyRequest`보다 먼저).
- FYI: 레이트리밋이 HandlerInterceptor라 401 폭주는 제한 안 됨(README에 적힌 설계).
- C3(known-open) 이제 닫힘: systemTest/performanceTest/testFixtures 리소스 10개 추적, untracked 없음.
- JWT lazy decoder: Security 7.0.7 `SupplierJwtDecoder`가 `SingletonSupplier` → 가상 스레드 pinning 없음.
#### 검토 범위
자세히: `auth/`, `settings/`, `notification/platform/**`, `ratelimit/`, `error/GlobalExceptionHandler`, `envelope/`, `filter/`, `mvc/{autoconfigure,filter,operation,context,budget}`, `operationasync/`, `proxy/`, `webflux/autoconfigure` + 관련 테스트·fixture; Spring 7.0.9 / Security 7.0.7 소스로 동작 확인. 대충: `advanced/`, `pagination`, `conditional`, `openapi`, `moduleboundary`.
### Group J — Inbound GraphQL
#### Per-leaf verdict
| leaf | verdict | one-line reason |
|---|---|---|
| `adapter-inbound-graphql` (`src/adapter/inbound/graphql`) | **NEEDS_FIX** | No HARD-STOP. One Critical DoS: the wired authorization stage expands fragments exponentially before any budget check. 5 Required: `__typename` rejected, errors masked and never logged, two security flags look active but do nothing. |
#### Findings (most severe first)
1. **[Critical][security/perf] Authorization stage walks fragments exponentially with no budget, before the cost stage** — `src/adapter/inbound/graphql/.../runtime/GraphQlDocumentAuthorizationHandler.java:63,109-137`, `.../cost/GraphQlDocumentShapeAnalyzer.java:190-216`
- Evidence: chain order `stable(selection, authorization, cost)` (`GraphQlPlatformAutoConfiguration.java:403-408`). Authorization calls `verifyIntrospection` → `selectsIntrospection` (analyzer 209-214) and `collect` (handler 130-135); both only guard cycles by push/pop on `expansionPath`, no visit budget. Only `walk` in the cost stage has the 200k budget. Attack document `query{...F0} fragment F0 on Query{...F1 ...F1} … fragment F39 on Query{_health}` ≈ 1.6KB (under the 16KiB cap) = 2^40 visits. `maxFragments=50` checked later in the cost stage. graphql-java 25 dedupes fragments in validation/execution (`OverlappingFieldsCanBeMerged:101-111`, `FieldCollector:85-95`, `NoFragmentCycles` memo) → the document validates fast; the blowup exists only in platform code.
- Why it matters: one request pins a request thread indefinitely (deadline never checked inside the walk). Anonymous in LOCAL/DEV; any authenticated principal in prod.
- Fix: in both walkers replace the path Deque with a document-wide never-popped `Set<String> visited` (results are set membership → same answers, linear walk), or run budgeted `analyze()` before authorization; systemTest with a 40× doubling fragment.
- Status: new (side effect of the GQL-007 fix)
2. **[Required][correctness] `__typename` treated as introspection** — `.../cost/GraphQlDocumentShapeAnalyzer.java:40,199`
- Evidence: `startsWith("__")`; shipped default `introspection-enabled: false` (`app-bootstrap/.../config/graphql.yml:18`), forced off in prod (`GraphQlPlatformStartupValidator.java:43-45`); no `__typename` test.
- Why it matters: Apollo `addTypename` and urql's cache add `__typename` everywhere → all their queries rejected (`GRAPHQL_DOCUMENT_SHAPE_REJECTED/INTROSPECTION`). Spring/graphql-java introspection-off blocks only `__schema`/`__type`.
- Fix: block only `__schema`/`__type`; systemTest `{ _health __typename }` succeeds with introspection off.
- Status: new
3. **[Required][correctness/ops] Every failure masked, none logged**
- Evidence: `GraphQlExceptionResolver.java:55` always returns the masked error → Spring treats it as resolved and logs only at DEBUG (`spring-graphql 2.0.5 ExceptionResolversExceptionHandler:99-101`); `GraphQlPlatformWebInterceptor.java:94`, `GraphQlPlatformInstrumentation.java:84` convert unknown RuntimeExceptions without logging; `grep -rl 'LoggerFactory|getLogger|System.Logger'` over main → 0 files, yet Javadoc says "the detail stays in the logs" (`GraphQlWireErrorMapper.java:30`, `GraphQlPlatformRejectionMapper.java:28`).
- Why it matters: a prod NPE/DB error leaves no trace; the executionId the client gets matches nothing.
- Fix: log once at ERROR with executionId + exception on the masking branch.
- Status: new
4. **[Required][correctness] Wired DataLoader path and blocking bridge errors reach the client as `INTERNAL_ERROR`** — `.../runtime/GraphQlPlatformRejectionMapper.java:114-121`
- Evidence: `GraphQlBatchTimeoutException`, `GraphQlMissingKeyException`, `GraphQlBlockingBridgeFullException` declare `CODE`s unknown to mapper and catalogue; `GraphQlErrorCodeCatalogTest.java:54-60` calls batch timeout "modelled-only" while `CLAUDE.md:171` marks DataLoader `wired`; bridge-full goes through Reactor `ExecutorScheduler.schedule` → `Schedulers.handleError` (ERROR log / uncaught handler on every rejection) → arrives wrapped in `RejectedExecutionException`; tests only assert `getErrors().isNotEmpty()` (`GraphQlBatchLoaderRegistrationTest.java:150`).
- Fix: `GraphQlBatchTimeoutException extends GraphQlRequestCancelledException`; unwrap bridge-full → `DEPENDENCY_UNAVAILABLE`; assert wire codes.
- Status: new (gap in M1 closure)
5. **[Required][security] Stable extensions allowlist accepts `idempotencyKey`, which nothing honours** — `.../http/GraphQlExtensionsPolicy.java:44` (only consumer `mutation/GraphQlMutationIdempotencyInterceptor.java:48` is not on the request path). Why: retrying clients believe retries are deduplicated; the mutation runs twice. Fix: `stableDefaults()` → `none()` until wired. Status: new (side effect of M6 closure)
6. **[Required][security] `GraphQlClientPolicy.persistedOperationOnly` does nothing** — `.../policy/GraphQlClientPolicy.java:50` (`grep -rn 'persistedOperationOnly()'` → only `GraphQlClientPolicyTest:75` getter assert). Why: adopter sets true, arbitrary documents still execute (fail-open allow-list bypass). Fix: remove the field or reject `true` in the compact constructor until wired. Status: new
7. **[Consider][perf] Preparsed cache scans every entry under a global lock per lookup** — `.../execution/BoundedPreparsedDocumentProvider.java:142,167-178` (`expireIdle` iterates the whole map, default 1000). Fix: access-ordered `LinkedHashMap` → stop at first unexpired entry. Counter: tens of µs at 1000 entries.
8. **[Consider][security] Root type names hard-coded for authorization** — `.../runtime/GraphQlDocumentAuthorizationHandler.java:80-83` (literals "Query/Mutation/Subscription" vs cost stage `schema.getQueryType()` at `GraphQlDocumentComplexityScorer.java:258-260`; renamed roots → rules never match, `denyByDefault(false)` allow rules fail open). Fix: names from `context.schema()`.
9. **[Consider][correctness] Streaming path of the body-cap filter untested** — `.../runtime/servlet/GraphQlRequestBodyLimitFilter.java:85` (systemTest `GraphQlPlatformExecutionPathTest:224-241` covers only Content-Length fast path; `FrameworkServlet` likely wraps unresolved RuntimeException in `ServletException` → `catch (BodyTooLargeException)` misses → 500 instead of 413; unverified; cap itself holds). Fix: chunked-transfer test, catch via cause chain.
10. **[Consider][structure] ~57% of the code unreachable** (known-open GQL-025) — static reference closure from the four Spring entry points reaches **136 of 400 files / 11.2k of 25.9k LOC**. Candidates: `advanced/**` 134 files / 5,986 LOC (nothing outside references it, nothing binds `GraphQlAdvancedFeatureFlags`); packages `compat` 1,562, `release` 540, `fetch` 434, `mutation` 683 LOC; unreachable classes in wired packages — parallel HTTP layer in `http/` (11 classes: `GraphQlRequestEnvelopeValidator`, `GraphQlHttpResponseFactory`, `StatusMapper`, `MediaTypes`, …), `error/GraphQlRequestErrorMapper` (second error vocabulary), `autoconfigure/GraphQlPlatformActuatorEndpoint` + `…ConfigurationReport` (not beans), 16 of 19 `schema/` files; `moduleboundary/` (535 LOC) test-only data in the production jar. Fix: delete or move to a non-shipped artifact (keep `pagination` if wanted). Counter: earlier review judged them tested parts awaiting wiring.
11. **[Consider][docs]** `CLAUDE.md:168` marks Accept negotiation `wired` (`GraphQlMediaTypes` unreachable, unit test only); `CLAUDE.md:59,132-133`, `README.md:46-61` describe non-existent `GraphqlExceptionResolver` and "return null → Spring default" (reality: everything masked). Fix: grade `modelled`, document current resolver + masking.
12. **[Consider][ops] Production boots with the anonymous principal resolver, then refuses every request** — `.../autoconfigure/GraphQlPlatformAutoConfiguration.java:319` (prod requires a `GraphQlAuthorizationPolicy` bean but not a `GraphQlPrincipalResolver` → startup OK, 100% `AUTHENTICATION_REQUIRED`). Fix: startup validator fails in prod without a principal resolver bean.
#### Leads checked and dismissed
- HARD-STOP 1-5: edges `shared-contract` + `application-core` only; no persistence/Spring Data imports; no GraphQL types in application-core/domain-core/shared-contract.
- No internal detail leaks (allowlist + masking). `GraphQlBlockingBridge` pool and queue finite, opt-in (Nit: thread factory always names `-0`, `GraphQlBlockingBridge.java:56`).
- Subscription lease/backpressure and `advanced/persisted` bypass: modelled only, not on request path; lease release idempotent (CAS).
- Preparsed cache key uses server-side client profile + memoized schema hash → no cross-profile leak. BigDecimal expansion bounded (`GraphQlDecimalBounds`); `GraphQlRequestSize` recursion bounded by Jackson depth.
- H1-H3, M1-M7: no regressions.
#### Coverage note
Read in full: wired path (`GraphQlPlatformAutoConfiguration`, all `runtime/*` incl. servlet filter), `cost/GraphQlDocumentShapeAnalyzer`, DataLoader executor/factory, `error/*`, client/extensions/authorization policies, preparsed cache, related tests/systemTests; checked graphql-java 25, spring-graphql 2.0.5, reactor 3.8.7 sources. Skimmed: `advanced/*` (except persisted + subscription lease), `compat`, `release`, `fetch`, `mutation`, `pagination`, `schema`, `observation`, `scalar`. Reachability computed with an inline script.
### Group K — Inbound realtime (WebSocket) + gRPC adapter
#### Per-leaf verdict
| leaf | verdict | one-line reason |
|---|---|---|
| `adapter-inbound-websocket` | NEEDS_FIX | No HARD-STOP; the live STOMP channel is sound. The platform library breaks advertised guarantees (authentication, size limits, backpressure, shutdown ordering, concurrency) — none has a production caller yet. |
| `adapter-inbound-grpc` | NEEDS_FIX | Deny-by-default config and error cleanup solid, but the exception interceptor drops `onComplete` and late `onCancel`. |
#### Findings (most severe first)
1. **[Required][security] `HandshakeAdmissionPipeline` admits cookie/bearer handshakes on header presence alone** — `src/adapter/inbound/websocket/.../handshake/HandshakeAdmissionPipeline.java:131-136,140-144`
- Evidence: `request.bearerToken().isPresent() ? Optional.of(anonymousReference())` (cookie branch the same); every caller gets the fixed actor `"0".repeat(64)`; comment says "the runtime replaces" it — nothing does. `grep -rn 'HandshakeAdmissionPipeline\|\.admit(' main` → no callers outside the class; no `BEARER_HEADER` test.
- Why it matters: `HandshakeDecision.actor` documented as "who is connecting" — any `Authorization: Bearer x` is admitted; all users collapse into one actor (`maxConnectionsPerActor=4` becomes node-wide 4).
- Fix: required `WebSocketCredentialVerifier` constructor parameter for COOKIE/BEARER; refuse 401 when absent.
- Assumption / Counterargument: adopters call `admit()` without separate verification / Javadoc says verification is the security integration's job — but the return type means "verified".
- Status: new
2. **[Required][correctness/security] Message/fragment limits never enforced on the servlet runtime** — `.../servlet/PlatformWebSocketHandler.java:37,84`, `.../inbound/FragmentAssembler.java:75`
- Rule → Evidence: no `supportsPartialMessages()` override → Spring default `false` → `MessageHandler.Whole`, `isLast()` always true → `FragmentAssembler` returns early, never enforces `maxMessageBytes`/`maxFragments`; effective limit is Tomcat's 8 KiB default (no `maxTextMessageBufferSize`, grep 0). `WebSocketRuntimeContract.oversizedMessageClosesWith1009` passes only via the frame check. `:75` compares UTF-16 char counts against a byte budget.
- Fix: `supportsPartialMessages() { return true; }`; composition sets container buffer to `maxFrameBytes`; running UTF-8 byte count; fragmented-message runtime test.
- Status: new
3. **[Required][performance/security] `ReactiveFrameSink` buffer is unbounded, contradicting its own comment** — `.../webflux/ReactiveFrameSink.java:41-43`
- Rule → Evidence: `Sinks.many().unicast().onBackpressureBuffer()` → reactor-core 3.8.7 `Queues.unbounded()`; `tryEmitNext` never fails, `write()` returns immediately, `releaseInFlight` returns both reservations → per-connection and node budgets do nothing on the reactive stack. No callers, no tests.
- Fix: bounded `onBackpressureBuffer(Queues.<String>get(n).get())` + slow-consumer test, or delete until the reactive stack is assembled.
- Status: new
4. **[Required][correctness] `SerializedOutboundWriter.drain()` can strand a message** — `.../outbound/SerializedOutboundWriter.java:110-127,156-158`
- Race: A polls empty, releases `queueLock`, still holds `writeLock`; B enqueues, `tryLock()` fails; A unlocks → B's message waits for an unrelated future offer. Javadoc `:106-108` ("nothing is stranded") false; `SerializedOutboundWriterTest.java:110` hides it by calling `drain()` manually.
- Fix: after `writeLock.unlock()` recheck `queue.size()>0` under `queueLock` and loop (or WIP counter); remove the manual drain from the test.
- Status: new
5. **[Required][correctness] `PlatformWebSocketHandler.stop()` runs after the web server stops** — `.../servlet/PlatformWebSocketHandler.java:162-167`
- Rule → Evidence: `getPhase()` = `Integer.MIN_VALUE + 1` ("Stopped before the web server"); Spring stops higher phases first; Boot 4.0.8 `GRACEFUL_SHUTDOWN_PHASE = MAX-1024`, `START_STOP_LIFECYCLE_PHASE = MAX-2048` → this bean stops last, promised 1001 never sent. `TomcatWebSocketAbuseIT.java:71` calls `stop()` directly.
- Fix: phase above `GRACEFUL_SHUTDOWN_PHASE` (e.g. `SmartLifecycle.DEFAULT_PHASE`); IT closes the context.
- Status: new
6. **[Required][correctness] gRPC exception interceptor drops `onComplete` and late `onCancel`** — `src/adapter/inbound/grpc/.../GrpcExceptionHandlingInterceptor.java:72-86,95-97`
- Rule → Evidence: `close()` sets `closed=true` on every close incl. OK; `runGuarded` returns early when closed → transport's `onComplete`/`onCancel` swallowed → grpc-stub 1.68.1 `ServerCalls` never runs `onCloseHandler`/`onCancelHandler` (`:193-215,277-307`) → streaming cleanup/metrics hooks broken. No test.
- Fix: always pass `onComplete`/`onCancel` through (try/catch only); keep the gate for `onMessage`/`onHalfClose`/`onReady`; wire test asserting `setOnCloseHandler` runs.
- Status: new
7. **[Required][correctness] `WebSocketSessionRegistry` register/deregister race breaks the per-actor cap** — `.../session/WebSocketSessionRegistry.java:63-64,94-100`
- Race: register gets map M via `computeIfAbsent`; concurrent deregister empties M and `byActor.remove(actor, M)`; register inserts into detached M → `connectionsOf` misses it, next register creates M2 → cap exceeded. `countByEndpoint` (`:75`) same race. Test only runs concurrent registers.
- Fix: register via `byActor.compute(...)` (atomic check+insert); deregister via `computeIfPresent(... isEmpty ? null : held)`; `compute`/`merge` for endpoint counter.
- Status: new
8. **[Consider][correctness] Conflation replacement skips the per-connection bound; loss reported as success** — `.../outbound/OutboundQueue.java:78-79,193-210` (`replace` only checks `globalBudget.reserve`; on failure both values dropped but `offer` returns `CONFLATED`). Fix: same local check as `offer`, return `DROPPED`. Assumption/Counter: conflated values can grow / typical conflatable payloads fixed-size.
9. **[Consider][performance] `WebSocketCorrelationRegistry.countFor` scans all node in-flight entries per request** — `.../handler/WebSocketCorrelationRegistry.java:90,116-122` (O(total in-flight), check+insert not atomic). Fix: per-connection nested map and `size()`.
10. **[Consider][security] gRPC authentication policy is yes/no and swallows exceptions silently** — `.../GrpcAuthenticationInterceptor.java:35-41` (`boolean isAuthenticated(Metadata)`, no `Context` identity, `catch (RuntimeException ignored)`). Fix: `Optional<Principal> authenticate(Metadata)` + `Context.Key`; WARN log without credentials. Counter: P1 is loopback-only.
11. **[Consider][performance] Live STOMP channel uses Spring's unbounded defaults** — `.../stomp/WebSocketConfig.java:57-66`, `.../stomp/LiveEventStompBroadcaster.java:174-208` (inbound `ThreadPoolTaskExecutor` `queueCapacity` = `Integer.MAX_VALUE`; synchronous `@EventListener` O(subscribers) fan-out on the publishing thread). Fix: set `queueCapacity` for inbound/outbound channels. Counter: docs label R1 best-effort.
12. **[Nit][docs]** `moduleboundary/WebSocketStableModule.java:89-90` duplicated `HANDSHAKE` Javadoc (known-open L3); both leaves' `CLAUDE.md` say Boot 4.0.0 (catalog: 4.0.8).
#### Leads checked and dismissed
- HARD-STOP 1-5: websocket imports only `application.realtime.*` ports and `@DomainEvent`; gRPC only `shared.error.*`; edges match modules.json.
- gRPC interceptor order: authentication outermost — correct. TLS/deadlines/external bind: documented P2; loopback enforced by `@AssertTrue`.
- STOMP authz: SUBSCRIBE exact match, SEND limited to `/app/`, `/app/../topic` does not reach broker; origin config rejects wildcards.
- `ResumeTokenCodec` constant-time compare + expiry after signature. Rollback phantom events documented P2. M4 no regression. `LateResponseTombstone`/`FanoutDeduplicator` have eviction.
#### Coverage note
Read in full: entire gRPC main tree, STOMP scope, advanced STOMP/SockJS/Rabbit configs, safety configs, outbound, servlet, handshake, session, inbound, webflux sink + related tests; checked grpc-java/Spring/Reactor sources. Skimmed: codec/cbor/protobuf, evidence, release, presence, cluster, graphql-bridge.
### Group L — Composition root + sample consumer
Path shorthand: `B/` = `src/app-bootstrap/src/main/java/dev/caskeleton/bootstrap/`, `BR/` = `src/app-bootstrap/src/main/resources/`, `S/` = `src/sample-portfolio/`.
#### Per-leaf verdict
| leaf | verdict | one-line reason |
|---|---|---|
| app-bootstrap | NEEDS_FIX | Outbox relay only scheduled when idempotency=jdbc (Critical); the prod block on error-detail exposure exists in docs, not code. |
| sample-portfolio | NEEDS_FIX | Actuator loggers accept unauthenticated writes; sample dropped out of the central ArchUnit suite and its replacement has only 3 rules. |
#### Findings (most severe first)
1. **[Critical][correctness] Outbox relay never runs unless idempotency provider = jdbc** — `B/autoconfigure/outbox/OutboxRelayScheduler.java:36`, `B/autoconfigure/idempotency/IdempotencyConfig.java:17-19`
- Evidence: grep of all `src` for `EnableScheduling|ScheduledAnnotationBeanPostProcessor|SchedulingConfigurer` → in production code only the sample, `composition/fileserver-platform` (not shipped) and `IdempotencyConfig`, which is `@ConditionalOnProperty(...idempotency.provider, havingValue="jdbc")` with default `disabled` (`BR/config/capabilities.yml:30`). The relay is `@Scheduled` (`OutboxRelayScheduler.java:36`); `OutboxReaper` (persistence-jpa `OutboxReaper.java:39`) too. `BootstrapCapabilityReachabilitySystemTest` enables outbox only with `relay-enabled=false` and checks the scheduling processor only in the jdbc-idempotency case.
- Why it matters: outbox + relay-enabled + a broker (the documented fork path) with idempotency disabled/redis/postgresql → relay bean created, never runs; events pile up PENDING, nothing logged.
- Fix: `@EnableScheduling` on `OutboxConfig` (already gated by `ca-skeleton.outbox.enabled`); one `ApplicationContextRunner` test: outbox+relay on, idempotency=disabled → `SCHEDULED_ANNOTATION_PROCESSOR_BEAN_NAME` exists.
- Status: new
2. **[Required][security] Prod profile claims error details are blocked; no code checks it** — `BR/application-prod.yml:13`
- Evidence: comment "`StartupSafetyValidator` error-detail exposure and body-capture logging stay off"; `B/runtime/StartupSafetyValidator.java:30-32` only calls `validateMultiInstance()` (`validateProdSafety` deleted in 9da00eb7; the deleted `APP_ERROR_DETAIL_EXPOSURE_ENABLED` toggle was a no-op anyway). `BR/application.yml:26,28` lets `spring.web.error.include-stacktrace/include-message` be env-overridden in prod; grep `include-stacktrace` → only yml/.env/README; no custom `ErrorController`/`ErrorAttributes` → Boot `BasicErrorController` owns `/error`. Same false claim in `src/app-bootstrap/README.md:280-281` (contradicts `:785-788`), `src/README.md:219,223` ("5종"), `BR/config/runtime.yml:10-12`.
- Why it matters: `APP_SERVER_ERROR_INCLUDE_STACKTRACE=always` in prod boots fine while docs say it is blocked.
- Fix: one prod validator shaped like `JpaSchemaSafetyValidator` (exit 71 when `include-stacktrace`/`include-message` ≠ `never` or `include-exception=true`), registered in `RuntimeSafetyConfig`; fix the 4 doc locations. Unchecked: whether `/error` body is envelope-wrapped.
- Status: new
3. **[Required][security] Anyone can change the sample app's log levels without auth** — `S/src/main/resources/application.yml:94`
- Evidence: exposure `include: ...,loggers`, `management.endpoint.loggers.access` unset (unrestricted); `S/.../bootstrap/security/SamplePublicAccessSecurityConfig.java:24-27` `securityMatcher("/**")` + `csrf.disable()` + `permitAll()` at `@Order(-100)`; `SampleManagementSecurityConfig.java:24-27` also permitAll; `src/Dockerfile.sample:131` `EXPOSE 8080 9001`.
- Why it matters: unauthenticated `POST /actuator/loggers/{name}` → TRACE (data leak) or OFF. Production denies this in `ManagementSecurityConfig`.
- Fix: `management.endpoint.loggers.access: read-only`.
- Status: new
4. **[Required][architecture/tests] Sample left the central ArchUnit suite; replacement does not cover HARD-STOP 2/3** — `S/src/test/java/dev/caskeleton/sample/portfolio/architecture/SampleArchitectureTest.java:25-69`
- Evidence: ace8aaae deleted `sampleFixture project(':sample-portfolio')` (was `build.gradle:179` at 944a1e34); current `src/app-bootstrap/build.gradle` has 0 sample references → `DOMAIN_IS_PURE`, `WEB_DTOS_STAY_IN_WEB_ADAPTER`, `CONTROLLERS_DO_NOT_ACCESS_REPOSITORIES_OR_JPA_ENTITIES`, `APPLICATION_METHODS_DO_NOT_ACCEPT_WEB_DTOS` no longer see the sample. Replacement has 3 rules (framework bans for domain/application; web → sample persistence). Missing: domain → application/adapter/bootstrap; application → `..adapter.inbound.web..dto..`; controller → `org.springframework.data..`/`jakarta.persistence..`. Current code is clean (0 such imports). `S/CLAUDE.md:38-41` cites a non-existent rule `production_code_does_not_depend_on_sample_portfolio`. Also to check: `adapter:inbound:grpc` not on the architectureTest classpath and has 0 ArchUnit tests.
- Fix: add the 3 `noClasses()` rules or `architectureTestRuntimeOnly project(':sample-portfolio')`; fix CLAUDE.md.
- Assumption / Counterargument: sample is the reference adopters copy (README/CLAUDE say so) / disposable fixture.
- Status: new (regression from ace8aaae)
5. **[Consider][structure] The remaining `StartupSafetyValidator` check can never fail** — `B/runtime/StartupSafetyValidator.java:42-61` (`B/autoconfigure/lock/DistributedLockConfig.java:23-29` creates `distributedLockProvider` under the same condition; missing dependency fails bean creation first; `CapabilityDependencyValidator.java:100-105` already requires JPA). Fix: delete it and the `RuntimeSafetyConfig.java:16-20` bean; #2's validator replaces it.
6. **[Consider][structure] Relational-dependency list lives in two places and has drifted** — `B/autoconfigure/jpa/DataSourceRequirement.java:28-60` vs `B/activation/CapabilityDependencyValidator.java:57-116` (Fileserver entry `:52-58` missing from the validator → fileserver on + JPA off likely fails with a missing-bean error not naming the switch, unverified; `B/runtime/SecretSourceValidator.java:39-40,93-101` always requires `APP_DATASOURCE_PASSWORD` even for a no-DB prod deployment). Fix: when `!jpa`, reject each `DataSourceRequirement.reasons(env)` (removes 4 ifs); require DB password only when `isRequired(env)`.
7. **[Consider][security/tests] Actuator endpoints that require login cannot be reached** — `B/management/security/ManagementSecurityConfig.java:25-54` (no `oauth2ResourceServer`/`httpBasic` on the actuator chain → `loggers` GET and `show-details: when-authorized` (`BR/config/observability.yml:45`) unreachable with any credential; `ActuatorSecurityHttpTest.java:95-96` injects `.with(user("ops"))` and asserts 200). Fix: drop `loggers` + `show-details: never`, or add `.oauth2ResourceServer(jwt())` and test with a real bearer token. Counter: fails closed.
8. **[Consider][sample] Sample defaults to schema settings prod forbids** — `S/src/main/resources/application.yml:45,53` (`out-of-order: true`, `ddl-auto: ...:update` on a Flyway-owned schema; `BR/application-local.yml:79-83` records this combination hid 2 mapping/migration mismatches). Fix: `validate`, `out-of-order: false`.
9. **[Nit][structure]** `B/autoconfigure/notification/NotificationApplicationCompositionAutoConfiguration.java` has 12 `@Bean`s (leaf CLAUDE.md: split at ~8), orphaned Javadoc at `:80-90`, many inline FQNs → split into submission/admin/dispatch/callback.
10. **[Nit][docs]** `BR/config/outbox.yml:7-8` says "Bound into OutboxProperties" and "NO env placeholders" (actual `OutboxSettings`, placeholders at `:18,19,22`); `BR/config/idempotency.yml:8` names non-existent `IdempotencyProperties` (→ `IdempotencySettings`); `src/app-bootstrap/README.md:7` "19개 leaf" (registry: 73; stated counts are defects); `src/app-bootstrap/CLAUDE.md:80,89` point to deleted `RedisCapabilityConfig`/`RedisKeys` in bootstrap; `S/README.md:418-420,434-437` describe the deleted `TestEnclosedConfigurationFilter` and a two-location customizer; `src/app-bootstrap/build.gradle:90-92` orphaned comment.
11. **[Nit][tests]** `src/app-bootstrap/src/test/resources/application-test.yml:113-117` sets `ca-skeleton.rate-limit.*` but the binder is `app.rate-limit` (`EdgeRateLimitTransportSettings.java:14`) — same in sample `application.yml:187-192`; `:112`/`:124` pair `api-base-path: ""` with `/api/healthcheck` (harmless, `addFilters=false`); `developer-experience/broken-readme.md` referenced by nothing (grep `broken-readme|definitelyMissingDxTask` → only itself) — delete.
#### Leads checked and dismissed
- All seven graph-pass leads confirmed; doc/test ones are #10/#11, the `StartupSafetyValidator` false assurance is #2.
- No sample leak into production (no `allowed_dependencies` or production build.gradle reference).
- `ProviderCallbackIngressBridge.java:43-88` deliberately closed as Core C2. 0-B `PersistenceJpaRootAutoConfiguration` internal-name reference resolved.
- Deleted Outbox/DistributedLock/IdempotencyUniqueScope contract tests moved to persistence-jpa `postgresqlIntegrationTest`.
- Sample systemTest idempotency regression looks fixed (provider=jdbc in `S/src/main/resources/application.yml:176-177` and systemTest `application-test.yml:105-106`; not run).
- H5 commit inclusion can be closed (`git status --short src/app-bootstrap src/sample-portfolio` empty). `build.gradle:119-122` "Shipped so an operator can enable it" comment is known-open 0-A #2.
#### Coverage note
Deep: app-bootstrap main activation, runtime, outbox, idempotency, lock, notification, security, management; every `application*.yml`, `config/*.yml`, logback file; build.gradle; key ArchUnit declarations; sample entry point, bootstrap config, yml, ArchUnit test, relevant README/CLAUDE sections. Skimmed: tracing/logging/metrics code, test bodies, sample domain/application bodies (boundaries checked by grep).
### Group M — Vendored messaging platform (`messaging:*`)
#### Per-leaf verdict
| leaf | verdict | 한 줄 사유 |
|---|---|---|
| messaging-core-api | PASS | 문제 없음. 쓰이지 않는 타입 잔존 (#9) |
| messaging-schema-api / -json / -avro / -protobuf | PASS | codec decode 크기 제한 있음 (avro/protobuf 훑음) |
| messaging-cloudevents | PASS | 훑음. starter runtime 잔존은 known-open M13 |
| messaging-policy | NEEDS_FIX | `DeadLetterOrchestrator`가 인코딩된 envelope을 `publish`로 넘김 (#1) |
| messaging-transport-spi | PASS | 미사용 `MessagingLifecycle` (#9) |
| messaging-runtime-core | NEEDS_FIX | #1 원인 지점, #4, #5 |
| messaging-observability | PASS | 훑음 |
| messaging-security | PASS | leaf 자체 이상 없음; 적용 누락은 #2 |
| messaging-kafka | NEEDS_FIX | #2, #3(H6 회귀), #7 |
| messaging-kafka-share-experimental | PASS | 훑음; 옛 prefix 오류 메시지 (#8) |
| messaging-rabbit | PASS | 선택 경로가 막혀 도달 불가; #10 |
| messaging-reliability-api | PASS | — |
| messaging-outbox-jdbc-postgresql | NEEDS_FIX | #1, #4, #6 |
| messaging-inbox-jdbc-postgresql | NEEDS_FIX | #6 |
| messaging-claim-check | PASS | 조립처 없음; 무결성 검사 fail-closed |
| messaging-admin-api / -admin-runtime | PASS | HMAC 상수시간 비교, plan digest·operation·source 바인딩 |
| messaging-pulsar/nats-experimental | PASS | 훑음; SDK 없음, 기본 runtime classpath에 없음 |
| messaging-spring-cloud-stream-bridge | PASS | 훑음; prefix drift (#8) |
| messaging-spring-boot-starter | NEEDS_FIX | #1 조립, #5, #8 |
| messaging-spring-boot-starter-kafka | NEEDS_FIX | #2, #5 |
| messaging-spring-boot-starter-rabbit | PASS | validator만, broker 선택은 startup에서 거부 |
| messaging-testkit | PASS | — |
실험 leaf 기본 runtime 포함 여부: starter 의존 목록에 없고 `app-bootstrap/gradle.lockfile`에 pulsar/nats/kafka-share 없음. `kafka-clients`는 test/architecture classpath에만(lockfile:203).
#### Findings (심각도 순)
1. **[Critical][correctness] 플랫폼 내부 재발행 경로가 이미 인코딩된 payload를 다시 인코딩 → 전부 REJECTED** — `src/messaging/messaging-outbox-jdbc-postgresql/.../OutboxRelay.java:223-229`
- Evidence: `OutboxRelay.publish`, `DeadLetterOrchestrator.java:101-104`, `KafkaRetryTopicPublisher.java:68-70` 세 곳이 `MessageEnvelope<EncodedMessage>`를 `MessagePublisher.publish`로 넘김. 자동 조립 구현 `DefaultMessagePublisher`(`MessagingCoreAutoConfiguration.java:459-477`)는 `encode(message)` 무조건 호출(`DefaultMessagePublisher.java:177,303-308`) → `JacksonMessageCodec.encode`의 `registered.isInstance(payload)` 실패 → `PAYLOAD_TYPE_MISMATCH`(`JacksonMessageCodec.java:115-119`) → `PUBLISH_PREPARATION_FAILED`=REJECTED(`:178-179`) → relay가 `markFailed`로 terminal(`OutboxRelay.java:204-211`). bridge는 `publishEncoded`로 고쳐졌으나 이 세 곳은 아님. 관련 테스트(`OutboxRelayTest.java:416`, `DeadLetterOrchestratorTest`)는 가짜 publisher.
- Why: platform outbox를 켜면 모든 row FAILED — business commit은 성공, 메시지는 0건. DLQ publish도 confirm 없이 영원히 requeue.
- Fix: 세 호출자를 `EncodedMessagePublisher.publishEncoded`로(`messaging-schema-api` edge 허용됨); relay 뒤에 실제 `DefaultMessagePublisher`를 붙인 테스트.
- Status: new
2. **[Critical][security] 검증은 통과시키는 broker 보안 설정이 실제 producer에 적용 안 됨** — `src/messaging/messaging-spring-boot-starter-kafka/.../KafkaMessagingAutoConfiguration.java:128-154`
- Evidence: producer bean은 `spring.kafka.bootstrap-servers`, serializer, acks, idempotence만 설정; `KafkaSecurityConfigurer.configure` 호출 0건(bean 생성 `:109-111`뿐) → `app.messaging.security.<b>` TLS/SASL, `app.messaging.brokers.<b>.bootstrap-servers` 미사용. per-broker `access.publishable`는 compile만(`MessagingConfigurationCompiler.java:243-246`), 실제 publish 판정은 `DeclaredDestinationAccess.of(destinations.all())`(`MessagingCoreAutoConfiguration.java:414-419`). configurer를 연결해도 SASL username에 credential 참조 id가 들어감(`KafkaSecurityConfigurer.java:149-151`).
- Why: production profile이 `TLS_REQUIRED`·credential-provider 검사를 통과하지만 실제 연결은 PLAINTEXT·무인증. 보안을 검증했다는 결과만 남음.
- Fix: producer config를 compile된 `KafkaBrokerProfile` + `BrokerSecurityProfile`에서 만들고 `KafkaSecurityConfigurer`로 병합, 하나라도 없으면 startup 실패; ACL은 `DestinationAccessPolicy`에 교집합; username은 provider가 반환.
- Status: new (bootstrap 중복은 10-messaging §8-2에서 번호 없이 언급)
3. **[Required][correctness] handler timeout 시 취소가 아무것도 멈추지 않음** — `src/messaging/messaging-kafka/.../KafkaConsumerRegistrar.java:313-329`
- Evidence: `CompletableFuture.cancel(true)`는 실행 중 작업을 인터럽트하지 않고 완료 전이면 항상 true → `!cancel && !isDone` 분기 도달 불가 → permit 즉시 반환 → `requeueAfterFailure`가 같은 offset 재읽기 → 첫 handler가 돌고 있는데 같은 메시지 재처리. 새 테스트는 `isCancelled()`만 확인(`KafkaConsumerRegistrarTest.java:201`).
- Fix: handler 실제 종료(`whenComplete`) 시 permit 반환, 또는 우리가 소유한 `Future`에서 실행 후 `Future.cancel(true)`로 인터럽트; 테스트는 "첫 handler 종료 전 재dispatch 없음" 확인.
- Status: regression of H6 (closed 기록, 수정 무효)
4. **[Required][correctness] admission 거부가 동기 예외로 던져져 relay pass 전체 중단, poison row 무한 반복** — `src/messaging/messaging-runtime-core/.../DefaultMessagePublisher.java:227`
- Evidence: `admission.admit(...)`가 `try` 밖 → `MessageTooLargeException`/`MessageBackpressureException` 동기 throw; `OutboxRelay.runOnce`에 record 단위 catch 없음; CLAIM SQL은 `attempts` 미증가(`JdbcOutboxRepository.java:118-145`) → 크기 초과 row, 다른 writer가 넣은 reserved header row(`OutboxEnvelopeFactory.java:70-77`)가 FAILED/EXHAUSTED로 안 가고 매 lease 재claim; 같은 batch 나머지도 lease 종료까지 묶임. `OutboxRelayWorker.java:184-190`은 로그 없이 삼킴, `failedPasses()` 호출 0. relay는 REJECTED를 `retryable` 무관 terminal 처리. `DefaultBatchMessagePublisher.java:254`만 자기 쪽에서 막음.
- Fix: `DefaultMessagePublisher` 안에서 admission 예외 → REJECTED(backpressure는 retryable=true); relay record별 try/catch + lease mark; 실패 pass 로그·metric.
- Status: new
5. **[Required][performance] publish가 deadline 적용 전 caller thread를 최대 약 65초 막음(reactive thread 포함)** — `DefaultMessagePublisher.java:227-233`
- Evidence: `InFlightLimiter` 최대 5초(`MessagingCoreAutoConfiguration.java:436-437`) + `KafkaProducer.send` metadata/buffer 대기 `max.block.ms`(기본 60초), 둘 다 `withDeadline` 전 동기. producer bean에 `MAX_BLOCK_MS` 없음(IT는 직접 설정: `KafkaAmbiguityChaosIT.java:148`, `KafkaBrokerCertificationIT.java:410`). `DefaultReactiveMessagePublisher.java:37`이 이 전체를 subscriber thread에서 실행.
- Fix: `max.block.ms` ≤ publish timeout; reactive facade는 `subscribeOn(boundedElastic)` 또는 대기 없는 admission.
- Status: new
6. **[Required][performance] cleanup job이 제한 없는 단일 DELETE** — `OutboxCleanupJob.java:50`, `InboxCleanupJob.java:56`
- Evidence: 1-인자 `purge*Before(cutoff)` = `DELETE ... WHERE < ?` 한 번; limit CTE 버전(`JdbcOutboxRepository.java:485`, `JdbcInboxRepository.java:141`)은 main 호출 0; `InboxCleanupJob.DEFAULT_BATCH_SIZE` 미사용; javadoc은 "bounded batches".
- Fix: `(cutoff, batchSize)` 버전 호출, 삭제 건수 < batch면 중지; 1-인자 버전은 port에서 제거.
- Status: MSG-006 §6 불완전 종결
7. **[Required][correctness] consumer poison message가 지연 없이 무한 재시도(현재 잠재)** — `KafkaConsumerRegistrar.java:115,283-286,341,609`
- Evidence: 기본 quarantine `unavailable()`; decode 실패·handler 예외 → `enqueueRequeue(Duration.ZERO)` → 즉시 seek+resume → 매 poll 같은 record, partition 영구 차단. `DefaultDeliveryProcessor.java:70-97`도 시도 제한 없이 requeue; `RetryDecisionEngine` bean 미사용.
- Fix: `BackoffCalculator` 지연, 시도 한도 초과 시 DLQ/park; production profile에서 quarantine `unavailable()`이면 startup 거부.
- Assumption/Counter: adopter가 README대로 consumer 직접 조립 / 유실보다 막힘이 낫다는 설계 — 단 지연 0은 그 선택의 요구 조건 아님.
- Status: new
8. **[Consider][docs/config] 옛 prefix `backend.messaging` 잔존** — `messaging-spring-boot-starter/src/test/resources/application-valid.yml:6`, `application-invalid-ordering.yml`, `KafkaShareProfileValidator.java:35`, `StreamBridgePolicyGuard.java:34` (`grep -rnE "application-valid|invalid-ordering" src docs build-logic` → 과거 리뷰·plan만; 테스트 활성 profile은 `prod`/`production`뿐이라 fixture 미로드; 오류 메시지 두 곳이 `MessagingPrefixMigrationValidator`가 거부하는 키를 설정하라고 안내; `app.messaging.experimental.*` flag는 바인딩만). Fix: fixture 삭제, 메시지를 `MessagingSettings.PREFIX` 기준으로. Status: known-open L9 + 메시지 부분 new
9. **[Consider][structure] 미사용 API 타입** — `messaging-core-api/.../BatchMessageHandler.java`, `CapabilityRegistry`, `DelayedMessagePublisher`, `ManualMessageHandler`, `PauseResumeController`, `ProcessingGuarantee`, `transport-spi/.../MessagingLifecycle.java` (main 참조 0). Fix: 삭제. Counter: public surface라 breaking — 구현체 0. Status: known-open L7/L8
10. **[Consider][correctness] Rabbit publish sequence 예약 비원자적** — `RabbitMessagingTransport.java:148-161` (`nextPublishSequence()`와 `publish()` 사이 lock 없음 → 같은 sequence, `pending.put` 덮어쓰기 → 한 caller stage 미완료). Fix: `synchronized` 또는 단일 스레드 계약 명시. Counter: 현재 구현 없음·broker 선택 거부로 도달 불가.
11. **[Consider][correctness] outbox relay가 ordering key 순서를 지키지 않음** — `JdbcOutboxRepository.java:118-145` (CLAIM이 `ORDER BY created_at` + `SKIP LOCKED`, `ordering_key` 무시 → replica 여럿이거나 AMBIGUOUS row backoff 중 같은 key의 나중 row가 먼저). Fix: ordered destination + outbox 조합 startup 거부, 또는 같은 key의 더 오래된 미완료 row가 있으면 claim 안 함. Counter: relay는 "at-least-once, nothing more" 명시(`OutboxRelay.java:25-27`).
#### Leads checked and dismissed
- `application-valid.yml` lead: 사실 → #8. `MediaTypeVerifier`/`Sha256Verifier`/`ScriptableContentPolicy`는 `adapter/outbound/fileserver` 소속(범위 밖).
- M9 spring-kafka 코드상 해소(build.gradle·lockfile grep 0; `SpringKafkaTransactionalProcessor` 이름만 오해). M13: `cloudevents-api/core`가 `app-bootstrap` productionRuntimeClasspath에 아직 있음(lockfile:85-86) — known-open.
- HARD-STOP: messaging main의 domain/application/shared/adapter import 0; 모든 edge 일치.
- H4(Rabbit 분류 typed), H7(종료 commit 실패), H8(replay 승인) 해소 확인. `messaging-inbox-jpa`/`messaging-outbox-jpa` 디렉터리는 gitignore된 `build/` 산출물(M16 회귀 아님).
- outbox CLAIM fencing token, `next_attempt_at`, attempt budget·partial index, inbox `ON CONFLICT DO NOTHING` + 트랜잭션 강제 정상.
#### Coverage note
깊게: runtime-core publisher, starter core/kafka/rabbit/reliability 자동 설정·provider selection, outbox relay/worker/JDBC·SQL, inbox, `KafkaConsumerRegistrar`/transport/security configurer, admission/limiter, admin guard·HMAC, claim-check. 훑음: schema codec 3종, cloudevents, observability, rabbit(transport·classifier), 실험 leaf 3개, stream bridge, testkit.
### Group N — gRPC 플랫폼 (`grpc:*` / `grpc-advanced:*`)
#### Leaf별 판정
| leaf | 판정 | 이유 |
| --- | --- | --- |
| grpc-core-api | NEEDS_FIX | `GrpcDeadlineBudget`이 시간이 흘러도 만료되지 않음 (#2) |
| grpc-policy | NEEDS_FIX | FAILED_TERMINAL 재시도 (#1), writer 백프레셔 (#6), rotation race (#5) |
| grpc-operation-ledger-jpa | NEEDS_FIX | #1의 상대편 계약, claim 경로가 검증 우회 (#11) |
| grpc-server | NEEDS_FIX | admission race (#5), `live()` 가드 무효 (#2) |
| grpc-spring-boot-starter | NEEDS_FIX | environment 기본값 LOCAL → 설정 누락 시 보안 해제 (#3) |
| grpc-admin | NEEDS_FIX | drain 이후 SERVING 복귀 race (#5) |
| grpc-client | NEEDS_FIX | `rotate` get/set race (#5) |
| grpc-advanced-resilience | NEEDS_FIX | resolver가 retry/hedging 설정 주입 가능 (#4) |
| grpc-advanced-compat | PASS | envoy.yaml은 Consider (#8) |
| grpc-advanced-streaming | PASS | Nit (#12) |
| discovery, observability, codegen, proto-contract, testkit, advanced-bootstrap/diagnostics/edition | PASS | 의존 방향·순수성 확인, 큰 결함 없음 |
#### 발견 (심각도 순)
1. **[Critical][correctness] FAILED_TERMINAL 뒤 재시도가 claim 없이 PROCEED** — `src/grpc/grpc-policy/…/idempotency/GrpcIdempotencyInterceptor.java:109-114`
- 근거: 기존 FAILED_TERMINAL 행(insert 0행) → `decide`가 `proceed(...)` 반환(호출자가 행을 소유하지 않음) → `recordCommit` → `GrpcOperationLedgerEntity.markCommitted:123-124` → `requireInProgress()`가 `IllegalStateException`. fake `InMemoryOperationLedger.java:50-60`은 FAILED→COMMITTED 허용해 차이를 가림. JPA 테스트 `GrpcOperationLedgerRepositoryTest:190`("a failed claim is terminal")은 반대 계약 검증.
- 왜 문제인가: ledger 쓰기가 business tx 안이면 키 영구 차단(모든 재시도 INTERNAL); tx 밖이면 mutation 커밋 + 행 FAILED 유지 → 다음 재시도가 같은 작업 재실행; 동시 재시도 두 개가 모두 PROCEED(원장이 막으려던 중복 실행). production caller 0(잠복, prior C1과 같은 종류).
- 수정: port에 `reclaimFailed`(`UPDATE … SET state='IN_PROGRESS' … WHERE storage_key=:k AND state='FAILED_TERMINAL'`, 1행이면 소유); fake도 같은 전이 규칙; JPA·fake 공유 계약 테스트.
- 상태: new
2. **[Required][correctness] `GrpcDeadlineBudget`이 생성 시점 Duration을 들고 있어 만료 미감지** — `src/grpc/grpc-core-api/…/deadline/GrpcDeadlineBudget.java:13,82-83`
- 근거: `expired()`는 `remaining.isZero()`뿐(시계 입력 없음). 사용처: `GrpcRequestContext.live():84-85`, `GrpcServiceAdapter.handle:69-74`("deadline has passed" 가드), `GrpcContextSnapshot.live(now):69-71`, `GrpcClientCallContext:54`. 테스트는 `Duration.ZERO` budget만.
- 왜 문제인가: 큐에서 deadline을 넘긴 호출도 use case 진입; 워커에서 `GrpcDeadlineCalculator.forDependency`가 downstream에 inbound보다 긴 deadline 부여; `GrpcRetryCoordinator` backoff 비교도 오래된 값.
- 수정: 절대 시각 저장(`long deadlineNanos` + 주입 `LongSupplier`, 또는 `Instant` + `Clock`) → `remaining(now)`/`expired(now)`; 시간 경과 테스트.
- 상태: new
3. **[Required][security] starter가 `environment` 미지정 시 LOCAL로 떠 보안 검사 전부 꺼짐** — `src/grpc/grpc-spring-boot-starter/…/boot/GrpcPlatformProperties.java:26`
- 근거: TLS off·trust-all·reflection ENABLED 거부 검사(`GrpcPlatformStartupValidator.java:87-102`)가 모두 `tlsRequired` 조건 안; `defaultFor(LOCAL)`은 ENABLED이고 bean이 됨(`GrpcPlatformAutoConfiguration.java:86-88`); `enabled=true`만 켜고 environment 누락한 운영 배포 → schema 전체 공개 + trust-all로 통과. 테스트는 전부 environment 명시(`GrpcPlatformAutoConfigurationContextTest:35,53`).
- 수정: 기본값 제거, `enabled=true`인데 environment 없으면 violation(또는 기본값 PROD).
- 상태: new
4. **[Required][security] resolver service config의 retry/hedging 미검증** — `src/grpc-advanced/grpc-advanced-resilience/…/discovery/GrpcResolverSafetyPolicy.java:56-65,86-91`
- 근거: `GrpcResolverUpdate.java:8-11`은 "retry policy를 사람이 쓴 것처럼 검증" 약속; 실제 검사는 authority·revision·credential 유사 문자열 정규식뿐 → `methodConfig[].retryPolicy`/`hedgingPolicy`가 `acceptedServiceConfig`로 통과; 반대로 정상인 `"retryThrottling":{"maxTokens":…}`는 `token` 부분일치로 거부.
- 수정: JSON 파싱 + top-level 키 allowlist; retry/hedging은 `GrpcMethodPolicyCatalog` + `GrpcHedgingEligibility`로 검증하거나 거부; substring 정규식 삭제.
- 상태: new
5. **[Required][correctness] 상태 보유 클래스들의 check-then-act / get-then-set race**
- `src/grpc/grpc-server/…/GrpcAdmissionController.java:51-60`: `get() < max` 확인 후 `incrementAndGet()` → 동시 수만큼 bound 초과.
- `src/grpc/grpc-policy/…/security/GrpcCredentialRotationManager.java:89-103,118-121`: `AtomicReference` get/set만 → `completeDrain`이 `rotate`와 겹치면 교체된 credential로 되돌아감; rotator 둘이 겹치면 문서(`:82`)가 금지한 역방향 rotation 통과.
- `src/grpc/grpc-client/…/GrpcChannelRuntimeRegistry.java:91-103`: 같은 패턴, replacement 하나가 drain 목록에서 빠져 누수; `:114,128` synchronizedList를 락 없이 순회.
- `src/grpc/grpc-admin/…/GrpcServiceHealthRegistry.java:50-53` vs `:85`: `draining` 확인과 `put(SERVING)` 사이 drain 시작 → drain 후 global SERVING(`:45` 주석이 막겠다던 상황).
- `src/grpc/grpc-policy/…/streaming/GrpcStreamAdmission.java:42-49`: `computeIfAbsent`가 global bound 검사 앞 → 포화 시 거부된 신규 caller마다 count 0 엔트리 영구 잔류.
- 수정: `synchronized` 또는 `updateAndGet`/CAS; `GrpcStreamAdmission`은 global 검사를 앞으로; `GrpcStreamAdmissionConcurrencyTest` 패턴으로 클래스별 동시성 테스트.
- 상태: `GrpcAdmissionController` known-open(prior M1 근거에서 지목, 종결 때 `GrpcStreamAdmission`만 고침); `GrpcStreamAdmission` 잔류는 M1(b) 미완; 나머지 new.
6. **[Required][correctness] `GrpcSerializedStreamWriter`가 백프레셔 신호를 잃음** — `src/grpc/grpc-policy/…/streaming/GrpcSerializedStreamWriter.java:47,73,100-103,126-139`
- 근거: (a) `LongSupplier payloadSizer`는 실제 payload 크기를 모름(테스트는 공유 `AtomicLong`을 미리 set — 동시 producer에선 다른 메시지 크기); (b) `PAUSE`를 `ACCEPTED`로 반환; (c) `flush()`가 isReady 무관하게 큐 전체 전송. `GrpcFlowControlPolicy.java:11`이 경고한 "isReady를 무시하는 writer"가 바로 이 구현.
- 수정: `ToLongFunction<? super T>`; `ACCEPTED_PAUSE` 결과 추가; `flush(BooleanSupplier ready)`.
- 상태: H3 잔여분(`ToLongFunction` 미적용) + (b)(c) new
7. **[Consider][correctness] `GrpcContextBinder.wrap`이 io.grpc `Context`(취소·deadline)를 스레드 경계에서 버림** — `src/grpc/grpc-policy/…/context/GrpcContextBinder.java:12,60,86-98` (SNAPSHOT 값만 캡처해 실행 스레드의 `Context.current()`(보통 ROOT)에 붙임). 수정: `Context.current()` 캡처 후 `captured.wrap(work)`. 반론: platform 자체 token·`parentRemaining`은 따로 전달.
8. **[Consider][config] `envoy.yaml` 참조 설정이 그대로는 동작하지 않음** — `src/grpc-advanced/grpc-advanced-compat/src/main/resources/envoy/envoy.yaml:32-39,50-53` (`DownstreamTlsContext`에 인증서 없음 — `envoy --mode validate` 확인 필요; `allow_headers`에 `x-user-agent`, `grpc-timeout` 없음; `expose_headers`가 `error-code,error-category`인데 `GrpcErrorMapper.java:30-39` 실제 trailer는 `error-reason`, `error-execution-id`, `completion-outcome`, `grpc-status-details-bin` 미노출; `timeout: 30s`로 server stream 30초 절단; 테스트 `GrpcWebCompatibilityGateTest:124`는 부분 문자열만; main resources라 jar에 포함). 수정: test resources/docs로 이동, header 목록을 Java 상수와 대조하는 테스트, `grpc_timeout_header_max: 0s`.
9. **[Consider][architecture] starter의 미사용 의존과 아무것도 검사하지 않는 runtime 불변식** — `src/grpc/grpc-spring-boot-starter/build.gradle:21-22`, `GrpcPlatformAutoConfiguration.java:55-67,129-133` (main에서 `grpc.contract`/`grpc.codegen` import 0; `requireValid(..., Set.of())`로 advanced 격리 검사 항상 통과; `requireStableStarterIsClean` main caller 0 → `src/grpc/CLAUDE.md` "runtime에 강제" 설명 거짓; server profile bean이 `stableNetty` 값을 리터럴 복사). 수정: 두 edge + registry edge 삭제, 파라미터·`validateAdvancedIsolation` 삭제(Gradle gate 기준), 문서 정정, 팩토리 재사용. 반론: reflection에 descriptor 필요 가능.
10. **[Consider][structure] 같은 한계값이 여러 곳에 따로 정의** — 1 MiB/8192가 `GrpcServerProfile:70-71`, starter `:59-60`, `GrpcMessageSizeProfile:43`, `GrpcNamedChannelProfile:80-81,95-96`, `GrpcMetadataBudget:33`에 각각(정합 검사 없음); `GrpcStreamProfile`은 자기 테스트에서만 참조, `GrpcFlowControlPolicy`의 256/8MiB 복제. grpc-policy 62파일 규모 자체는 과하지 않음(실제 비용은 production 조립 부재로 fake가 실제 adapter와 달라지는 것, #1). 수정: `GrpcTransportLimits` 하나로, validator에 "method profile ≤ server profile", `GrpcStreamProfile` 삭제.
11. **[Consider][correctness] ledger claim 경로가 entity 검증 우회, storage_key가 컬럼 길이 초과 가능** — `src/grpc/grpc-operation-ledger-jpa/…/JpaGrpcOperationLedger.java:39-53` (native insert가 `sha256:` 형식 검증 안 거침; `callerFingerprint` 길이 무제한 → storageKey 최대 585자 vs `VARCHAR(512)`; insert 0행 + 조회 실패를 소유 획득으로 처리). 수정: `claim()`에서 형식 검증, fingerprint ≤256자, 키 해시로 고정 길이, 0행+조회 실패는 예외.
12. **[Nit][performance]** `GrpcClientMessageDeduplicator.java:11-17,75,112` ("constant space"라지만 outcome이 세션 내내 누적, `endSession`이 O(N) `removeIf`); `ReactiveGrpcClient.serverStream:50`은 List 래퍼일 뿐(삭제).
#### 확인했지만 문제 없던 항목
- HARD-STOP 1-5: core-api 의존 0, framework-free 5개 leaf에 외부 import 없음, 모든 edge 허용, grpc→advanced 참조 0.
- 0-A/0-B(ledger IT CI 미호출) 닫힘: `optional-platforms/build.gradle` `integrationCheck` + `integration-main.yml:313`. H3, M2 회귀 없음.
- ledger SQL named param, `ON CONFLICT`가 PK·unique 모두 커버; NON_IDEMPOTENT retry 차단, hedging은 READ_ONLY unary만; `mapUnknown` description opaque; envoy admin 블록 없음·CORS exact; resume token HMAC + 상수시간 비교; `GrpcDrainCoordinator` 미거부는 prior §2에 기록됨.
#### 커버리지
자세히: ledger-jpa 전체, grpc-policy idempotency/streaming/context/deadline/error/security/resilience, server 핵심, starter 전체, admin, client, advanced resilience/diagnostics/compat + envoy, 944a1e34 이후 바뀐 main 파일 전부. 훑음: codegen, proto-contract validator, observability, edition, bootstrap, testkit.
### Group O — 빌드 시스템, 아키텍처 registry, 의존성 게이트
#### 영역별 판정
| area | verdict | 이유 |
|---|---|---|
| root lifecycle / settings | NEEDS_FIX | `ca.developer-bootstrap`가 설정 단계에서 실패해 Docker/release 이미지 빌드 불가 |
| CI workflows | BLOCKED | `redis-sdk` job이 항상 실패 → `release-gate` 매번 빨간불 |
| architecture gates (build-logic) | NEEDS_FIX | 방향·purity 게이트 정상. `verifyNoIgnoredSourcePackages`는 범위가 좁고 git 오류 시 통과(fail-open) |
| conventions (java/quality/lanes) | PASS | 품질 도구 배선이 리팩터 전 groovy 버전과 같고 약화 없음 |
| modules.json / leaf build.gradle | NEEDS_FIX | registry 밖 production edge 0건. app-bootstrap test 입력 미선언, 좌표 중복 |
| build-tools | PASS | 전부 fail-closed. 중복·규모는 known-open |
| supply chain | PASS (Consider) | 버전 고정 + STRICT lock 있음, checksum 검증 없음 |
#### Findings (심각도 순)
1. **[Critical][correctness] root plugin 적용 시점에 compose overlay가 없으면 빌드 전체 실패** — `src/build-logic/src/main/java/dev/caskeleton/buildlogic/bootstrap/DeveloperBootstrapPlugin.java:40-50`
- Evidence: `apply()` 안에서 `../docker-compose.<profile>.yml`이 없으면 `GradleException`. root에서 적용(`src/build.gradle:29`). release는 context `src`로 `docker build --file src/Dockerfile ... src`(`.github/workflows/release.yml:329-338`); Dockerfile은 `WORKDIR /build/src`(`src/Dockerfile:29`)에서 `RUN ./gradlew verifyDependencyLocks`(`:57-60`) — `/build/docker-compose.local.yml`은 존재할 수 없음. 208077fd에서 추가.
- Why: root 설정 단계에서 죽어 어떤 task도 실행 안 됨 → release 이미지 빌드 차단, Dockerfile 헤더의 로컬 명령도 실패.
- Fix: overlay 확인을 `bootstrapDockerPreflight`의 `doFirst`로 이동, `apply()`는 부수효과 없음; TestKit 회귀 케이스(compose 파일 없는 root에서 `help` 성공).
- Status: new (regression 208077fd)
2. **[Critical][CI] `redis-sdk` job은 통과 불가, `release-gate`가 이 job의 success를 요구** — `.github/workflows/ci-quality-gates.yml:120`, `.github/workflows/_reusable-gradle.yml:58-63`
- Evidence: 입력 `":shared-contract:test --tests '*EdgeRateLimitProviderNeutralContractTest'"`를 reusable workflow가 `./gradlew ${GRADLE_TASKS}`로 따옴표 없이 펼침 → bash는 확장 결과의 따옴표를 제거하지 않음(`printf '[%s]'` 재현 → `['*EdgeRateLimitProviderNeutralContractTest']`) → 어떤 클래스와도 매칭 안 됨 → `failOnNoMatchingTests` 기본값으로 실패. `release-gate`는 `REDIS_RESULT == success` 요구(`:240,246-258`). 9da00eb7에서 변경. 이 클래스는 tag가 없어 이미 `:ci` → `:shared-contract:check` → `test`로 실행되므로 job 자체가 중복.
- Fix: `redis-sdk` job과 `release-gate` 해당 줄 삭제(남긴다면 따옴표 제거).
- Status: new (regression 9da00eb7). CI 미실행, bash/Gradle 동작에서 도출.
3. **[Required][HARD-STOP 6] "Git이 커밋할 수 있는 소스" 게이트가 main/test만 보고 git 오류 시 통과** — `ArchitecturePlugin.java:191-203`, `VerifyNoIgnoredSourcePackagesTask.java:70-87,102-119`
- Evidence: 대상이 `List.of("src/main/java","src/test/java")` 하드코딩 → 나머지 Java source root 45개(testFixtures 10, integrationTest 10, systemTest 5, architectureTest 등; H9가 다룬 종류) 미검사. `git rev-parse`/`check-ignore` exit code 무시 → 128(fatal)이면 stdout 비어 "not a Git checkout" 또는 위반 0으로 통과; `IOException`도 null → 통과. `optional-platforms`의 `ci`는 이 task에 의존하지 않아 grpc 소스 미검사. 테스트 0(`grep -rl IgnoredSource build-logic/src/test app-bootstrap/src build-tools/src/test`).
- Fix: 모든 project `sourceSets*.java.srcDirs` 수집; `.git`이 있는데 git이 0/1 외 코드로 끝나거나 실행 파일이 없으면 실패; `integrationTest`에 ignore 대상 패키지를 넣은 TestKit 케이스.
- Status: new (groovy 시절부터의 공백, regression 아님)
4. **[Required][correctness/cache] app-bootstrap `test`가 선언하지 않은 저장소 파일을 읽음 → 로컬 UP-TO-DATE 거짓 통과** — `src/app-bootstrap/build.gradle:270-287`
- Evidence: `ca.repository.root`로 `docs/runbooks/**`(`RunbookCoverageContractTest.java:129`, `ContainerRuntimeOomContractTest.java:110`), `src/.env.example`(`EnvProfileMatrixContractTest.java:30`)을 읽는데 선언 입력은 `docs/registries`뿐(`:282`). `org.gradle.caching=true`에서 runbook만 수정하면 `:app-bootstrap:test` UP-TO-DATE. 같은 패턴 `adapter/outbound/messaging/build.gradle:30-34`(schema 경로 sysprop, `inputs.file` 없음). CI에선 `GITHUB_SHA` → version `0.0.1+<sha>`(`ReleaseProvenancePlugin.java:24-37`) → 모든 jar manifest(`JavaConventionsPlugin.java:45-48,66-75`) → 커밋마다 test runtime classpath가 달라 cache miss(가려짐; CI의 test/SpotBugs cache 이득 0). cache를 살리려고 `normalization.runtimeClasspath.metaInf.ignoreAttribute`를 넣는 순간 거짓 통과가 CI로 이동.
- Fix: `inputs.dir(docs/runbooks)`, `inputs.file(src/.env.example)`(RELATIVE), messaging `inputs.file(schema)` 선언 후 manifest normalization 검토.
- Assumption/Counter: setup-gradle이 `caches/build-cache-1`을 유지 / 유지하지 않아도 로컬 focused test 거짓 통과만으로 HARD-STOP 6 위험.
- Status: new
5. **[Consider][supply chain] artifact checksum 검증 없음, included build 의존성 미lock** — `src/gradle/`, `src/build-logic/settings.gradle:6-9`, `src/build-tools/settings.gradle` (`verification-metadata.xml` 없음; build-logic은 spotless/spotbugs/errorprone/dependency-management plugin을 lock·hash 없이 `gradlePluginPortal`에서; `settings-gradle.lockfile`은 catalog만). Fix: `--write-verification-metadata sha256` 커밋, build-logic/build-tools `dependencyLocking`. Counter: Trivy·dependency-review·고정 버전이 CVE 위험은 막음, metadata 갱신 부담.
6. **[Consider][correctness] 프로젝트 좌표 `dev.caskeleton:messaging-kafka` 중복** — `:adapter:outbound:messaging-kafka`(9da00eb7) vs `:messaging:messaging-kafka` (같은 이름, 같은 group(`JavaConventionsPlugin.java:44`); `VerifyDependencyPolicyTask`는 `group:name` 매칭이라 구분 불가; 같은 jar 파일명 → 한 composition에 둘 다 들어가면 `BOOT-INF/lib` 충돌; Gradle 충돌 해결이 둘을 합치는지 미확인). Fix: adapter leaf 이름 변경(예: `:adapter:outbound:kafka-sender`) + modules.json·디렉터리.
7. **[Consider][simplicity] runtime membership 검사가 방향 게이트와 중복** — `RuntimeMembershipPlugin.java:35-44`, `VerifyCompositionRuntimeMembershipTask.java:166-188` ("등록된" 집합 = sample 제외 main 모듈 전부; settings는 registry 모듈만 include(`ModuleRegistrySettingsInstaller.java:278-280`) → "미등록" 가능한 건 `:sample-portfolio`뿐이고 방향 게이트가 이미 차단; build-logic README의 "registry's memberships" 필드는 없음). Fix: plugin, alias task, `ArchitecturePlugin.java:216-225`, `ConditionalTransportQualificationPlugin.java:56` 연결 삭제. Counter: 싸고 미래 보호막.
8. **[Consider][simplicity] "지정 테스트가 실행됐고 skip 없음" 검사 구현이 5개** — `ca.strict-test-lane` `requires`/`rejectSkipped`; `ca.strict-qualification`(RequiredClasses task + listener + XML 재검사, 10개 leaf); `ConditionalTransportQualificationPlugin` XML 재읽기; `GraphQlPlatformPlugin` `graphqlStableTest` 자체 XML 검사; build-tools `JUnitEvidenceReader`(build-logic `JUnitEvidence` 복사본, README는 "one JUnit XML reader"). 규모: build-logic main 83파일/5,995 LOC, build-tools 98파일/5,888 LOC. Fix: 모두 `strictTestLanes { lane { requires ...; rejectSkipped = true } }`로, 나머지 삭제. Status: known-open (13-build §0-A/0-B)
9. **[Consider][performance] quarantine job이 테스트 0개를 돌리려고 저장소 전체 컴파일** — `ci-quality-gates.yml:212-217`, `src/build.gradle:88-92`, `JavaConventionsPlugin.java:149-165` (leaf마다 test JVM, up-to-date 판정 없음; `Tag("quarantine")` 0건). Fix: job·task 삭제, 첫 quarantine 테스트 때 재추가.
10. **[Nit][readability]** httpclient가 같은 sysprop 3줄을 7번 반복(`adapter/outbound/httpclient/build.gradle:131-229`); `-Duser.timezone=UTC`가 leaf build 파일에 18번 → `JavaConventionsPlugin`에 `withType(Test).configureEach { systemProperty 'user.timezone','UTC' }` 한 번(jvmArgs 아님 — `StrictTestLanePlugin.java:176-178` `setJvmArgs`가 덮어씀).
11. **[Nit][docs]** CLAUDE.md `ci` 정의에 `sampleWebContractCheck`(`build.gradle:146`) 누락, `verifyDependencyLocks`는 `ci` 밖(workflow에서만, `ci-quality-gates.yml:62`); `libs.versions.toml:27-31`은 BOM override가 `src/build.gradle`에 이유와 함께 있다고 적었지만 실제 `JavaConventionsPlugin.java:90-92`이고 이유 사라짐, `netty.version`은 Boot 4.0.8 BOM 값(4.2.17.Final)과 같아 무효; `optional-platforms/README.md:3` 모듈 개수("18") 기재(no-count 규칙 위반).
#### 확인했지만 문제 아닌 것
- M6 (parallel): 방향/purity/membership task가 소유 project 안에서 `Provider<ResolvedComponentResult>`로 resolve(`ArchitecturePlugin.java:96-189`), `parallel=true` — 닫힌 것으로 보임(`--parallel` TestKit 없음).
- H9 (commit 포함): untracked는 graphify-out/.playwright-mcp뿐, 모든 leaf에 lockfile — 닫힌 것으로 보임.
- 이전 H1/H2, gRPC ledger lane 호출자: 닫힘.
- 정적 edge 감사: 모든 leaf `<conf> project(...)`를 스크립트로 registry와 대조 → 허용 밖 production edge 0.
- 품질 게이트 약화 없음(checkstyle severity=error라 `maxWarnings=MAX` 무영향; errorprone/`-Werror`/`ignoreFailures`/`onlyIf` 해제 0). build-logic/build-tools catch 블록 전부 재던짐/실패.
- tree-sitter 파싱 실패 22건: 유효한 Groovy(파서 한계); `fileserver/build.gradle:10`의 `\`만 불필요.
- wrapper sha256 검증, action SHA 고정; quarantine·http3 job은 의도된 advisory.
#### 커버리지
깊게: root build/settings/properties, ArchitecturePlugin·task들, ModuleRegistry, RuntimeMembership, Java/Quality conventions, StrictTestLane/StrictQualification, DependencyPolicy, ReleaseProvenance, DeveloperBootstrap, ci-quality-gates·reusable workflow, Dockerfile builder stage, 큰 leaf build 4개, modules.json(스크립트). 훑음: build-tools, 나머지 leaf, integration/release workflow.