feat(ql3): prove shadow resource rollback

This commit is contained in:
whyour
2026-08-19 02:36:06 +08:00
parent b739d92347
commit c60a86d1bf
12 changed files with 1564 additions and 89 deletions
+15 -1
View File
@@ -11,6 +11,20 @@
最新增量证据(2026-08-19):
- D-359/ADR-0451、ADR-0452(已接受):为 D-358 的闭合窗口审计增加真实 compiled-backend 资源与关闭回滚门。证据 schema 固定为
`qinglong/legacy-shadow-resource-rollback-evidence@v1`edge 为 8 candidates/1 pagestandalone 为 128 candidates/4 pages8 个样本分别精确执行
16/64 条查询,SQLite database/WAL/SHM/journal 的 logical/allocated bytes 和 file count 前后必须一致。full 模式经过独立进程重启:enabled 与 off
都执行真实 `ScheduleService.runTask` Legacy child 且 exit 0Shadow Run 增量必须从 1 变为 0off 还要求零 fact factory、默认 observer/Repository
未加载。Linux arm64 router stress 保持 128 MiB/0.5 CPU/0 swap/64 PIDcgroup peak `80,416,768` bytesedge audit p95 `3.914 ms`Edge release
保持 256 MiB/1 CPU/0 swap/128 PIDcgroup peak `139,079,680` bytesedge full/standalone audit p95 分别为 `32.056/19.265 ms`,两档
`memory.events max/oom/oom_kill` 增量均为 0。双 Node crash harness 留在 256 MiB 发布门,完整 Edge+Standalone Workflow/Prompt crash matrix 仍通过;路由门只跑
单进程产品/恢复 workload,不提高内存或 timeout。只读 mount 验证改为解析路径实际生效的最具体 mount:`/workspace` 可继承只读 `/`,但任一可写 descendant
仍失败。干净镜像同时暴露历史 `build:back` 被 stale artifact 掩盖的 TypeScript common-root 错位;新构建器在隔离 staging 编译,只原子发布
`output/back` 到既有扁平 `static/build/*`,重写 source map,拒绝携带 `output/packages`,失败保留上一版。D-359 不新增 package、生产依赖、schema、
migration、数据库 authority、timer/watcher、端口或部署对象。完整 backend 与 18-package clean build/test 退出 0;静态契约 `126/126`,四项可执行
架构审计与 14/14 artifact budget 全部通过,制品字节保持 D-358 基线。D-359 不把 Docker arm64 冒充物理路由/断电/flash 证据。D-360 继续负责可信的
Legacy→Shadow capture authority、具体 exporter 与正式 Primary gate。
- D-358/ADR-0450(已接受):新增显式、只读、一次性的 Legacy Shadow 闭合窗口终态审计。调用方必须提供 origin 与
`[windowStartMs, windowEndMs)`,cohort 固定为窗口内创建的 legacy-owned Shadow Run;只有窗口经过默认五分钟 settling、候选 keyset 与
Legacy evidence 都完整且分母非零时,才输出 terminal agreement/full comparability permille。关联只接受 direct Run/Attempt reference、同 Cron
@@ -9129,7 +9143,7 @@ flowchart LR
| PR-1 Run Schema | Incubating | Run/RunAttempt/RunEvent schema、nullable cancel request 与 Attempt deadline 字段及恢复索引、CancellationDispatch 状态/version/lease/backoff schema、Repository port、临时 Sequelize adapter、统一事件大小/分页上限、跨 adapter RunRepository contract suite(原子事务、回滚、Run/Attempt/RetryPolicy CAS、唯一错误、分页与取消恢复);ADR-0041 的 `pg-0003-run-retry-policy`、capability v2、driver-neutral PostgreSQL Run Repository 与真实 `pg.Pool` 上的共享 Repository/rollback/SQLSTATE contractADR-0063/0069/0071/0073/0074/0076 的独立 Node 24 local-sqlite typed schema、十二条 reviewed migration、capability v6、共享 operation authority、readiness/RunRepository/API credential repository/receipt journal/dispatch plan/encrypted Secret envelope/Project Policy/security audit/authorized mutation/stable Identity catalog、Drizzle↔真实 catalog table/column/index/CHECK/FK lockstep、base/adopted/application edge/standalone 产物门禁;ADR-0064 的 legacy baseline/plan digest、Online Backup recovery、side-by-side target migration、staged manifest、双库栅栏 activation、source 生命周期写栅栏、target stable identity 和重启语义;ADR-0065 的独立 cutover authority、外部副作用停机 evidence、append-only journal、start/restart/stop barrier 与 unknown→manual_required 收敛;ADR-0066 的 adopted storage→Run reconciliation→receipt maintenance→domain recovery→lifecycle→admission application gate、严格有界 recovery summary 与 admission-first reverse stopADR-0067 的 SQLite 事实驱动 Run 候选源、256 条硬上限、截断失败关闭和唯一 Repository authorityADR-0068 的 receipt-first Reconciler、callback token/sequence fence、exact local-process identity、Attempt/Run/双 Event 原子终态推进和最终 verifierADR-0069 的 local-process 单向包边界、pre-spawn journal、受审 POSIX launcher、immutable receipt、exact identity 和 Profile-aware cleanup lifecycleADR-0070 的独立 local-execution、spawn 前后双 transaction CAS、callback digest、exact stop 补偿与 fail-closed starting 保留;ADR-0071 的独立 local-dispatch、不可变 revision/context、Secret-first materializer、Profile Artifact admission、4/64 MiB output hard quota 和窄 application facadeADR-0073/0074 的 Project-bound SecretRef、AES-256-GCM、外置 keyring 生命周期、双 SQLite authority CAS、application preflight、强 Principal/Policy 和 envelope+audit 原子提交;ADR-0086 的本机 Owner provisioning/challenge/claim/delivery acknowledgement/credential recovery CLIADR-0377 的 Local/Cluster 同构、Profile-aware、Project-scoped Artifact range read | fresh database/pepper setup、credential rotation/GC 运维编排与 Secret/Project/Role/Approval 管理 CLI/API/UI、备份/rekey、2.x/target process controller、人工 recovery、target 写后 reconciliation 与完整 cutover/rollback 演练;retry 产品策略、Artifact retention/tombstone stack、具体本机 lifecycle 和 target executableLinux x64/arm64、PID namespace、断电与固定路由设备门禁;PostgreSQL 16/18 双连接并发与 failover integrationTask revision/context 跨方言 contract/并发压力与引用感知 retention、Keyv 数据迁移 |
| PR-2 Run 状态机 | Incubating | 纯转换表、终态/时间/错误/执行器元数据规则、Run version 与 event sequence CAS、事务性 RunCommandService、回滚测试 | 重复 Worker callback/fencing、并发数据库压力测试、Primary 执行链接入 |
| PR-3 Executor 端口 | Incubating | ADR-0003、ExecutionSpec/Context/Handle/Result、Executor port、LocalProcessExecutor、进程组取消/超时升级、流式背压、Legacy Cron spec builder、真实进程 contract tests、可复现 edge 基准入口 | 固定 edge/多架构设备基线、Legacy builder 与 makeCommand 差异审计、Primary 生产流量接入 |
| PR-4 Shadow Run | Incubating | origin 三态策略;默认关闭的 `QL3_SHADOW_ORIGINS`manual、scheduled_node、boot、subscription、system 与 script 现有 ChildProcess 旁路观察;system crond 显式 origin marker、Shell execution ID、finish-only 准入、确定性 Run/Attempt 与 exact replay`@once` 保持 manual、gRPC transport 不冒充 origin 的准入裁决;每个 worker 懒加载;Run/Attempt/Event 影子生命周期;稳定且不复制 caller 原文的 task identity/revision 与有界日志引用;同 worker 有界注册表和跨 worker 持久化候选关联;stop all/stop instance、Shell callback、乱序/迟到/歧义处理;监听前一次性、Profile-aware 的 keyset Startup Reconciler,终态证据补齐、lost/abandoned/pending 分流与 terminal Attempt response-loss 修复;origin-bounded 且逐级守恒的版本化 startup difference report、固定字段 metric batch 与一次性 collector;显式、只读、闭合窗口且 Profile-bounded 的 Shadow→Legacy 终态差异审计;失败开放和契约测试 | Legacy→Shadow capture authority、具体 exporter、资源压力、回滚演练和 Primary 门禁 |
| PR-4 Shadow Run | Incubating | origin 三态策略;默认关闭的 `QL3_SHADOW_ORIGINS`manual、scheduled_node、boot、subscription、system 与 script 现有 ChildProcess 旁路观察;system crond 显式 origin marker、Shell execution ID、finish-only 准入、确定性 Run/Attempt 与 exact replay`@once` 保持 manual、gRPC transport 不冒充 origin 的准入裁决;每个 worker 懒加载;Run/Attempt/Event 影子生命周期;稳定且不复制 caller 原文的 task identity/revision 与有界日志引用;同 worker 有界注册表和跨 worker 持久化候选关联;stop all/stop instance、Shell callback、乱序/迟到/歧义处理;监听前一次性、Profile-aware 的 keyset Startup Reconciler,终态证据补齐、lost/abandoned/pending 分流与 terminal Attempt response-loss 修复;origin-bounded 且逐级守恒的版本化 startup difference report、固定字段 metric batch 与一次性 collector;显式、只读、闭合窗口且 Profile-bounded 的 Shadow→Legacy 终态差异审计;128/256 MiB Linux arm64 资源门、SQLite 零增长与 Shadow enabled→off 进程重启回滚;失败开放和契约测试 | Legacy→Shadow capture authority、具体 exporter 和 Primary 门禁 |
| PR-5 Primary LocalExecutor | Incubating(默认不激活,仅 manifest-gated manual | runtime-owned Run 创建器;持久化先于 spawnRun/Attempt 完整成功、失败、取消、超时与 lost 闭环;Executor handle 身份校验;spawn 后激活写失败的 stop+lost 补偿;completion rejection 安全收敛;独立 Primary 幂等查询与唯一索引竞态裁决;durable `run.cancel_requested`、stop-before-signal、首次请求幂等、晚到完成裁决与待取消有界恢复查询;最多 64 条一页的 cross-worker cancellation source;独立 CancellationDispatch Repository 原子 claim/result、lease expiry 接管、owner/token/version fencing、指数退避与结果 RunEvent;最多 64 页的单周期 cancel supervisor;显式 start/stop、无重叠、错误隔离、停止等待有上限且 timer unref 的 lifecycle runnerLinux durable handle 的 PID/boot/start ticks/process-group 复验与 TERM/KILL controller;完整有界分页且 fail-closed 的 startup Reconcile supervisorRunningInstance nullable `run_id/attempt_id` 关联;Primary 专用组合 Repository 在同一 SQLite 事务提交前投影 Crontab/RunningInstance,失败整体回滚;有界且防穿越的 legacy log output refmanual owner seam、真实本机装配、单 spawn/fail-closed;严格 manual-only rollout manifest loader、短期审批/gate、配置哈希审计;HTTP worker 已接轻量 lazy bootstrapaccepted 后按 receipt-first reconcile→completion receipt lifecycle→timeout intent lifecycle→cancel dispatch lifecycle→router 顺序激活,失败撤销,监听失败和 shutdown 有界停止;Primary timeout 在 spawn 前持久化绝对 deadline,有界 source/requester/supervisor 只提交 timeout 意图并复用 CancellationDispatch;代码级 edge/standalone Profile 为各 lifecycle 提供不同 cadence 与页上限,cluster-control/worker 拒绝误装本机 SQLite Primary;统一 CompletionService 原子提交 Attempt/Run/双 Eventspawn 前保存 callback token hash、终态推进 sequence,实时回调与 receipt consumer 共享入口并覆盖两个清理 crash windowmanual Primary 已接入受限 POSIX launcher、`0600` direct-file stdout/stderr、父进程退出后续写、不可覆盖 receipt 生产、回执环境清除、TERM 转发等待及 live transaction 后清理;Startup Reconciler receipt-first 双检查并在确定 exited 后执行 profile 化的单次 50/100 ms publish grace`0007` 独立 CompletionReceiptJournal 在 spawn 前登记、为升级前 active Attempt 补登记并驱动周期扫描,使终态残留继续可发现;确定无效的已知 Attempt receipt 先持久化隔离状态,再进入确定性私有分片 quarantine;终态 missing 与 quarantine 按 edge/standalone retention 有界清理;非 Journal 文件具备只读优先、固定分片/条目上限、overflow fail-closed、显式同盘隔离的 Node 24 运维 CLI;扫描具备页上限、resume cursor、timer unref、无重叠、有界 stop 和低敏计数;ENOSPC 与 launcher receipt 存储失败有代码门禁 | 部署配置写入/审批入口与用户可见状态;PostgreSQL CancellationDispatch adaptercluster-control 生产启动拓扑;固定 edge/Linux 多架构与真实磁盘压力基线、完整 2.x API 契约和回滚演练 |
| PR-7 Worker Session、Run Lease 与启动协议基础 | Incubating(默认关闭,独立入口显式 opt-in | ADR-0012/0013/0014/0021/00570061/01080121/02310239/0377;有界 capability/Placement/DispatcherSQLite 协议孵化与 PostgreSQL v9 Session/Run Lease/credential/attestation authorityimmutable revision Placement、数据库时钟 keyset candidate、认证 Worker Pull、digest-only offer recoveryversioned capability-free ExecutionSpec response、stable claim 跨重启退避、单 owner 原子 inbox 准入与 TLS 1.3 mTLS/`ql3w` HTTPS client;同一 package journal 上 revision-fenced starting/spawn/started/running/completion 状态、callback digest、tagged no-spawn 与 ambiguous recoveryPostgreSQL starting/running/start-failure/completion 数据库权威事务、精确重放与 cancellation/timeout 优先终态;batch Secret delivery 在 Attempt advisory lock 下复验 Session/Lease/revision 完整围栏并复用单 AgentSecret-before-Artifact materializer 将同一 log ID 交给 Executor/journal/running ACKoffer-scoped `wlog-*` 私有文件 spool、Edge/Node 容量策略、append/quota/path 防护、barrier 后 output ownership、受审 POSIX Executor、truncation fact、固定内存流式 source、认证 Artifact stream、共享 immutable store port、S3-compatible SSE/checksum/条件 promotion adapter、upload-before-completion 协调,以及 Local/Cluster 同构、Profile-aware、ETag-fenced range read;用户取消 run.stop mutation 以数据库时间写 intent/Event 并在事务内复验 Project/RoleBinding fence;非执行取消 convergence lifecycle、运行期 expiry 与安全 lost retry 已接入 cluster-control 单一全局 cadence;完整 generation/version/token/Attempt fencing;独立最小权限 Worker ingress、CA/CRL 与连接 generation 热重载;offer journal、spawn barrier、receipt-first recovery;独立 `@qinglong/worker-runtime` 的本地 P-256 CSR、key/chain/trust 验证、generation + active pointer 安装和持久退避;默认关闭的 production process 已装配具体 execution graph、完整 Session heartbeat/drain/offline、direct-file bootstrap、单 Agent/单 cadence、startup reconciliation、证书 maintenance、transport fail-close/recovery 与 Edge/Node 有界预算;真实 PostgreSQL 18 + Linux Node 合约已覆盖 Run completion、credential 和 CA 双轮换且保持同一 Session;真实 K3s 合约已覆盖 TLS/credential Secret 分权、双对象 CAS、Recreate 顺序、identity generation 与单节点 PVC recovery;所有能力默认不可达且受 edge/cluster import audit 约束 | 具体 cert-manager/Vault/SPIFFE/离线 CA adapter 与模板、ingress reload controller、生产 RBAC、证书到期告警和 `ql3w` credential recovery 产品面;具体 KMS/Vault Secret provider、对象存储 credential/temporary lifecycle 与 retention/tombstoneWorker 管理 API;真实 Kubernetes 多节点 CSI/node-loss/production 360 秒 drain 与固定 edge 文件系统 suspend/时钟/断电、x64/arm64 资源门禁 |
| PR-8 Project/Policy/Approval Core | Incubating(默认拒绝、无生产业务执行入口) | ADR-0028;统一六类 ActorRef 与 exact-shape 校验;`0017` ownerless default Project 和 append-only versioned RoleBindingowner/admin/operator/viewer 固定矩阵;Project 内 mutation 幂等、expected-version CAS、双 SQLite 连接竞争门禁;archived read-only、revocation、存储损坏 fail-closedAgent 写/Secret/Tool `require_approval`ADR-0047 把六类 subject、role/permission matrix 与 fence 抽到 runtime-core`pg-0004-project-policy`/capability v3 建立 ownerless PostgreSQL baseline、严格 role/state CHECK、append-only runtime 权限、SERIALIZABLE Project lock、mutation replay、双连接单 winner 和 cluster admission authorizerADR-0049/`pg-0005` capability v4 建立 stable IdentitySubject、append-only digest-only API credential、真实 cluster bearer authenticator、write-only durable security audit 与最小权限 runtime role,且已验证 HTTP→credential→Policy→audit→handler 纵向链路;ADR-0051 建立 `/api/v3` 认证前 peer/global 双预算、transport-peer-only、无 timer 且有界内存的 overload shieldADR-0027 Artifact authorizer adapterADR-0029 `AuthenticatedPrincipal` contract、`0018` digest-only versioned challenge、CSPRNG/TTL、同事务消费 challenge + 写首 owner、精确重放与双连接竞争/崩溃回滚门禁;ADR-0030 `0019` stable identity/binding、legacy HS384 + current-session membership、logout/platform/revoke/disable、single-factor 与损坏 fail-closed 门禁;ADR-0031 `0020` digest-bound ApprovalRequest、User-only decision、Project/Role version fence、精确 expiry/重放/并发裁决及同事务 immutable dispatchADR-0032 `0021` execution backfill、三表原子 consume、稳定 due keyset、claim/renew/start/result fencing、pre-start takeover/post-start recovery-required、attempt budget、handler inspect/digest barrier 和 bounded dispatcherADR-0033/`0022` control/resolution backfill、start/renew/completion 原子联动、稳定 recovery keyset、双 resolver claim/takeover、finding/result 精确重放、自动/人工终结、迟到 completion 单 winner 和 evidence-only bounded reconcilerADR-0034/`0023` 首个 `run.create` canonical plan、Run/Attempt/Event/receipt 同事务、幂等 collision fail-closed、renew/终态 fence、真实 SQLite handler 与 automatic evidence providerADR-0035/`0024` 独立 `approval.recover` 矩阵、稳定 User + 五分钟强认证、Project/RoleBinding fence、human resolution + authorization fact 原子提交、撤权竞态与回滚门禁;ADR-0036 recovery-first 单 timer lifecycle、edge/standalone 独立 cadence/页预算、跨周期 cursor、非重叠与有界 stopADR-0074 以新的 Node 24 SQLite v5 ownerless Project/RoleBinding/audit authority 和独立 local-secret-admin 提供强 Principal、`secret.manage`、撤权 fence、envelope+allowed audit 原子提交及不回显语义;ADR-0086 以可信 POSIX console 和 staged delivery 完成本机首 Owner 产品 ceremony | fresh database/pepper setup 与安全迁移向导;`shareStore`/Express 到 authentication core 的 production migrationcredential rotation/revocation API、mTLS/Worker enrollment、恢复码;Project/Role/Approval/Secret 管理 CLI/API/UI、audit retention/query/export/alert、preview Artifact/digest/immutable plan builder、真实 MFA/hardware adapter、人工 recovery API/UI/独立 rate limit 与审计事件、handler/provider registry、lifecycle startup/shutdown/指标/admission gatePostgreSQL action/receipt/provider/recovery-authorization 与 OPA adapter、缓存 version 失效;Tool/Package/Secret/Shell 各自的 handler/evidence contractSecret/Run/Tool/Workflow waiting_approval 全入口装配;完整回滚演练 |