docs(audit): batch-46 — upload-lane quotas/lifecycle, tracker #8/#1, release-pipeline smoke

Real-machine closure for everything batched since batch 45: #75/#76 verified
end to end on auditfix77 (cooldowns 429, pending cap 403, 2GiB budget 403 via
the real accounting path, failure-does-not-burn-window, withdraw/admin-delete
row+blob double clean, panel two-step confirms in CDP), tracker #8 and #1
landed and verified (#77 panel redirect, #78 felis converge), plus #79 docs.
Also: pgint's new assertions first run on real PG (17/17) and the release
pipeline's first-ever run — tag v0.1.0-rc1, multiarch assets published and
executed on the target arch, /releases/latest deliberately untouched.

.dockerignore

@@ -31,3 +31,14 @@ Dockerfile
 *.key
 felis
 felis.exe
+
+# macOS materializes extended attributes as ._<name> sidecars (BSD tar uploads,
+# Finder copies, network volumes) and leaves .DS_Store behind. Neither is
+# source, and one is actively harmful: a ._*.sql beside the migrations is
+# //go:embed-ed into the binary and makes every `felis migrate` fail
+# ("non-numeric version") — observed live on a Mac-staged tree. Same exposure
+# for any tree the other //go:embed patterns walk (deploy/, plugins/).
+._*
+**/._*
+.DS_Store
+**/.DS_Store

.github/workflows/ci.yml

@@ -98,3 +98,42 @@ jobs:
 
       - run: npm run typecheck
         working-directory: panel
+
+  plugins:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+
+      # The other jobs never touch the Java layer: the plugin jars were only ever
+      # compiled by bootstrap on a live host, and the three test mains under
+      # plugins/*/test were run by hand. JDK 21 plus the Gradle major the plugin
+      # Dockerfiles pin (8.14) is that same toolchain, in CI.
+      - uses: actions/setup-java@v4
+        with:
+          distribution: temurin
+          java-version: '21'
+
+      - uses: gradle/actions/setup-gradle@v4
+        with:
+          gradle-version: '8.14'
+
+      - run: bash plugins/test.sh
+
+  mods:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+
+      # The three loader mods (Minecraft 1.20.1 / 1.20.4, Java-17 lines) compile
+      # through their vendored Gradle wrappers, which fetch their own Gradle. Until
+      # this job nothing ever built them: no install path touches them, and their
+      # gradlew scripts were committed without the exec bit, so the README's
+      # one-liners failed on a fresh clone.
+      - uses: actions/setup-java@v4
+        with:
+          distribution: temurin
+          java-version: '17'
+
+      - uses: gradle/actions/setup-gradle@v4
+
+      - run: bash plugins/test-mods.sh

.github/workflows/release.yml

@@ -40,6 +40,31 @@ jobs:
       - run: go vet ./...
       - run: go test ./...
 
+      # The same reason, for the Java layer the binary EMBEDS: the release asset is
+      # the tree's plugin sources (bootstrap_asset.go), and a tag whose plugins don't
+      # compile turns every install of that release into a failed bootstrap. JDK 21
+      # gates the install-time plugins + codec/invite tests; JDK 17 gates the loader
+      # mods (their vendored wrappers fetch their own Gradle).
+      - uses: actions/setup-java@v4
+        with:
+          distribution: temurin
+          java-version: '21'
+
+      - uses: gradle/actions/setup-gradle@v4
+        with:
+          gradle-version: '8.14'
+
+      - run: bash plugins/test.sh
+
+      - uses: actions/setup-java@v4
+        with:
+          distribution: temurin
+          java-version: '17'
+
+      - uses: gradle/actions/setup-gradle@v4
+
+      - run: bash plugins/test-mods.sh
+
       # Both architectures, because bootstrap's default release channel DOWNLOADS these
       # rather than compiling on the target host — an arm64 host with no asset silently
       # falls back to a slow source build. Neither stage is emulated: the Dockerfile pins

AUDIT-2026-09-22.md

@@ -1,6 +1,6 @@
 # Felis 生产就绪审计 — 2026-09-22（真机 E2E + 混沌）
 
-分支：`audit-fixes-20260922`（已推送）。环境：CentOS Stream 9 / aarch64 / k3s v1.36.4，
+分支：全部修复**逐 commit 直接推 `main`**（不再用功能分支）。remote = `[email protected]:FelisMC/Felis.git`（2026-09-23 起；旧 `MliroLirrorsIngenuity/Felis` 仅靠 GitHub 301 兜住）。环境：CentOS Stream 9 / aarch64 / k3s v1.36.4，
 IPv6-only 接入（`ssh -6 -i ~/.ssh/id_ed25519 root@fdb2:2c26:f4e4:0:21c:42ff:fede:69ec`），
 面板经 `socat TCP6:443 → 127.0.0.1:30443` 中继（手动启动，重启 VM 后需重开）。
 
@@ -17,21 +17,120 @@ IPv6-only 接入（`ssh -6 -i ~/.ssh/id_ed25519 root@fdb2:2c26:f4e4:0:21c:42ff:f
 | 17 | **`ListPendingOpLogins` PG 少列**：接口注释承诺「joined to its staff username」，handler 输出 `username`/`created_at`，fake 正确填充；PG SQL 未 JOIN 也未取 `created_at` → 真机 pending 列表 username 为空、created_at 为 `0001-01-01` | 真机 internal `/op-login/pending` 响应 | `pgrepo.go` SQL 改为 JOIN users + 取 created_at。提交见分支 |
 | 18 | **绑定码并发兑换 500**：`RedeemPlayerBindCode` 无 `FOR UPDATE`（同文件 `VerifyLinkCode` 有），且裸 INSERT。6 路并发同码兑换 → 3×HTTP 500（`users_username_key` 唯一冲突）+ 1×400 + 2×200；跨码并发同 UUID 同样会撞 | 真机四组并发测试 + API 日志 `unmapped error ... duplicate key` | 同码：码行 `FOR UPDATE`（输家干净地 400 invalid_code）；跨码：`INSERT users ... ON CONFLICT (username) DO NOTHING`+重读、`account_links ON CONFLICT (mc_uuid) DO NOTHING`（两路汇聚同一 user）。复测：同码×6=1×200+5×400、双码×2=2×200 同 ID、DB 干净、日志零 unmapped |
 | 19 | **reaper CronJob 渲染位置错误，永远无法调度**：`felis manifests` 把 CronJob 渲染到控制 ns，却引用 minecraft ns 的备份 PVC（Pod 不能跨 ns 挂 PVC：真机 `FailedScheduling: persistentvolumeclaim "felis-backups" not found`）；修正 ns 后又发现 ServerAccount 也不能跨 ns 使用（`serviceaccount "felis-reaper" not found`）。而 reaper 的 Role/RoleBinding 本就在 minecraft ns | 真机三层取证（PVC/SA/调度）| CronJob 与 SA、RoleBinding subject 全部移到 `MinecraftNamespace`（提交 `e4f2cff`+`c839454`）。**修复后完整演练通过**（见下） |
+| 6 | **默认安装无备份能力 + 回收链路不可达/不可用**：(a) 没有任何环节渲染归档 PVC，`FELIS_BACKUP_PVC` 永远为空 → backup/restore 恒 503；(b) bootstrap 从不下传 reaper 三旗标 → 官方安装路径根本无法启用回收；(c) 即便启用，stock k3s 的 `<root>/<pvc>` 布局不存在，且 k3s storage root 是 `0700 root:root`、reaper pod 以 uid 1000 运行 → 真机 `lstat /worlds/…: permission denied`（fail-closed 跳过，但纯空转）；(d) README 口径（“定时备份”）不实 | 真机全链路：默认渲染 PVC+env → 建服→marker→备份 202→jobs 端点 running→succeeded→改 marker→restore→读回原值 ✅；再建 resolvecheck 世界（20d idle，marker）→ CronJob 手动 Job：`evaluated=2 reaped=1`，归档含 marker、PVC+宿主目录回收、`world_backups` 得 `inactive_15d` 行、servers 行/CR 保留 ✅ | `platform/workloads.go` 渲染归档 PVC（minecraft ns、RWO 10Gi、默认 SC），`felis manifests --backup-pvc` 默认 `felis-backups`（`=` 空为显式关闭），bootstrap 统一下传 env/旗标并写 [archive] local_path；`resolveWorldDir` 新增精确 local-path 目录解析（读 PVC `spec.volumeName`，非 glob，杜绝陈旧 PV 目录顶替）；ReaperRole 增 `pvc:get`；bootstrap 设 `FELIS_WORLDS_HOST_PATH` 时给 uid 1000 授 traverse（setfacl/o+x）；README/故障手册改写。提交 `fd33fd0`+`2b87a5a` |
+| 8 | **磁盘打满灾难链**：DiskPressure → kubelet 驱逐控制面（无 PriorityClass 保护）→ 镜像被 GC（无外网、registry 空）→ 全部 ImagePullBackOff；释放后数分钟才恢复调度。恢复依赖人工 `docker save | k3s ctr images import -` | 三轮真机 drill（原缺陷复现 + 两轮修复验证）：①自定义 1e6 类：游戏 pod 先走、控制面随后仍被驱逐（kubelet 日志逐条列出 ranked/evicted），随后镜像被 GC → ErrImagePull；②内建 `system-cluster-critical`（2e9）：填盘至 1.7G free，kubelet 对 felis-api/operator/registry 全部报 **“cannot evict a critical pod”**，三者在整个 DiskPressure 期间保持 Running；游戏 pod（0）被驱逐；③释放磁盘后：`DiskPressure` 经 ~5 分钟（`eviction-pressure-transition-period`）转 False——即“恢复慢”的主因；被 GC 的游戏镜像按 runbook 重导入后 25s 恢复 ✅ | 控制面四类 pod 挂内建 `system-cluster-critical`（用户自定义类值上限 1e9，达不到 2e9 临界阈值；preemption 保留=管理面可调度，已记录权衡）；troubleshooting §13b 固化事件链、5 分钟条件过渡与镜像恢复路径。提交见下 |
+| 9 | **升级策略 Recreate**：单副本 + Recreate：任何控制面升级=停机；坏升级（错 tag）中断约 95s 且需人工 `rollout undo`（无自动回滚） | 复核部署模板（Recreate + 单副本 + 无 leader election 的注释理由成立） | 不改策略（Recreate 是对无 leader election 的正确取舍），改为固化 runbook：troubleshooting §15 = 升级即重跑安装器；停机窗口=rollout 时长；坏镜像在安装器 180s 等待内以 `kubectl describe` 诊断呈现；回滚 `kubectl rollout undo`（镜像被 GC 时先按 §13b 重导入）。提交见下 |
+| 10 | **备份语义**：归档包含整个 /data（jar、libraries、cache），167MB；是否符合 world backup 定位待评估 | 真机 tar 清单（cache/、config/、eula.txt、server.properties、jar）+ 恢复语义（overlay+prune） | 评估结论：**整卷备份是正确语义**（回滚=整服状态回滚，含配置/插件），保留行为；文档化：troubleshooting §10「backup contains」+ OpenAPI/README 口径改为“整服数据卷”。提交见下 |
+| 20 | **验证邮箱唯一性只存在于注释里**：errors.go/repo.go 都声称迁移 0010 建了 `users_verified_email_unique` 部分唯一索引、`VerifyEmailOTP` 会返回 `ErrEmailTaken`；实际上**索引从未创建**、`ErrEmailTaken` 全仓库从未被返回 → 两个账号可同时验证一个邮箱，而登录门正是按 verified email 解析账号 → 同一邮箱的登录码归谁由数据库任意决定 | pgint 首跑即红：第二账号验证同邮箱返回 `nil`；`grep users_verified_email_unique internal/store/migrations/*.sql` 零命中。真机复验（auditfix17 邮箱 409 演练）：同码二次 verify 仍 409（码未被消费） | 新增迁移 `0020_verified_email_unique.sql`（`lower(email) WHERE email_verified`）；`VerifyEmailOTP` 在消费码**前**查「他人已验证」→ `ErrEmailTaken`（不消费码、不扣次数），并把并发唯一冲突映射为同一答案；handler 新增 409 `email_taken`。提交 `b6ef27c`；PG 契约测试基建 `2a55a0d`（`-tags pgint`，见 CONTRIBUTING） |
+| 21 | **admin 改邮箱不清 verified**：`UpdateUser` 写入新地址但保留 `email_verified=true` → 面板改错一个字符就能让登录码寄到别人邮箱（该 flag 正是邮箱登录解析/寄码的依据）；fake 同错 | pgint 契约测试 | 改地址时同写 `email_verified = email_verified AND email IS NOT DISTINCT FROM 新值`（同值 no-op 保留证明；换值即清除）；fake 同步。提交 `d1ec40f`。真机验证：PATCH→`f`、PATCH 回原值仍 `f`、玩家重验证→`t` |
+| 22 | **`role='owner'` 是死信**：迁移 0011 加了 owner 角色并把全部用户管理路由压在 `IsOwner()` 上，但**没有任何代码写过 `owner`**——breakGlass(`UpsertOwner`)、setup MC 绑定(`CompleteOwnerSetup`)、重置路径一律写 `admin` → 全新安装的整个 owner 层（用户列表/创建/编辑/禁用/删除/配额/会话）不可达；且新角色没进各处 staff 判定（op-login 只认 admin、玩家邮箱门只拒 admin、reclaim 保护与 AdminExists 只认 admin）；面板一旦有 owner 行还可被降级/删除/禁用 | 真机：owner 会话加载 `/api/v1/users` 403；promote 后 200。op-login：修复前 owner start 中立不发码，修复后铸请求+finish 得 `role=owner` 会话 | `UpsertOwner`/`CompleteOwnerSetup` 改写 `role='owner'`（冲突臂重断言，即 0011 文档的 promote 路径）；op-login 双端改 `staffRole`；玩家邮箱门改 `role != 'user'`；`IsProtectedAdminLink`/`AdminExists` 计入 owner；面板新增守卫：owner 行不可降级/删除/禁用（用户名/邮箱编辑仍可）。提交 `e0d2378` |
+| 23 | **配额门非原子 + storage 缓存被清零**：(a) audit #4：`QuotaCheck` 与 `ClaimServer` 两条语句，同一用户并发认领两台无主服可双双通过 `max_servers`（deferred-seams 曾把这条挂为“只能在真 PG 上关闭”）；(b) 更隐蔽：PATCH 资源时 `UpdateServerResources(..., 0)` 把本不能改的 storage 缓存写 0，而缓存列是配额聚合的**唯一**输入 → 此后该服的 storage 维度在配额里凭空消失 | (a) 新增 pgint 并发测试：修复前两台全赢；修复后恰 1 赢 + 1 `ErrQuotaExceeded`，DB 只 1 行 owned；(b) hermetic 测试 `TestPatchServerPreservesStorageCache`（修复前 `resourceUpdates` 里 storage=0） | (a) 门槛进 `ClaimServer`：同一事务内 `pg_advisory_xact_lock(hashtext(user_id))` + 四维重查（与 `QuotaCheck` 共用 `quotaAllows` 防漂移）+ 行 `FOR UPDATE`，两个 claim handler 把 `ErrQuotaExceeded` 映射为与串行一致的 403；(b) resize 前读取现值并透传 storage。提交 `bb68fef`；deferred-seams 对应条目核销 |
+| 24 | **reaper 警告信从不真正投递**：`felis reaper` 从不装配任何 Warner（`r.Warner` 恒 nil），而 `maybeWarn` 对 nil warner / 投递失败一律照样 `MarkWarned` + `warned++` → 每位有主的服都在**无人收到提醒**的情况下 15 天后被静默回收，跑批日志还谎报"已警告 N 台"。红线⑤的"best-effort 不阻塞回收"被误读成了"失败也要记成已通知" | hermetic 单测 3 例（坏 notifier / nil warner → 不盖章、成功才盖章）。真机 drill（auditfix20）：13d idle 实收 1 封（收件人=owner 的验证邮箱、subject 含 3d）；重跑不重发；14.5d 补发 1d（elif 档位）；SMTP 端口打坏 → 日志 `warn delivery failed; will retry next run` 且**不盖章**，恢复后补发；邮箱未验证 → `has no verified email` 不盖章；owner NULL → 静默跳过；边界 11d23h 不发 / 12d1m 发；全程 `world_backups` 保持 4 条不动（纯警告零备份副作用） | `mail.SendNotice` + `mailWarner`：查 owner 的 verified email → `SendNotice`；nil/失败不盖章、下次重试；reaper pod 模板加 optional `FELIS_SMTP_PASSWORD` env；`felis setup` 复制 felis-smtp 镜像并在「configure email」刷新 minecraft ns 的 felis-smtp+felis-config 镜像；docs/troubleshooting §10 更新。提交 `8e7c7bb` |
+| 25 | **idle auto-stop 从未触发（三层复合缺陷）**：条件齐备的空载服永远不停。① CRD status schema 未声明 `emptySince`，apiserver **pruning** 掉计时戳（`unknown field "status.emptySince"`），计时器每次读回都是 nil；② 即使戳幸存，Running 空载稳态**没有任何 watch 事件**（玩家进出不碰 CRD、RCON 只在 Reconcile 内探），盖章一次后 reconcile 链停摆，无人叫醒；③ auto-stop 用整对象 `Update` 写 spec 且 OperatorRole 从未有 `minecraftservers:patch/update` → 403 `cannot update resource`。三层任一都让功能永久失效，而单测（fake client 不剪 schema、手动驱动、无 RBAC）全部覆盖不到 | 真机逐层实锤：schema 修复后 `empty=2026-09-22T20:02:13Z` 首次成功持久化；静置 88s+ 无动作、operator 日志 90s 零行（②实锤）；修复前日志 5 条 403（③实锤）；全修后场景 1：超时戳触发即 Stopped；场景 2：起服 → 20:11:43 盖章 → **全程无干预** → 20:12:17 自动 Stopped；翻转 Running→Stopped→Running 收敛 Running；`idle=null` 后不再自停 | ① schema 补 `emptySince`（`c04a3f0`）；② `reconcileRunning` 尾部返回 RequeueAfter——空载=到点精确唤醒、有人=30s 探针周期，+3 个单测（`1c89a5e`）；③ auto-stop 改 merge-patch（防 status clobber，与 reaper 的 Stop 同型）+ OperatorRole 补 `patch` + rbac 测试锚点（`f650bf8`）。真机 auditfix22 全通；troubleshooting §11 增自检三连 |
+| 26 | **RCON Secret 被删 → 服务器永久锁死**：删掉 per-server RCON 密码 Secret 后，(a) operator 没有任何 watch 能看到删除（Running 稳态零事件），secret 一直不重建；(b) 一旦被任意事件带到 reconcile：新密码铸造成功，但运行中的 pod 仍持有旧密码、sts 模板无变化 → pod 永不重启 → 探针用新密码连旧密码 pod，**永久 `RconNotReachable`** 直至 300s `ReadinessTimeout`；恢复只有人工删 pod。`ensureRconSecret` 的注释却声称"heals on the next pass" | 真机：删 secret 后静置 60s 无重建、无日志；annotate 触发后 96s+ 持续 RconNotReachable（新密码 `91dd62…` vs pod 旧密码）；删 pod 后 17s 恢复 Running——根因=密码漂移实锤 | ① `SetupWithManager` 增 `Owns(&corev1.Secret{})`（controller-owned，删除事件映射回 CR）——真机日志见 `source="kind source: *v1.Secret"`；② pod template 新增 `RconSecretAnnotation` = 当前密码 SHA-256 指纹（64-bit）：secret 重建指纹变 → sts 自动 rollout 用上新密码，未重建则恒定不抖动；测试 2 例（跨 reconcile 稳定 / 重建必变）。提交 `56f3abd`。真机复测：删 secret 后**同秒**察觉+重建+换戳，31s 全自动恢复 Running |
+| 27 | **陈旧启动锚点 → 已恢复的服务器被误判 StartupTimeout + Provisioned 永久 False**：`status.startRequestedAt` 只在 Stop 时清，成功（markRunningReady）不清——一旦服务器曾经历一次长 Starting，其 300s 预算就悬在健康运行中；此后任意一次 pod 波动（rollout/崩溃）都会立即套用旧戳判 `Failed(StartupTimeout)`。且 `markFailed` 写入的 `ConditionProvisioned=False` 无人复位，恢复后仍永久挂着失败标记 | 真机：20:19 已恢复 Running 的 test-one，20:21 因一次 stamp 注入引发的 pod rollout 被标记 `Failed StartupTimeout`（锚点残留自 20:16）；恢复后 status.conditions 里 `Provisioned=False … StartupTimeout` 持续存在 | `markRunningReady`：清 `StartRequestedAt`（每次启动/恢复尝试各有独立预算）+ 复位 `ConditionProvisioned=True`；测试 2 例（Ready 清锚点、Failed→Running 后 Provisioned 恢复）。提交 `82b5a60`。真机复测：清戳生效（`startRequestedAt: None`）、pod blip 删→33s 恢复全程无 Failed、锚点重新盖章后成功清除 |
+| 28 | **cfsetup 把 "policy already exists" 当成功吞掉 → fail-closed 保证可被旧策略顶替**：Cloudflare Access 的 `CreateAccessPolicy` 在收到 already-exists 时直接返回 nil；若该 app 上已有一条更宽松的旧策略（改 identity 后重跑、或此前手工配置），守卫 op.console 的仍是旧策略，而 Setup 报告成功——`validateFailClosed` 只校验过"我们构建的策略"，从未校验证留在线上的那条 | 代码审查（集成侧无真实 CF 账号，无法真机）：httptest 3 例复刻——已存在同名策略、缺席、POST 竞态。修复前第 1 例吞错返回 nil 且不发 PUT | 改为按名 upsert：lookup → PUT 覆盖守卫体 → 缺席才 POST（POST 撞 already-exists → 重查后 PUT，绝不吞）。`apiPost/apiPut` 共用一个 `apiWrite`。提交 `30857df` |
+| 29 | **"configure email" 的工作负载镜像复制从不生效**：`replicateSMTPToWorkloadNamespace` 把硬编码 `namespace: felis` 的 felis-smtp manifest 用 `-n minecraft apply` 发出——kubectl 拒绝 namespace 冲突（`the namespace from the provided object ... does not match`）→ 第一段直接 return err，**第二段 felis-config 的镜像复制根本不会执行**。于是任何"装好后再改 SMTP"的部署，reaper 的 pre-reap 警告永远拿不到新配置（这正是 8e7c7bb 添加该刷新要解决的事），且失败仅 warning 不中止 | 真机 kubectl 行为实验：`-n minecraft apply` 带 `namespace: felis` 的 manifest → `error: … does not match …`；修复后（manifest namespace=minecraft）→ `accepted-namespace=minecraft`（server dry-run）；felis-config 渲染+apply 序列同为 accepted | `smtpSecretManifest(password, namespace)` 显式参数（控制面调用传 "felis"，镜像传工作负载 ns）；回归测试 `TestSMTPSecretManifestCarriesTargetNamespace`。提交 `ed722d5` |
 
 ## 待决策台账（未修）
 
-| # | 主题 | 说明 |
+前两日台账的 #7、#11–#15 已全部修复（本批核销，证据见下），不再挂在"未修"里：
+
+| # | 主题 | 状态 |
 |---|------|------|
-| 6 | 默认安装无备份能力 | backup/restore 端点默认 503（需 `FELIS_BACKUP_PVC`+PVC），reaper CronJob 需 `--backup-pvc/--archive-local-path/--worlds-host-path` 三旗标渲染，bootstrap 一个都不传；文档无说明；README 与"自动备份"口径不符。另：归档 3 个月过期依赖 reaper 清理，未启用则磁盘只增不减 |
+| 7 | 异步失败不可感知 | ✅ 修复：`GET /api/v1/servers/{name}/jobs`（`ff7c57c`，含 RBAC `jobs:list` 与 OpenAPI）；真机 drill 用它看到 running→succeeded |
-| 7 | 异步失败不可感知 | backup/restore 失败后无状态出口：restore 行不变、backup 无行；只有集群侧 Job/日志可查。建议状态字段或 `?failed` 查询 |
+| 11 | PG 断连报 401 | ✅ 修复：会话存储故障改为 503（`2a8f897`） |
-| 8 | 磁盘打满灾难链 | DiskPressure → kubelet 驱逐控制面（无 PriorityClass 保护）→ 镜像被 GC（无外网、registry 空）→ 全部 ImagePullBackOff；释放后约 8 分钟才恢复调度。恢复靠 `docker save felis:* | k3s ctr images import -`（docker 守护进程存储是唯一副本，需固化回源路径）。建议：PriorityClass、镜像入内置 registry、磁盘告警 |
+| 12 | ready 门滞后 | ✅ 修复：operator 启动期内每 2s 重探 RCON（`a2df2f2`） |
-| 9 | 升级策略 Recreate | 单副本 + Recreate：任何控制面升级=停机；坏升级（实测错 tag）服务中断约 95s 且需人工 `rollout undo`（无自动回滚）。建议 runbook/文档化 |
+| 13 | setup URL 截断 | ✅ 修复：TUI 折行（`abce381`） |
-| 10 | 备份语义 | 归档包含整个 /data（jar、libraries、cache），167MB；是否符合"world backup"定位待评估 |
+| 14 | controller-runtime 日志噪音 | ✅ 修复：SetLogger 接 slog（`a415246`） |
-| 11 | PG 断连表现 | 会话查询失败报 401 而非 503（fail-closed 但误导；用户以为没登录） |
+| 15 | reaper 启用未演练 | ✅ 已演练（见"第二日"节；CronJob 仍 `suspend=true` 防误删） |
-| 12 | ready 门滞后 | 容器 Ready 后 6~10s 内 API 仍 409 not_running |
+
-| 13 | setup token 截断 | 43 字符 token + 长域名，80 列终端下 TUI 截断显示（复现：tmux 80 列） |
+## 可达性分级（#1–#79；#62 立案后剔除）——这些缺陷真实使用中到底谁能踩到
-| 14 | 日志噪音 | controller-runtime 未 SetLogger，首用打印整段堆栈；TLS handshake EOF 噪音（kubelet 探针） |
+
-| 15 | reaper 启用未演练 | **已演练完毕（2026-09-22）**：绑定挂载演练 world → 一次 pass 归档+删 PVC+保留 servers 行/CRD；过期归档驱逐（行转 deleted、文件删除、合法归档未动）。启用仍受 #6 三旗标约束；另注意 `--worlds-host-path` 的 `<path>/<pvc>` 布局在 stock local-path 下不成立（编排责任），且 VM 上演练用的 CronJob 已 **suspend** 防误删（真实世界目录不在 /srv/worlds-root）|
+回应质疑"是不是全在测边界条件 / 只有内部 hook 才能触发"：
+
+- **hook 的边界**：演练中使用的内部手段（读服务端日志取 OTP 码、service token 打内部面代 approve、本地 SMTP sink、tmux 驱动 TUI）只是**测试仪表**——代替"真实收邮件 / 键盘敲击 / 操作台点击"；触发路径本身是公开路径，真实用户做同一动作走同一段代码。真正**本环境不可复现**的只有 #28（无真实 CF 账号，仅 httptest 复刻），另有 #4/#5/#14 属工具/卫生级（无运行时触发面）。
+- **分级口径**：① 日常＝普通使用或默认安装下即会踩到；② 运维＝升级/重启/故障恢复/磁盘满/删资源/闲置回收等真实运维动作会踩到；③ 窗口＝需要并发或特定状态时序，真实但概率低；④ 工具＝仅工具判定/审查；决策＝非缺陷（评估/演练项）。
+
+| # | 真实触发路径（一句话） | 分级 |
+|---|------------------------|------|
+| 1 | 恢复失败后 10 分钟内重试（必然的恢复动作）→ 202 但 Job 不动 | ② |
+| 2 | 默认安装下第一次点「备份」→ Job FailedMount（felis-config 缺在 minecraft ns） | ① |
+| 3 | #1 修复所需 RBAC（单独不产生用户可见行为） | ② |
+| 4 | gofmt/staticcheck 漂移 + CI 门禁缺失 | ④ |
+| 5 | 依赖漏洞（govulncheck 判定可达）：需特定输入 | ④ |
+| 6 | 默认安装即踩：备份/恢复恒 503、回收链无旗标不可启用 | ① |
+| 7 | 任何异步操作（备份/构建）想查结果：无端点（功能缺口） | ① |
+| 8 | 磁盘打满事故链（控制面被逐 → 镜像 GC → 全体 ImagePullBackOff） | ② |
+| 9 | 非缺陷：升级策略评估 → runbook | 决策 |
+| 10 | 非缺陷：备份语义评估 → 文档化 | 决策 |
+| 11 | PG 掉线期间任何会话校验 → 401 误导 | ② |
+| 12 | 每次起服：就绪感知滞后 | ① |
+| 13 | 首次安装向导：长 URL 截断 | ① |
+| 14 | operator 日志噪音（卫生项） | ④ |
+| 15 | 流程项：reaper 启用补演练 | 决策 |
+| 16 | 邮箱登录主路径：错码/重放 → 500、5 次锁定失效 | ① |
+| 17 | 面板 op-login pending 列表字段缺失 | ① |
+| 18 | 绑定码并发兑换（双击/双端即可）→ 部分 500 | ③ |
+| 19 | 按文档启用回收 → CronJob 永远无法调度 | ② |
+| 20 | 两账号验证同一邮箱 → 登录码归属不定（安全） | ③ |
+| 21 | 面板改任一用户邮箱 → verified 未清、码寄旧地址 | ① |
+| 22 | 全新安装 owner 层不可达（用户管理全路由） | ① |
+| 23 | (a) 并发认领两服可双双过配额；(b) PATCH 任意服资源 → storage 缓存清零 | ①（a=③） |
+| 24 | 有主服闲置 12–15d：警告信从不投递、静默回收 | ② |
+| 25 | 启用空载停机后等超时：三层复合缺陷永不生效 | ② |
+| 26 | 运维删/轮换 RCON Secret → 永久锁死至人工删 pod | ② |
+| 27 | 曾长启动的服 + 任意 pod 波动 → 误判 StartupTimeout | ② |
+| 28 | 重跑 cfsetup 且线上有旧策略 → fail-closed 被顶替；⚠无真实 CF 账号，仅 httptest | ④ |
+| 29 | 装后改 SMTP（configure email）→ 复制链从未生效 | ② |
+| 30 | 对不存在/刚被删的用户 id 操作 → 500 而非 404 | ③ |
+| 31 | 真实 owner 打开 /admin/builds 被 mock 残留误判 | ① |
+| 32 | 新浏览器首访 /admin/builds 播种两个假 404 | ①（轻） |
+| 33 | 禁用/已删账号重走邮箱登录门 → 可复活（安全） | ① |
+| 34 | 同族：死账号游戏内身份面仍存活 | ① |
+| 35 | 任何真实服务器保存过一次后：所有备份/fileread 必败 | ① |
+| 36 | 任何构建失败的提交者永远看不到结果 | ① |
+| 37 | 管理端 fleet：login/lobby 行全是死操作 | ①（轻） |
+| 38 | felis manifests 输出残留过期指导 | ②（文档） |
+| 39 | 断玻璃 reset owner 用非在位名 → 铸第二 owner 永不可清 | ② |
+| 40 | 断玻璃 add operator 撞名 → 裸 SQLSTATE | ② |
+| 41 | 断玻璃 Sync 选系统服 → 裸内部错误 | ② |
+| 42 | 对无世界盘服备份/恢复 → 202 后静默卡 30 分钟 | ② |
+| 43 | #42 配套：409 一刀切文案 | ②（轻） |
+| 44 | 重跑 `felis setup` 完成 s/c reconfigure → 状态框退回“Setup complete”（信息一致性） | ①（轻） |
+| 45 | 评审看不到将被执行的 recipe（执行的是上传 blob 里的 Dockerfile）→ 盲批 | ① |
+| 46 | 大层推送打死 registry（256Mi 模板被 OOM kill；实测 475MB 层）——用户构建/安装器入仓即触发 | ① |
+| 47 | Mac 打包树构建：`._*.sql` 旁文件混入 embed → `felis migrate` 全挂 | ② |
+| 48 | 安装器重跑：每镜像 start/stop docker 触发 systemd 限流 → 批量入仓中途断 | ② |
+| 49 | 安装器重跑：`[registry]`/`[archive]` 运维配置静默回退（构建/S3/reaper 行为回默认） | ② |
+| 50 | 配过 SMTP 的安装按文档重跑安装器 → 生成的注释块被 `[smtp]` carry 吞并并每次 +1（纯膨胀） | ② |
+| 51 | 改配置/轮换 DB 凭据后：工作负载 ns 的 `felis-config` 副本永不刷新 → backup/reaper 静默读旧配置 | ② |
+| 52 | 向导内 `s`/`c` 重配置后：工作负载 ns 的 `felis-config` 镜像停在上一次运行快照，直到下次 `felis setup`/安装器才收敛 | ② |
+| 53 | 默认安装或 `felis update` 的 felis-api 检查：坐标仍指向已迁移的旧仓库，现仅靠 GitHub 301 兜住——redirect 一退休，安装器默认 URL 与更新检查全挂 | ② |
+| 54 | 完成态安装上照 `felis update` 指引升级：`sudo felis setup` 只开配置控制台，任何组件都不会动（三处 run 指引 + trailer 全假） | ②（文档） |
+| 55 | 配置的源回 200 但档案字段不可用（马虎自建/三方 Yggdrasil）：坏源在前 → 经梯子的登录被静默吞掉（零日志、后续源不被询问） | ② |
+| 56 | 给"还没进过服"的玩家加白名单/封禁/踢出（面板说成功、vanilla 实际拒绝）——任何新玩家第一次被管理就撞 | ① |
+| 57 | 控制台发任何命令后无反馈（回复被丢弃、pod 日志也不回显） | ① |
+| 58 | 对从未启动过的服务器点"文件"页（90s 卡死 → 误导性 504） | ① |
+| 59 | 装了 LuckPerms 的服打开 LP 页：读不到被报成"未分配任何权限"（假事实），且历史伪造 `[RCON]` 输出 | ① |
+| 60 | 新建服务器输入非法名/子域名（前端预检比后端弱）→ 对话框直出 Go 英文错误；账户验证撞已占邮箱同理 | ① |
+| 61 | 给"还没进过服"的玩家预授权限：面板提示"仍然会实际生效"（过度承诺），实际 LuckPerms 静默丢弃（未进服的名字无法解析） | ① |
+
+统计：① 26 ｜ ② 25 ｜ ③ 3＋#23(a) ｜ ④ 4 ｜ 决策 3 ＝ 61。第三批构建链另有 3 处未编号修复（Job requests 超小节点上限 → 永远 Pending、Kaniko chown、Trivy DB egress 被锁）——均属 ①/②。
+
+第四十批追加：#63（demo-up 单起点化——旧镜像臂能产出"起了但无处路由"的演示机；dev/demo 路径）②；#62 经立案复核**不成立**（全部安装路径都构建 felis-velocity.jar，证据见该批节），已剔除、不计入。
+
+第四十一批追加：#64（三个 vendored `gradlew` 以 100644 提交、无可执行位——README 教的构建命令在全新 clone 上直接 `Permission denied`；无任何 CI/安装路径跑过这三个模块）①；#65（三个装载器 mod 的 `license` 仍写 MIT、Forge/NeoForge 的 `issueTrackerURL` 是 `example.invalid` 占位，与仓库 AGPL-3.0-only 相悖——fabric loader 启动时会打印该字段）②（低）。
+
+第四十二批追加：#66（英文 README 缺中文版"使用方式"里的私有仓库安装 workaround 与重跑升级说明——英文读者照文档第一步即 404、无任何指引）①（轻）。
+
+第四十三批追加：#67（build Job 从不设 `ttlSecondsAfterFinished`——每构建一次就永久留下一个完成 Job+Pod，完成 Pod 计入节点 pod 预算（stock k3s 110），构建量上来后新构建全 Pending）②（时间维度；随构建量从②滑向①）。
+
+第四十三批追加（二）：#68（构建 context 拉取单次尝试——控制面滚动/重启/驱逐恰好撞上构建窗口时，fetch initContainer 一次 `connection refused` 直接打成终态 `Failed`；`BackoffLimit=0` 无第二次 Pod，代价 = 人工重审重提）②（运维：升级/重启/故障恢复撞上正在进行的构建；构建量大时概率上升）。
+
+第四十五批追加：#69（README 中英"审批通过后自动构建**并部署**"过度承诺——数据模型无目标服务器、部署实为"选用该镜像"；① 轻）；#70（kaniko 拉内建底座默认 HTTPS——`--insecure` 只覆盖推，任何 `FROM registry.felis.svc:5000/…` 构建必败 ①）；#71（kaniko 以 drop-ALL 解包底座层，chown 必败——任何非 scratch 底座必败 ①）；#72（trivy 扫 jar 必拉 Java DB、被 egress 锁拒绝——含 jar 即所有真实模组包的构建必败 ①）；#73（bootstrap 测试在跑 k3s 的主机上必假失败 ④ 工具）；#74（console 断连测试读写竞态 flaky ④ 工具）。
+
+第四十六批追加：#75（提交上传面无上限——pending 无个数上限、无存储预算、create/upload 无节流；一个账号可无限堆积上下文刷爆 uploads PVC ①）；#76（提交无撤回/管理员删除路径——提交者无法自救、运维无法回收占用 ①）；#77（未完成引导的会话触发受保护操作 → 面板显示"无权执行此操作"而非送往 /setup；tracker #8 ①）；#78（create-if-absent 使新增 CR 字段在已装机永不落地——lobby 无 RCON 故控制台死、玩家数恒 0；tracker #1 ②）；#79（troubleshooting [INERT] 图例指向已不存在字段 ④ 文档）。
 
 ## 已验证事实（正向清单）
 
@@ -46,6 +145,807 @@ IPv6-only 接入（`ssh -6 -i ~/.ssh/id_ed25519 root@fdb2:2c26:f4e4:0:21c:42ff:f
 - op-login 全状态机（start→status→approve→finish；早 finish 不烧码、错码扣次且请求保留、重放/重复批准/非管理员批准/未知 handle 全部按契约返回）✅
 - 并发：OTP 风暴 8 路 = 1×202 + 7×429 且仅铸 1 码；绑定码并发（同码/双码）见 #18 ✅
 - **reaper 全链路真机演练**（修复 #19 后）：绑定挂载假世界 → `felis reaper` 一 pass：归档 tar 落盘（内容含标记文件）、`world_backups` 行 `inactive_15d`+90d 过期、**PVC 删除且宿主目录回收**、servers 行保留且 activity 时钟重置（红线②）、CRD 保留；第二 pass：伪造过期归档被驱逐（`expired=1`，文件删、行转 `deleted`），合法归档未动；`evaluated=2 reaped=1` 只动到期的世界 ✅
+- **S3 存储向导全链**（第三十二批）：错误凭据预检零副作用、正例落地（Secret + 两 toml + API 滚 + 状态行）、上传 → MinIO 桶 → 内部面取件、UI 回滚归位、控制面/工作负载镜像对齐 ✅
+- **NetworkPolicy 栅栏真机**（第三十五批）：pod 源"身份×端口"矩阵全对（允许行与拒绝行同源对照）、转发型外部源白名单闭环（拒→放→拒，ipset 与 spec 双幂等）、host/ClusterIP 路径全通、Mac→NodePort 面板 200；两条归因注记见该批 ✅
+
+### 本轮新增真机证据（第二日）
+
+- **默认安装的备份闭环**：apply 新 bundle → 归档 PVC 自动创建（WaitForFirstConsumer，被首个 backup Job 拉绑定）→ `POST /backup` 202 → `GET /jobs` running→succeeded → 漂移 marker → `restore-backup` → 读回备份前内容 ✅
+- **回收闭环（真实 k3s 布局 + 权限）**：resolvecheck（20d idle）→ CronJob 派 Job：`evaluated=2 reaped=1`；归档含 marker、PVC+宿主目录回收、`world_backups` 记 `inactive_15d`；期间先经历 `lstat /worlds/…: permission denied`（k3s storage root 0700）→ 修 ACL/root 权限后通过 ✅
+- **磁盘压力三连 drill**：①1e6 自定义类：游戏 pod 先走、控制面随后仍被逐（kubelet ranked/evicted 日志）→ 镜像 GC → ErrImagePull；②改内建 `system-cluster-critical` 后：`cannot evict a critical pod` × felis-api/operator/registry，全部保持 Running；③恢复阶段实测 `DiskPressure` 条件 ~5min（eviction-pressure-transition-period）转 False，被 GC 的游戏镜像按 runbook 重导入后 25s 恢复 ✅
+- **混沌**：SIGKILL k3s 主进程 → kubectl 与 felis-api ready 均在 **10s** 内恢复（api pod 本身未被重启）✅；内存压力（2.6G tmpfs 满写）：无 OOM kill、无人被逐（swap 5.9G 吸收 + 内核回收），控制面不受影响 ✅
+- **构建链路定位（当时未修）**：Kaniko/Trivy 默认镜像是外部的 → 已加 `[registry] kaniko_image/trivy_image/build_cpu_limit/build_mem_limit` 覆写；上下文跨 ns 不可挂载（PVC 不能跨 ns）是设计级缺口，s3 lane 也缺凭据注入 ✅（证据：`internal/submit/blobstore.go:40`、`cmd/felis/api.go` contextBase 分支、jobspec 无 volumes/env）→ **第三批已修复，见下**
+
+### 本轮新增真机证据（第三批：探针 + 构建链路全通）
+
+- **控制面探针（`0c8e29b`）**：felis-api / felis-operator / registry 三个 Deployment 此前**完全没有探针**。修复后真机验证：api、operator `/readyz`+`/healthz`（internal 8081），registry `/v2/`（含命名端口解析）；三个 Pod `Ready=true`、`restartCount=0`、无 `Unhealthy` 事件；operator 新增 `--health-probe-bind-address`（8081，与 metrics 8080 分离，零检查也 404 → 已注册 ping）
+- **构建链路全通（`f79e5eb` + `02fd2de`）**：
+  - 传输：context_ref 改为 internal-face URL；`felis fetch-context` initContainer 走 service token（对象来自 `felis-build` 里按 `bootstrap`/`felis setup` 同款 Secret 复制机制落地的 `felis-service-token`；netpol 只放行控制 ns:8081）+ zip-slip 安全解包到限容 emptyDir；Kaniko `--context=/context` 只读挂载
+  - **真机 E2E**：玩家 OTP 登录 → 建 submission → 上传 gzip context（marker `felis-e2e-build-marker`）→ Owner approve → Job：fetch ✅ → Kaniko 构建并 push `registry.felis.svc:5000/user-uploads/<sub>:latest` ✅ → Trivy 扫描（内建镜像 DB）✅ → Job `Complete`；**从 registry 拉回镜像校验 `/hello.txt` 内容 = 上传的 marker** ✅
+  - 演练中真机抓到并修复 3 个单测看不到的缺陷：① Job requests=limits（2C/4Gi）在 4C/5.5G starter 上**永远 Pending**（`FailedScheduling/Insufficient memory`）→ requests 改为地板值（250m/512Mi，不超上限）；② Kaniko 重拷 Dockerfile 时 chown/chmod 到源文件 owner（distroless uid 65532）在 drop-ALL 能力下失败（`copying dockerfile: chown … operation not permitted`）→ fetch 容器以 root 解包（= Kaniko 自身 uid）；③ Trivy DB 默认 `mirror.gcr.io` 正被 egress 锁拒绝（fail-closed）→ 新配置 `[registry] trivy_db_repository` + 文档 §8e 固化镜像配方（`docker pull/tag/push aquasec/trivy-db:2` 进内建 registry；`--insecure` 已覆盖明文 HTTP）
+
+### 本轮新增真机证据（第四批：PG 契约测试 + 验证邮箱唯一性 + owner 角色）
+
+- **PG 契约测试首跑抓到 #20**：`internal/pgint` 对着真实 PG（`felis_pgint`，drop schema + 重放迁移）跑会话/OTP/op-login/绑定码/submission/build 全契约；`TestOnboardingEmailOTPRejectsTakenEmail` 当场变红 → 挖出「索引从未创建 + `ErrEmailTaken` 从未返回」。
+- **迁移 0020 真机应用**：`felis migrate up` → `applied 1 migration(s): [20]`；`schema_migrations` max=20；`pg_indexes` 出现 `users_verified_email_unique`。
+- **邮箱唯一性真机 409 演练（auditfix17）**：owner 已验地址被玩家 onboarding 流程二次验证 → 第一次 409 `email_taken`；**同码重放仍是 409**（证明码未被消费，符合契约；若被消费会是 400）。
+- **admin 改邮箱清 verified 真机演练（auditfix18）**：PATCH player.test → `[email protected]` 后 `email_verified=f`；PATCH 回原地址仍 `f`；玩家走 onboarding 重验证 → `t`。
+- **owner 角色真机闭环（auditfix18）**：owner 行按 0011 文档升级路径置 `role='owner'` → `GET /api/v1/users` 200（修复前 403）；owner op-login start 铸请求+发码、finish 得 `role=owner` 会话；PATCH owner role / DELETE owner / DISABLE owner 全部 403；玩家邮箱登录回归 200。
+- **配额原子门（第五批，`bb68fef`，auditfix19 已上线）**：pgint 并发实证（真实 PG 上两路并发认领：恰 1 赢 + 1 `ErrQuotaExceeded`；修复前两路全赢）；`docs/deferred-seams.md` 的 audit #4 条目核销。
+
+### 本轮新增真机证据（第六批：reaper 警告链路端到端演练，auditfix20）
+
+- **演练装置**：独立 hostNetwork Pod（`felis:auditfix20`，SA/卷结构复刻 live CronJob）+ 宿主本地零依赖 SMTP sink（捕获完整 RFC5322 原文）+ 专用 `felis-config-drill` Secret（SMTP 指向 `127.0.0.1:1025`）；场景服 `warntest`（独立 CRD，desiredState=Stopped，不占资源），期间 operator 暂停防 reconcile。
+- **投递**：13d idle → 恰 1 封，`To: [email protected]`（owner 的**验证**邮箱）、subject 含 `3d`、正文含 server/remaining；`warned_3d_at` 盖章、`warned_1d_at` 留空；`world_backups` 保持 4 条（纯警告零备份）。
+- **去重与档位**：重跑 0 封（已盖章不重发）；14.5d idle → 补发第 2 封 = `1d` 档（elif 顺序正确，不是重复 3d），`warned_1d_at` 盖章。
+- **失败语义（三连）**：SMTP 端口打坏 → 日志 `warn delivery failed; will retry next run`（err=`connection refused`）、**不盖章**，恢复端口后同一 run 补发成功；邮箱 `email_verified=false` → `resolve owner email: … has no verified email`、不盖章，恢复后补发；owner NULL → 静默跳过（无日志无警告）。
+- **边界**：idle=11d23h（< 12d 阈值）不发不盖章；idle=12d1m 发。`http_code` 冒烟：升级 auditfix20 后 panel 200 / op-login 200。
+- **环境还原**：warntest CRD+行删除、drill Secret/Pod 删除、operator 恢复；`servers` 表回到 resolvecheck+test-one，CRD 回到 lobby/login/resolvecheck/test-one。
+
+### 本轮新增真机证据（第七批：idle auto-stop 三层修复，auditfix21/22）
+
+- **发现路径**：operator 深挖时先怀疑"计时器无驱动"，真机实验立刻抓到 pruning（操作日志 `unknown field "status.emptySince"`）+ 触发后 88s 无动作 + RBAC 403，三层各自独立、各自足以致死。
+- **场景 1（超时点唤醒 + 写权限）**：EmptySince 已超时 11 分钟 → annotate 触发一次 → 秒级内 desiredState=Stopped、sts 0/replicas、EmptySince 清、phase=Stopped。
+- **场景 2（完整自驱，决定性）**：desired=Running → pod 起 → 20:11:43 盖章 → 静置无干预 → **20:12:17 自动 Stopped**（30s 到点后 ~4s 完成 stop+scaledown+markStopped）。
+- **翻转混沌**：3 秒内 Running→Stopped→Running，最终收敛 Running ready（期间 409 竞争为控制器正常噪音，controller-runtime 重试自愈）。
+- **收尾**：`spec.idle` 删除后再无自停；test-one 回 Stopped、空 sts、无 EmptySince；VM 镜像 auditfix22 = `f650bf8`。
+
+### 本轮新增真机证据（第八批：operator 自愈深挖，auditfix23/24）
+
+- **RCON Secret 删除自愈（`56f3abd`）**：修复前——删 secret 静置 60s 零察觉、触发后 96s+ 永久 `RconNotReachable`、仅人工删 pod 可恢复；修复后——删 secret **同秒**（20:22:16）察觉（Secret watch 日志）→ 新密码 + stamp 换值（`3b24164f…`→`c465b627…`）→ rollout → **20:22:47 全自动恢复 Running**（31s）。
+- **陈旧锚点（`82b5a60`）**：修复前——已恢复服务器因锚点残留被 `Failed(StartupTimeout)`、`Provisioned=False` 永挂；修复后——Ready 即清锚（`startRequestedAt: None`）、`Provisioned: True`；pod blip 演练：删除→20:24:48 重锚→20:25:19 恢复→锚清空，**全程无 Failed**。
+- **StatefulSet 误删自愈**：删除 sts → **秒级**重建（PVC `world-test-one-0` 保持 Bound，Retain 策略）、29s 后 Running、stamp 保持 `c465b627…`（未触发多余 rollout）。
+- **收尾**：期间 7 条 operator ERROR 全为 CR/sts 409 竞争噪音（自愈）；test-one 回 Stopped。
+
+### 本轮新增真机证据（第九批：未演练端点地毯覆盖 — 迁移 / access / window / credentials）
+
+- **account/migrate 四步状态机全链路（首个端到端演练，0 缺陷）**：造 user2 目标账户（绑定码 `FDTGKYAQ`）→ player.test claim test-one → 内部 start（未联动 UUID=404 not_linked；源=201 initiated）→ `passkey_required 409`（有 passkey 强制强因子）→ **passkey credentials list + DELETE 204**（顺带覆盖管理端点；删后 OTP 门自动解除）→ OTP start 202 / 错码 400 / 真码 confirmed → issue-code（self=400 / missing=400 / 真目标=201）→ redeem 负例×2（源持码=400；target 错码=403 setup_required——**onboarding 门正确拦截未验证账户**；验证邮箱后重试）→ **redeem 成功**（小写+空格码兼容，`servers_moved:1`）。
+- **retire 断言全对**：servers.owner→user2；migration=redeemed（含时间戳）；源 disabled+软删；源活跃 session=0；源旧 cookie 401；double-redeem 400；源再加迁移 409 account_retired。**环境复原**：源复活 + 新 OTP 会话、test-one 归还无主。
+- **access 玩家管理全组（0 缺陷）**：players/whitelist/banlist 三个读 projector 解析正确（`There are 0 of a max of 20 players online: `→[]）；kick/permission/group/luckperms-info 调用全通（demo 服无 LuckPerms → 原样透传 `Unknown or incomplete command`，API 不掩饰）；输入负例 4 连 400（bad player/bad action/bad node/bad world）；`whitelist add/remove` 在空档案服上得到 vanilla `That player does not exist`（fail-closed egress 无法解析 Mojang profile——**vanilla 约束非 API 缺陷**，命令已如实送达）；手写 whitelist.json + `command whitelist reload` 验证成功路径解析（`players:["E2E_Tester"]`）；**audit 7 条 access.\* 全记录**。
+- **updates/window**：unset→null；PUT 正例+读回；半设 400；倒置 400；clear→null 全对。
+- **/fleet**：4 服全量（含运行中 lobby/login 的 endpoint 地址）一次读全。
+- **bootstrap-assets crd**：嵌入式资产含 `emptySince`（auditfix24 二进制核对）。
+
+### 本轮新增真机证据（第十批：configure email 复制链修复，auditfix25）
+
+- **发现路径**：复核 `8e7c7bb` 新增的 replicate 逻辑时怀疑 manifest/-n namespace 冲突 → kubectl 真机实验裁决（hardcoded `namespace: felis` 版直接 `error: … does not match …`，证实第一段必错、第二段被跳过）。
+- **修复后行为验证**：两条命令序列（felis-smtp manifest 携带目标 ns + `felis-config` create--dry-run|apply）均被 server 接受且落点 `minecraft`。
+- **live 收敛项（随下次 `felis install/setup` 重跑）**：live CronJob 仍是旧模板（无 `FELIS_SMTP_PASSWORD` env）；`felis-smtp`/镜像两 ns 均未创建（SMTP 未配置，属正常）；与 operator Role 手动补 patch 同批处理。
+- **部署**：`felis:auditfix25` = `ed722d5`（api/operator/reaper 三处 set；api/operator rollout 完成，panel 200）。
+
+### 本轮新增真机证据（第十一批：submission 全路由负例矩阵 + 内部面取件 + owner 会话重铸）
+
+- **session 重铸路径（hook）**：owner 旧 cookie 过期（401）→ 临时插 `account_links`（owner↔假 UUID）→ op-login 三件套：start 202 → 日志取码 → 内部面 `op-login/{id}/approve`（hook，经 service token）→ status `approved:true` → finish 200 `role=owner` → op 域新 cookie 生效（`/fleet` 200）；**演习后已删除假链接行**。顺带复证：approve 失败时 status=false、finish 400 `op_login_invalid`（码不烧，复用同码二次 finish 成功）。
+- **submission 车道负例矩阵（23/23 PASS，0 缺陷）**：create 未知字段/空名/非法字符/超长 → 全 400；非 gzip 上传 400、未知 id 404；跨用户上传 → **404 不可见**（非 403）；玩家打 admin 三路由（queue/approve/reject）全 403；reject 空理由/1001 字/夹带字段 → 400，成功 200（`reject_reason`/`reviewed_by`=验主邮箱/`status=rejected` 全对）；重复 reject 409、reject 后 approve 409；**无 blob approve → 400 且行保持 pending_review（未 stranded）**；玩家列表严格只见自己。
+- **内部面取件**：`GET /api/v1/internal/submissions/{id}/context`（service token）→ 200 + gzip magic `1f8b` + 解包内容 = 上传 marker；无 token 401；未知 id 404。
+- **演习清理**：E2E-Lane-\* 8 行已删、假链接已删；队列只留历史 "E2E *"（approved）行。
+- 备注：port-forward 再次因 API pod 重建悬死（空响应）→ 按速查重启即恢复；本轮无需新镜像（纯验证）。
+
+### 本轮新增真机证据（第十二批：users 管理面全量矩阵 → 抓到并修复 #30）
+
+- **/users 管理面矩阵（修复前，39 PASS / 2 NOTE）**：create/get/patch/list 权限与负例全对（未知字段/非法用户名/owner 角色/未知 id → 400/404；owner 保护 403 复证）；sessions 单条吊销→旧 cookie 立即 401→重登恢复、revoke-all 同效（重登两次含 61s 冷却等待，全绿）；passkeys 无凭证 200 no-op；links 幂等/跨用户 409/非法 auth_source 400/解链 404 全对；disable/enable/delete/double-delete 全对。
+- **缺陷 #30（两处 FK 500 + 一处假 200）**：`PUT /users/{unknown}/quotas` 与 `POST /users/{unknown}/links` 触碰 user_id 外键 → **500 internal**（契约要求 404）；`GET /users/{unknown}/quotas` 回**零值视图（=unlimited）**，像 id 存在。修复：`PGRepo.requireLiveUser`（live 行 + `deleted_at IS NULL`）守卫三个方法，`handleGetQuotas`/`handleLinkAccount` 补 ErrNotFound→404 映射；单测 `TestAdminSubresourcesRequireLiveUser`（fake 同步 liveUserExists 契约）+ pgint 真库契约（ghost→ErrNotFound、live 对照、软删后→ErrNotFound）。commit `1918da2`。
+- **auditfix26 真机复验**：ghost 三连 → **404 not_found**；活用户对照 → PUT/GET quotas 200（读回 2）、POST link 200、delete 200；panel 200，api/operator 滚动完成（新 pod 24s Ready）。
+- 契约备注（非缺陷）：软删用户 `GET /users/{id}` 仍 200（带 `disabled=true` + `deleted_at`，列表已过滤）；`DELETE .../sessions/deadbeef` 幂等 200；revoke-all 对未知用户 200 no-op。
+
+### 本轮新增真机证据（第十三批：CLI / 镜像面 / 直接构建 / 面板 CDP 全路由 + #31/#32）
+
+- **CLI 面 19/19（VM，`felis-auditfix26`）**：`version`/`-h`/无参=exit2/未知命令=exit2；`manifests` 缺 `--velocity-cidr`、缺 `--felis-image`、坏 CIDR、坏 NodePort 全 fail-loud exit2，正例渲染 25 文档且 `--worlds-host-path` 分支出 CronJob+三条前置警示——整包 `kubectl apply --dry-run=server` **全部 `configured`**；`apply` 缺 `-f`/文件不存在/坏 JSON 各自 exit2/1/1；`migrate up` 二次幂等（`database already up to date`）；`bootstrap-assets crd`（含 `emptySince`）与 `game-stack`（tar）正常；`update` 正常。
+- **镜像面负例 + builds 负例**：`/images` 列表 200、空体 400、合法 201、重复幂等 201、无 `?ref` 400、删除 **204**（契约；我的 200 预期作废）、再删 404、坏 ref 400；`/images/build/unknown` GET/cancel 均 404；空体提交 400；玩家打 `/images*` 全 403。
+- **`/images/build` 正路径（直接构建）**：借遗留 submission blob 作 context，`FROM scratch` 内联 Dockerfile → **202 → 12s succeeded**；`/logs` SSE 实收 Kaniko 流；真机 registry `tags/list` 出现 `e2e/direct-probe:latest`（镜像确实被推入）。留档：该 drill 镜像保留在 registry。
+- **`auth/options`**：未知邮箱 → `{"methods":[]}`；玩家 → `["email_otp"]`（无 passkey 时）；坏邮箱 400。
+- **`PATCH /servers/{name}`**：空体/storage/空 policy/未知字段/坏名 → 400；未知服 404；displayName 设置+还原 200。**契约备注**：`displayName:""` 因 `omitempty`+merge-null 语义 = **清除该字段**（本次 drill 先误读为设空串，比对 patch 前 fleet 快照确认 test-one 原无 displayName，无数据损失）。
+- **面板 CDP 全路由巡检（owner 会话，14 路由）**：全部有标题/有内容/无崩溃；唯二发现 = #31/#32（见下）；`luckperms` 页对 stopped 服 409 → **优雅降级**为「服务器已休眠 + 启动」态（browser 级 409 日志属正常资源日志，非缺陷）。玩家会话 4 路由零错误且**导航分级正确**（不见管理/平台组）；未登录 `/login` 渲染正常（`/me` 401 为预期探测）、`/account` 未登录重定向 `/login`。
+- **缺陷 #31（面板 mock 残留·owner 判定）**：`ImageBuildPage` 用 `email==="[email protected]" || startsWith("owner@")` 猜 owner——真实 owner（`[email protected]`）**不被识别**（只能看 approved 提交），而任何 `owner@x` 邮箱都冒充 owner。改为消费 TierProvider 服务端 `isOwner`。commit `6907961`。
+- **缺陷 #32（面板 mock 残留·假构建种子）**：首次访问 `/admin/builds` 向 localStorage 播种 `["bld-1","bld-2"]` → 每个新浏览器打两个必然 404 的请求（页面注释自认“in production will 404”）。移除播种。同 commit。
+- **auditfix27 复验**：清掉该 localStorage 键后重访 → **只发 `/me`**、零错误、空态正确；面板 111 单测 + typecheck 全绿。
+- **passkey 全流程复验（CDP 虚拟认证器，面板 UI 实操）**：注册（对话框输入→仪式→列表出现 `E2E-Key`）→ API logout → 邮箱优先 passkey 登录（`/me` 回 `[email protected] user`、落回 `/`）→ `DELETE credentials/{id}` 204 → 列表清空（环境复原）。
+- **环境**：api/operator/reaper 镜像 = `felis:auditfix28`（= `6907961`）；面板 200。**教训**：`pkill -f "port-forward …"` 会匹配到**执行该命令的 ssh 自身 cmdline**（命令里同时含明文 `port-forward svc/…`）→ 自杀式断连；改为 `ss -ltnp` 取 PID kill，另起一条 ssh 启动转发。
+
+### 本轮新增真机证据（第十四批：#33 死账号复活漏洞 —— 全登录门闭环 + 资产切断）
+
+- **发现路径**：users 管理面演习收尾核对时发现软删用户仍留 `account_links`/`quotas` 孤儿行 → 顺藤摸瓜确认 `UserByEmail` 与 `SessionUser` 均不过滤 `disabled`/`deleted_at`。
+- **红证据（auditfix27，真机）**：禁用用户的邮箱门重登录 `verify=200`、`/me=200`（**禁用锁死可绕过**）；**已删号**用户 `start=202`（真实发码）、`verify=200`、`/me=200`（**删号可复活**）。
+- **修复（`5853589`，7 文件 / +306 行）**：① `UserByEmail` 只解析 live 账号（覆盖 email/passkey 邮箱优先/op-login/auth options 全部前置门）；② `SessionUser` 同样过滤（皮带：任何门铸出的死号会话都立即失效）；③ `RedeemPlayerBindCode` 两个解析臂拒绝死账号（新哨兵 `ErrPlayerAccountRetired` → 403 `account_retired`，**不消费码**，可逆重试）；④ `DeleteUser` 事务内 `DELETE webauthn_credentials` + `account_links`（凭证与 `UNIQUE(mc_uuid)` 占用不再外泄）；⑤ discoverable passkey resolve 补 liveness 检查。单测 `TestDeadAccountsCannotLogInOrKeepSessions` + pgint `TestDeadAccountsAreLockedOutInPG`（含"删后新铸会话不验证"与资产清点）。
+- **绿证据（auditfix28，真机）**：红期为死号铸出的会话 → **401**（皮带生效）；死号 `start=202` 中立且 **90s 日志窗口 0 条码**、`verify=400 invalid_code`；禁用中 `start` 中立（日志码数 1→1 不变）→ `verify=400`；**恢复启用后 `verify=200` + `/me=200`（可逆）**；bind 门死账号 → **403 `account_retired`** 且码保留 `count=1`（retryable）；删号后 `account_links`/`webauthn_credentials` 行数 = 0。
+- **演习清理**：红演练遗留（2 个测试账号的 quotas 行、1 条旧链接、2 条死号会话）已清除；`/fleet` 与 panel 冒烟保持 200。
+
+### 本轮新增真机证据（第十五批：#34 死账号同族收尾 —— in-game 身份解析与链接接管，auditfix29）
+
+- **发现路径**：#33 修完 Web 登录门后做同族复查——**in-game 面是 UUID 驱动、不走 Web 会话，因此 #33 的会话皮带盖不住它**。两处：① `UserByMCUUID`（claim / menu / wake 授权 / op-login vouch / QR link-status 五处共享解析）直接读 `account_links`、不过滤 users 的 liveness；② `VerifyLinkCode` 对「holder 已死」的接管语义未定义——一刀切 409 会把迁移退役这类真实场景锁死。
+- **修复（`bb9798e`，5 文件 / +182−7）**：① `UserByMCUUID` JOIN users 过滤 `disabled`+`deleted_at`——死账号的链接在游戏面**读作未链接**（claim 412、wake 落回 policy 门、vouch 403、status `linked:false`），绝不作为遗留身份存活；② `VerifyLinkCode`：**软删** holder 的链接可由新账号凭新 mint 码**接管**（账号已亡，码证明调用者仍持有该 UUID），**禁用** holder 仍 409（接管= 绕过锁死，不允许）、任何失败都不消费码。fake 单测 `TestLinkVerifyTakesOverDeletedLinkOnly`；pgint `TestVerifyLinkCodeTakesOverDeletedLink` + `TestDeadAccountsAreLockedOutInPG` 增「禁用链接无资格 / 复启用恢复」断言。
+- **绿证据（auditfix29 真机，三面六点）**：
+  1. `link/status`：活 `{"linked":true}` → 禁用 `{"linked":false}` → 恢复 `{"linked":true}`；
+  2. `claim`：活链 404（身份已解析、ghost 服不存在）→ 禁用 **412 not_linked** → 恢复 404；未知 UUID 对照 412；
+  3. **接管正例**：临时软删 holder（`de49df52`）→ 新 mint 码 `X3G3RZZM` 由 player.test verify → **200 linked:true**；链接行迁移到 player.test（`auth_source=mojang`）、码被消费（`account_link_codes` 0 行）；演练后 holder 与链接全部还原；
+  4. **接管负例**：临时禁用 holder（user2）→ 新码 `X5W94S42` verify → **409 already_linked**；同码重试仍 409（**不消费**）；恢复启用后**同码**由 holder verify → **200**（码保留、可逆）；
+  5. **wake 授权**：活 owner **202**（真实启动）→ 禁用 **403 forbidden** → 恢复 **202**；test-one 已停回 Stopped 且所有权释放；
+  6. **op-login vouch**：禁用 owner UUID → **403 not_admin**（与未链接/非 staff 同一个拒绝面）；活 owner UUID → 404 op_login_not_found（vouch 已解析、只是请求不存在）。
+- **环境刷新（部署态）**：热升级遗留 `FELIS_IMAGE=felis:auditfix16`（job 侧执行器 backup/restore/fileedit/forwarding-init 引用的镜像）→ 已 `set env` api/operator 双 Deployment 至 `felis:auditfix29` 并滚动完成；api/operator/reaper 三镜像一致 = auditfix29；port-forward 重启后内部面绿；panel 200、op 面 `/me` 200。
+- **运维备注**：`account_link_codes` 会积累过期行（"失败不消费"的另一面），巡检顺手 `DELETE FROM account_link_codes WHERE expires_at < now()`。
+
+### 本轮新增真机证据（第十六批：内部面残面地毯补测 —— ready / join-event / status / reclaim / blacklist / backup 负例）
+
+- **背景**：对内部面做覆盖盘点后发现五个端点此前无真机证据：`ready`、`join-event`、`status`、`player/reclaim`、`player/blacklist`，以及 `internal/servers/{name}/backup` 的负例臂。本批全部补测（auditfix29），**0 新缺陷**。
+- **`ready`**：test-one → **204**（advisory 语义）；坏名 → 400。
+- **`join-event`**（临时 UUID `3333…`）：首次 **204**、重复 **204**（幂等）；DB 实锤：`test-one.last_active_at` 15:02→05:53（重置）、`server_allowlist` 恰 1 行；缺 uuid → 400；未知服 → 404。演习后 allowlist 行删除、`last_active_at` 复原。
+- **`status`**：test-one → 200 全字段（phase/desiredState/endpointMode=fallback/…）；未知服 → 404。
+- **`player/reclaim`**：新建（临时 UUID `4444…`）→ 200 `hold_expires_at`=+30d；**幂等重试返回同一时间戳**（首窗保留，与存储行 21:53:51.726017Z 完全一致）；缺字段 → 400；**protected-admin 负例**：临时插 owner↔thirdparty 链接 → **409 protected_admin**（不 bar、不 stash）。清理后 `username_blacklist`/`player_data_holds`/`server_allowlist`/临时链接 = 0/0/0/0。
+- **`player/blacklist`**：bar 后命中 `{"blacklisted":true}`；陌生 UUID → `false`。
+- **`internal backup` 负例**：未知服 → 404；坏名 → 400；保留名（login）→ 400 `bad_name`（在入队前拒绝，RWO 闸门单测已覆盖）。
+- **探针**：`/healthz` 200、`/readyz` `{"status":"ready"}`。
+
+### 本轮新增真机证据（第十七批：hasJoined 多源会话校验器 —— 假 Yggdrasil 全链路 + 三方身份重写/改名）
+
+- **背景**：`hasJoined`（velocity 指向的 vanilla sessionserver 协议面，`handleHasJoined`）此前零真机覆盖。本批用「VM 主机假 Yggdrasil + 临时将 `[[auth_source]]` 换向」的方式把正/负路径全部打通。
+- **装置**：主机 python 假源（`:18099`，按 username 分流 `ftok`/`ftnotch`/`ftbadname`/`ftdown`/其余 204）；`felis-config` 的 littleskin 源临时改指 `http://10.42.0.1:18099/fake`（tag=`faketest`/prefix=`FT`），api 滚动后逐项打靶。**演练后配置已还原 littleskin、装置已清理**。
+- **结果（全绿，0 缺陷）**：
+  1. **三方身份重写**：`ftok` → 200 `{"id":"74409c3bbae93acabd2176e517b4f2a0",…}`，与本地按 `uuid.NewMD5(felisAuthNS, "faketest:native-123")` 的预算值**逐位一致**；
+  2. **保费名冲突改名**：`ftnotch` → `47c5527a18b03fe5a49ec50c11154dcd` + `name="FT_Notch"`（Mojang 实查 Notch=200 premium）；`ftok` 的 `FtPlayer` 恰也是真实 Mojang 名（`640c1672…`）→ `FT_FtPlayer`，改名按设计触发（非保费名保持原名）；
+  3. **敌意插件名拒绝**：假源返回 `§4admin` → **204**（不落 proxy 玩家列表）；
+  4. **源故障不静默**：假源 500 → **503**（velocity 报 auth servers down），对照未知玩家 → 204；
+  5. **形状负例**：缺参 / 超长参 → 204；声明 body → **400 + `Connection: close`**（防 drain 挂连接）。
+- **顺带覆盖**：`isPremiumName` 的 Mojang 实查（`api.mojang.com`，超时/错误 fail-closed=改名）——404→非保费、200→保费判别实测成立。
+
+### 本轮新增真机证据（第十八批：缺陷 #35 —— world 执行器 uid 1000 读不了游戏服写的世界 + 面板备份模块补全）
+
+- **发现路径**：给备份页新增「立即备份 + 最近操作」后做第一次真机 drill——面板链路全对（按钮/成功消息/运行态→终态、零坏请求），**但 backup Job 真的失败了**：`Job has reached the specified backoff limit`，pod 日志 `felis backup: archive: backup: tar walk: open /world/world/level.dat: permission denied`。
+- **根因（缺陷 #35，跨模块）**：世界卷的属主是**游戏镜像自己的 UID**（我们发布的 Paper 镜像都是 root），而 Paper 保存 `level.dat` 用的是 **0600**（Files.createTempFile 默认权限）→ 固定 uid 1000 的 **backup / restore / fileedit / reaper** 四类执行器：读不了（归档 `permission denied`）、也覆盖不了（restore 写不进 600-root 的 level.dat）。此前 drill 侥幸全绿，是因为当时的世界文件由 uid-1000 工具（restore/夹具）写的；**服务器真实启动保存过一次之后**，所有备份从此必死。测试全部是 shape 断言（无集群），这条只能真机抓。
+- **修复（`2010961`）**：四类执行器统一改**以 root 运行**（`runAsNonRoot:false`，省略 fsGroup 防误 chgrp），容器保持除 `DAC_OVERRIDE` 外 drop-ALL——与 operator `init-forwarding` 容器的既定先例同源（"只有 root 能可靠读写这些文件"）；DAC_OVERRIDE 兜住"游戏镜像是非 root UID"的任意镜像场景。四份 shape 测试同步改断言。troubleshooting §10「Permissions」段落重写（uid-1000 + setfacl 时代结束）。
+- **绿证据（auditfix31，真机四联 drill）**：
+  1. **backup**（面板 CDP 实操，红→绿同场景）：点击「立即备份」→ `进行中` → **`成功`**（同页面 reload 后仍在）；Job pod `runAsUser:0` + DAC_OVERRIDE；新归档 `test-one-1790115428472416736.tar.gz` 实测含 `world/level.dat`（471B）与 server.properties，共 499 条。
+  2. **restore**：`POST restore-backup`（bk-9df0…）→ 202 → Job **Succeeded**（root 写路径过关），日志 `restored from … into /world`。
+  3. **fileedit**：`GET /servers/test-one/file?path=world/level.dat` → **200**（base64-gzip 内容 628B）——修复前该请求必然 permission denied。
+  4. **reaper**（从 live CronJob 派生一次性 Job + 复刻钻取世界：1Gi PV/PVC + `world/level.dat` 512B **0600 root** + marker + 20d idle 行）：pod **root+DAC_OVERRIDE**；`world reaped server=reapdrill2 …`、`evaluated=3 reaped=1`；归档 711B 含 marker 与 level.dat；PVC 删除；`world_backups` 得 `inactive_15d` 行；servers 行/CRD 保留（红线②）。**钻取现场全部清理**（job/CRD/行/PV/宿主目录/临时归档）。
+- **面板补全（`97a64c8`）**：备份页新增「立即备份」（仅 Stopped 可用，409 原文呈现）与「最近操作」卡（GET `/servers/{name}/jobs`，running 每 5s 自刷新，failed 显示 Job 失败文本——正是它把上面这次失败暴露出来的）。面板 113 单测 + typecheck 全绿。
+- **部署注意**：backup/restore/fileedit 的 Job spec 是 api **运行时渲染**，随镜像即生效；**reaper CronJob 的 pod 模板是安装期静态渲染**——本轮已按新形状热补丁 live 对象（root + DAC_OVERRIDE），`felis install/setup` 重渲染时收敛。
+
+### 本轮新增真机证据（第十九批：面板补齐 owner 解绑通行密钥入口）
+
+- **背景**：`DELETE /users/{id}/passkeys`（owner-tier 凭据补救：密钥丢失/被盗时切断登录脚架，且不锁死账号——邮箱码/游戏内审批仍可用）后端早已实现并审计，但面板无入口，owner 只能靠 API。属"功能缺口"而非缺陷。
+- **修复（`11ac4f5`）**：用户详情页危险操作区新增「解绑通行密钥」行 + 确认对话框（`api.unbindUserPasskeys` 客户端方法 + 双语 i18n + wire-shape 测试）。
+- **绿证据（auditfix32，CDP 真机全链）**：player.test 注册虚拟认证器密钥 `E2E-Unbind` → `/auth/options` 从 `["email_otp"]` 变为 `["passkey","email_otp"]` → owner 在 `/admin/users/<id>` 点「解绑通行密钥」→ 对话框 → 确认 → **凭据列表清空、options 回落 `["email_otp"]`**（passkey 门确实关闭）；审计 `user.unbind_passkeys` 落账（同批还可见 `account.passkey.registered`/`backup.create`/`backup.restore`/`reap_world` 各审计行）。面板 114 单测全绿。
+
+### 本轮新增真机证据（第二十批：缺陷 #36 —— 提交者看不到构建结果）
+
+- **缺陷 #36（提交/构建模块·结果不可见）**：`/me/submissions` 只回审核状态；构建的成功/失败（含失败原因）只有 admin-tier `/images/build/{id}` 看得到 → **提交者永远不知道自己的包构建死了**。修复 `72c4aa3`：两个列表路由（玩家 `/me/submissions` + 管理 `/submissions`）对 `build_id` 非空的行附 `build_status`/`build_error`，来源是只读 `Builder.Get`（**绝不调 Sync**——状态推进归 15s reconcile 循环，列表渲染不碰集群）；构建行已消失（ErrNotFound）→ 字段省略；其他存储错误照常 500，绝不静默吞。OpenAPI 的 Submission schema 同步。
+- **绿证据（auditfix33，真机双例）**：① 失败例（上下文 Dockerfile `COPY does-not-exist`）→ 提交 → owner approve → `/me/submissions` 返回 `build_status:"failed"` + `build_error:"build job failed or scan found a CRITICAL CVE"`；② 成功例（`FROM scratch`+LABEL）→ `build_status:"succeeded"`。面板 CDP（玩家会话）：展开行显示「构建状态」徽章（构建失败/构建成功）+ 失败原因 + build_id，全程零 4xx。**演习残留已清**（2 行 submission+build、2 个 context blob、1 条 whitelist 条目；`sub-bf7dc1…` 那条是更早 E2E 遗留，未动）。
+
+### 本轮新增真机证据（第二十一批：面板文件编辑器补齐 —— 缺口而非缺陷）
+
+- **背景**：`GET /files`、`GET/PUT /file` 后端早已全绿（可读写 `level.dat`），但面板无入口——「一行 server.properties 写错导致起不来」的修复路径只有 API。属功能缺口。
+- **修复（`0a36b3f`）**：新增 `/servers/:name/files` 页：面包屑目录浏览、编辑器对话框（[]byte ↔ base64 编解码）、二进制文件打开即只读（NUL/非 UTF-8 拒绝 round-trip）、>256KiB 禁用保存；**停服门前置**（世界卷 RWO，未停服时整页显示「服务器正在运行」+ 停止动作，而不是让每个调用 409）；控制台右侧新增门口卡。i18n `files` 命名空间（en/zh）+ 3 条 wire-shape 测试。
+- **绿证据（auditfix34，CDP owner 全链）**：根目录 → `world/` 导航；`felis-e2e-marker.txt`（原 `v1\n`）打开 → 追加 → 保存「已保存 …」→ **API 读回一致** → 重开一致 → 还原原始字节 → 落盘复核一致（零残留）；`level.dat` 打开为只读 + 二进制提示；控制台门口卡存在；`audit_logs` 两条 `file.write`（edit/restore 各一）；全程零 API 4xx/5xx。面板 117 单测 + typecheck 全绿。
+
+### 本轮新增真机证据（第二十二批：缺陷 #37 —— 系统服务在面板里全是死操作）
+
+- **缺陷 #37（面板·fleet 死操作）**：`login`/`lobby` 是平台自建系统服务、名字在保留名单里，于是**每个 per-server 路由都用 `ValidateServerName` 拒绝**（400 `bad_name`）——但管理端 fleet 表格给这两行渲染 认领/停止/唤醒/控制台，全是死操作（控制台链接点进去也是一页 `invalid server name: reserved`）。修复 `2f90851`：`naming.IsSystemServer` 作为唯一事实源；fleet 行附 `system:true`；面板把这两行渲染为「系统服务」纯标签（owner 列 + 操作列），不再给任何动作。玩家侧不受影响（`/me/servers` 本就不含系统服务）。
+- **绿证据（auditfix35，真机）**：`GET /fleet`（owner）→ `lobby system:true`、`login system:true`、`test-one/resolvecheck` 无 flag；面板 CDP：两行「系统服务」、**0 按钮 0 链接**；`test-one` 行照常 认领/控制台、无系统标记；零 4xx。Go 侧新增 `TestIsSystemServer` + `TestFleetAdminRead` 的 system 子测试。
+
+### 本轮新增真机证据（第二十三批：缺陷 #38 + 多节点回收缺口 —— reaper 提示过期 / 钉节点）
+
+- **缺陷 #38（CLI·提示过期）**：`felis manifests` 渲染 reaper 时的 stderr 提示还在教“uid 1000 需要 `setfacl -m u:1000:x` 才能遍历存储根”——#35 之后 reaper 已改为 **root + DAC_OVERRIDE**，这条指导已失效且会误导运维（照做无害但白做，真问题被掩盖）。同批落 **多节点回收缺口**（结论第 5 条）：新增 `--reaper-node`，reaper CronJob 的 pod 渲染 `nodeSelector kubernetes.io/hostname=<node>`；多节点集群必须钉在存世界的节点，否则可能调度到 hostPath 为空的节点。`--reaper-node` 无 `--worlds-host-path` 时 fail-loud exit 2。修复 `daf7602`（双测：`TestReaperCronJob_NodePin`、`TestManifestsReaperNodePin`）。
+- **绿证据（auditfix36，真机 render + dry-run + 收敛 diff）**：
+  1. **固定渲染**：`felis manifests --felis-image felis:auditfix36 --velocity-cidr 10.211.55.6/32 --panel-node-port 30443 --worlds-host-path /var/lib/rancher/k3s/storage --archive-local-path /var/lib/felis/archives --reaper-node localhost.localdomain` → exit 0；bundle 内 `kubernetes.io/hostname: localhost.localdomain`；stderr **0 处** uid 1000 / setfacl，改为「已钉到节点 …」；
+  2. **负例**：`--reaper-node` 无 `--worlds-host-path` → exit 2 + 原文；不传 node 的渲染 stderr 仍完整保留「NO nodeSelector … 多节点必须传 --reaper-node」警示，bundle 内 0 个 selector；
+  3. **`kubectl apply --dry-run=server -f -`**：整包 **全部 configured**（含新 nodeSelector 的 CronJob）；
+  4. **再安装收敛性 diff**（`kubectl diff`，本批新增的收敛证据）：与 live 对比只剩 **一个对象**（reaper CronJob）两处实质增量——`+env FELIS_SMTP_PASSWORD`（热升级期未补的模板字段）与本次显式传入的 `+nodeSelector`；其余 24 份文档（api/operator Deployment、RBAC、NetworkPolicy、PVC、registry）**零差异**——即“热补丁过的 live 对象”与“当前代码重渲染”已收敛（reaper 安全上下文 root+DAC_OVERRIDE 两侧一致，无 diff）。
+
+### 本轮新增真机证据（第二十四批：S3 上传通道真机演练 —— 此前只有单测的暗路径）
+
+- **背景**：`internal/submit/s3store.go`（S3ContextStore）此前只有单测，本装是 local 路径（`user_uploads_context = "/var/lib/felis/uploads"`）从未激活。本批用「VM 宿主 MinIO（quay.io 镜像，:9000）+ 临时把 felis-config 切到 `s3://felis-user-uploads` + `[registry.s3] endpoint=http://10.211.55.6:9000` + 创建 `felis-uploads-s3` Secret」把整条通道打通，**全程 0 缺陷**、演练后还原并逐项复核。
+- **绿证据（auditfix36，真机全链）**：
+  1. 切 S3 后 api 滚动启动 **无** “S3 user-uploads store not configured” 告警（凭据解析成功）；pod 内 busybox 探针实测可达 `http://10.211.55.6:9000/minio/health/live`（rc=0）；
+  2. 玩家提交 + 上传 → **200**，对象实测落桶：`sub-12c953e12b0cd144/context.tar.gz`（197B，mc ls 实见）；
+  3. owner approve → 构建 Job `build-bld-1790118222183394193` **status.succeeded=1**（fetch-context 从内部面流式取件 = api 自 S3 读回成功）；`/me/submissions` 的 `build_status` 收敛为 `succeeded`；
+  4. **回滚**：felis-config 还原（sha256 与演练前备份**逐字节一致**）、删除 `felis-uploads-s3`、api 滚动；再演练一次本地路径：新提交上传 **200** 且 blob 实测落在 uploads PVC（context.tar.gz 197B）；启动日志仅剩 smtp/jwks 两条既有提示；
+  5. **清理**：2 行 submission + 1 行 build 删除、回滚演练 blob 删除、S3 构建产物从 whitelist 摘除（204）、MinIO 容器 + 两个镜像移除；`/fleet` 200。
+- **遗留观察（非缺陷）**：registry 里保留本次推送的 `user-uploads/sub-12c953e12b0cd144:latest` 层数据（与早前 direct-probe 同类，filesystem registry 无删除接口）；`felis setup` 的「S3 存储」向导屏本身未演练（列入 TUI 逐屏待办）。
+
+### 本轮新增真机证据（第二十五批：breakGlass 控制台逐屏全量 + 备份/恢复门禁 —— 缺陷 #39–#43）
+
+- **背景**：breakGlass 此前只验过 Owner 首装与 Add Operator happy path。本批把菜单四操作（Owner reset / Add Operator / Halt / Sync）+ 首装（bootstrap）分支全部逐屏真机走完，并打穿 Sync/恢复背后的 API 门禁；共抓 5 个缺陷、全部修复复验。驱动方式：VM tmux（`remain-on-exit on` 才能读回 alt-screen 撕掉后的 durable summary）。
+- **先落的正向证据（无缺陷）**：Halt——`test-one` Running → 选中 → 卡「is stopping」→ CRD `desiredState=Stopped`、pod 收敛消失、审计 `break_glass.halt`；面板 `wake`/`stop` 两个恢复杠杆均 202（复验后还原）。Sync 正路径——test-one（Stopped）→ 卡 backup started → Job 6s Complete → 归档落 `felis-backups` PVC（167MB）→ `world_backups` 行 `present` → `/api/v1/backups` 可见。Sync 负路径——对 Running 选 → 友好 409 卡（不误烧冷却）。
+- **缺陷 #39（owner 席位可被静默复制，且不可清理）**：恢复模式下用非在位席位名做「reset」→ `UpsertOwner` insert 臂**铸出第二个 owner 行**、原席位继续存活；面板对任何 owner 行都删/降/禁 403 → 永久无法收敛回单席。修复 `55d515d`：`provisionOwner` 先查在位席位（新 `PGRepo.OwnerUsername`），非席位名 → `ownerSeatTakenError`（`Is api.ErrConflict` → TUI 路由回表单并**指名**应输入的用户名）；bootstrap（无席位）与同席位名复位原样。真机双验：新名被拒（表单原位显示 `an Owner already exists as "08595879-…" — enter that username to reset the Owner`）；改席位名复位成功（owner 恒 1 行、id/邮箱不变）；面板删除保护同步实证（对新 owner 行 DELETE → 403）。
+- **缺陷 #40（operator 撞名 = 裸 SQLSTATE，「换名重试」分支在真机从未生效）**：`InsertOperator` 冲突返回原始驱动错误（23505），TUI 却按 `api.ErrConflict` 判定「可恢复、换名重试」——fake 与 PG 漂移。真机复现：输入已存在用户名 → 控制台 exit 1 + 裸错误。修复 `55d515d`：`isUniqueViolation → ErrConflict` + pgint 契约断言。复验：撞名**回到表单**提示换名 → `drill-op-2` 成功（审计 `break_glass.operator_create` 落账，演练行已清）。
+- **缺陷 #41（Sync picker 死选项）**：picker 列出 login/lobby，而备份 API 对它们**永远失败**（保留名 + 无 servers 行）；真机选 lobby → 裸内部错误 + exit 1。修复 `ac3a557`：`backupPickable` 过滤系统服（halt picker 保留它们，断玻璃完整权力）。
+- **缺陷 #42（缺失世界盘 → 202 后静默卡死 30 分钟）**：对「从未启动/已被回收」的服备份或恢复：202 → Job → Pod `persistentvolumeclaim "world-<name>-0" not found` **Pending 至 deadline**，全程零失败记录。真机用已回收的 `resolvecheck` 复现（留证后删除）。修复 `508a1c0`：`Cluster.WorldVolumeExists`（直接 Get 与 Job 挂载**同名**的 PVC）+ 两 handler 409 `no_world_volume`（"start it once to create it, then retry"）。**修复首跑翻出配套 RBAC 洞**：felis-api SA 无 `persistentvolumeclaims:get`（403 被吞成 500「internal error」）→ `APIMinecraftRole` 补 get-only 规则 + rbac 测试锚点。复验（auditfix38）：resolvecheck 两面 409 + 友好文案；test-one 照常 202 → Job 10s Complete → 新行落库。
+- **缺陷 #43（409 一刀切文案）**：TUI 把所有 409 当停服门 → 世界盘拒绝会展示错误原因。修复 `ac3a557`：按 body 的 `error.code` 分流（无 code 的旧体仍按停服门）。复验：无盘 pick 显示 API 原文；停服门文案不变。
+- **同步完成**：① bootstrap 分支 scratch 库演练（`migrate up` 20 迁移 → 无菜单/无认证直接铸 owner；`local_auth_enabled=true`；审计 `break_glass.bootstrap`；exit 0；库/hba 规则/临时配置即测即清）；② 台面收敛：reaper CronJob `suspend=true/auditfix36/无 pin` → `suspend=false / felis:auditfix38 / nodeSelector=localhost.localdomain`；③ 镜像升级 `auditfix37→38`（api/operator + `FELIS_IMAGE`）。
+- **流程修正（教训）**：pgint 一度误用**本机 Docker Desktop**（启动 daemon + 临时 PG 容器）——已完全清理（容器/镜像删除、daemon 退出），并改为**经 ssh 隧道用 VM 的 postgres** 运行（`ssh -L 15433:127.0.0.1:5432` → `postgres://felis:***@localhost:15433/felis_pgint?sslmode=disable`）。勿再在本机跑容器。
+
+### 本轮新增真机证据（第二十六批：`felis setup` 重跑向导逐屏 + 构建链 pin 回验 —— 缺陷 #44）
+
+- **范围**：本机已装机，故覆盖"重跑状态屏 + c/s/e 三条 reconfigure 流"；首装屏（postgres/owner/connect/edge/storage/smtp/mc-bind/migration/preflight/summary）此前各批已有定点真机证据（安装闭环 / 第二~三批 / 第九批 / 第二十五批），本轮不重复。
+- **逐屏走查（auditfix38→39，VM tmux）**：
+  1. 重跑 → 直落状态屏「✓ Felis is already set up.」（owner/connect 不触碰；host bootstrap 已就绪跳过）✓
+  2. `c` → 三选一 chooser（Local / Cloudflare+Access / Reverse proxy + 警示语）渲染 ✓，esc 无损返回。
+  3. `s` → chooser 预选当前后端；S3 分支表单（Endpoint/Bucket/Region/AK/SK + 提示）渲染 ✓。**观察：reconfigure 非只读**——选「Local disk」即 apply（写 /etc 两文件 + 重渲染 felis-config + 滚 API）；从 S3 表单 esc 退回会把选择重置为 Local 预选。
+  4. `e` → SMTP 表单渲染 ✓；esc 直接回状态屏、零副作用（felis-smtp 未创建）✓。
+- **缺陷 #44（CLI·重跑框脱落）**：重跑后完成 `s`/`c` reconfigure，落回首装 summary「✓ Setup complete.」——丢了 alreadySetUp 框（smtp 路径有专门分支，storage/connect 漏）。修复 `abb5910`（`showSummary` 透传 `m.result.alreadySetUp`）+ 回归测试 `TestRootReconfigureStorageKeepsStatusFraming`。真机复验（auditfix39）：storage reconfigure 完成 → 「✓ Felis is already set up.」+ storage recap ✓。
+- **演练事故（自曝；环境 drift，非产品缺陷）**：首次 storage 走查意外触发 Local apply——它从 `/etc/felis/felis.pod.toml` 重渲染 Secret，而构建链 pin 值当初**只热补在 live Secret、不在 /etc 文件** → 重渲染清空 pin（放任则下次构建死在 `:latest` 拉取 + trivy DB egress）。当场修复：pin 值写回 `/etc/felis/felis.host.toml` + `/etc/felis/felis.pod.toml` → 重渲染 Secret → 滚 API；再做第二次 storage apply，重渲染后 pin 仍在（drift 修复耐久）。
+- **构建链回验（direct build，真机）**：Job 规格实证 `kaniko=gcr.io/kaniko-project/executor:v1.24.0`、`trivy=aquasec/trivy:0.74.0`、`--db-repository registry.felis.svc:5000/mirror/trivy-db:2`（registry 仍有 `mirror/trivy-db`）；`POST /images/build`（context=遗留 `sub-bf7dc18…` blob，内部面取件）→ 202 → **succeeded**；registry `e2e/pins-check` 落位；whitelist 条目已摘除（204）、Job 已清。
+- **文档修正（`ae6e925`）**：§8e 原「编辑 felis.toml 后重启 felis-api」不完整（API 挂的是 Secret）→ 改为「写进 /etc 两文件 → 重渲染 Secret → roll」，并写明三种无效/易损做法（只 restart / 只改 host 文件 / 只热补 live Secret——后者会在下次 reconfigure 被冲掉）。
+- 收尾：api 1/1、panel 200、tmux 全清。
+
+### 本轮新增真机证据（第二十七批：告警模块落地 —— 内部面 /metrics + 规则集 + 真实构建失败实弹演练）
+
+- **范围**：把"指标 → 规则 → 告警"链路从零补到可交付：API 内部面 `/metrics`（`94f71ee`）、`deploy/alerts/` 规则与 promtool 单测（`43df08b`）、真机实弹演练（本批）。
+- **/metrics（`94f71ee`）**：`felis_image_build_failures_total` 此前只在进程内存里、无任何 scrape 出口。修复：internal face（8081）新增 `GET /metrics`（服务面 Public 路由，语义同 /healthz）；单测断言外部面 404。真机：port-forward `svc/felis-api-internal 18081:8081` → 200 且含 `felis_image_build_failures_total`；operator `:8080` 提供 `felis_servers_total` / `felis_start_duration_seconds_*`。
+- **规则集（`43df08b`）**：`deploy/alerts/felis-alerts.yaml`（plain Prometheus）5 条——构建失败 increase>0 / 起服 p90>300s / 磁盘可用<15% / DiskPressure / 内存可用<10%；`felis-prometheusrule.yaml` 为 prometheus-operator twin（脚本比对两文件 groups 一致）；`felis-alerts_test.yml` 为 promtool 单测。VM 上 promtool 3.14.0 实跑：`check rules` + `test rules` 双 SUCCESS。
+- **实弹演练（真实构建失败 → pending → firing）**：
+  1. 打包含 `COPY does-not-exist` 的 Dockerfile 上传到 uploads PVC `sub-alertdrill` → `POST /images/build`（`e2e/alert-drill2:latest`）；
+  2. Kaniko `failed to get fileinfo for /context/does-not-exist` → Job Failed、build 行 `failed`；
+  3. 真实 Prometheus（宿主 `:19090`）scrape `127.0.0.1:18081`(api) 与 `:18080`(operator) 双 target up；`felis_image_build_failures_total{job="felis-api"}=1`；
+  4. `FelisImageBuildFailures` pending（activeAt 08:28:14Z）→ **08:33:14Z 准时 firing**（`for: 5m` 精确到期），labels/annotations 完整。
+- **清尾（残留全清）**：whitelist `e2e/alert-drill` 摘除（204）；`sub-alertdrill` 目录、`/tmp/drillctx`、两个演练 Job、tmux `prom`/`fwd`/`alertpoll`、`/root/prom-drill`（promtool+prometheus 二进制）全删；DB `%alert-drill%` 行删除（whitelist/build 复核 count=0）；宿主无残留监听/进程。演练期间 live api 进程内计数器=1（重启归零，属演练事实）。
+- **可达性追加**：#44 定级 ①（轻）——已装机环境重跑 `felis setup` 完成 storage/connect reconfigure 即触发。
+
+### 本轮新增真机证据（第二十八批：缺陷 #45 —— 审核门"盲批"：评审看不到将被执行的 recipe）
+
+- **缺陷 #45（构建 lane·审核语义）**：被执行的 Dockerfile 永远来自**上传上下文压缩包根目录的 `Dockerfile`**（`build/jobspec.go` 钉死 `--dockerfile=Dockerfile`），API 的 `dockerfile` 字段**仅审计存档**（`submit.auditDockerfile`、`build.Request` 注释均已声明）——但审核者没有任何路径能看到它：`GET /api/v1/submissions` 不含 blob 内容、面板只显示 `context_ref` 文本、内部面取件路由是 service-token（评审用不了）→ "人工审核是门禁"事实上是**盲批**。同批口径缺口：`POST /api/v1/images/build` 的 `dockerfile` 字段在 OpenAPI 里无任何说明（易被当成"将被执行"），面板表单也把该框呈现为"Dockerfile 内容 *"。
+- **修复（`168a375`）**：
+  - 新增 admin-tier `GET /api/v1/submissions/{id}/context`：评审下载与构建 Pod 同源同字节的 `context.tar.gz`；`Content-Disposition: attachment` + `nosniff`（攻击者提供的归档只下载、不渲染）；审计 `submission.context.download`（actor=评审者、target=submission id）。
+  - 内部面取件 handler 共享 `openSubmissionContext`/`streamSubmissionContext`（行为不变，原测试锁定）。
+  - OpenAPI：新路由 + `/api/v1/images/build` 字段描述补全（明说"执行的是 context 根目录的 Dockerfile；`dockerfile` 仅审计"）。
+  - 面板：SubmissionsPage 展开区新增「下载上下文」按钮（spinner/错误呈现，i18n en/zh）；ImageBuildPage 表单补审计说明行。
+  - 测试：`TestAdminSubmissionContextRoute`（流式/404/503）+ admin-only 矩阵加该路由 + OpenAPI parity 强制文档；面板 117 单测 + 构建、go vet/go test 全绿。
+- **真机验证（auditfix41 已部署；owner 会话经 op-login + 内部面代 approve 重铸）**：
+  - admin 下载 `sub-bf7dc18e9dd97dc2` → **200**，`attachment; filename="context.tar.gz"`、`application/gzip`、`nosniff`；sha256 `205496f2…` **与 uploads PVC blob 逐字节一致**；
+  - 内部面（Bearer=felis-service-token）同 blob → 200 + 同 sha256（重构未破坏构建取件路径）；无 token → 401；
+  - 有效玩家会话（console host 隔离，仅 adminOnly 生效）→ **403**；匿名 → 401；不存在 id（admin）→ 404；
+  - 审计落账：`audit_logs` = `[email protected] | submission.context.download | sub-bf7dc18e9dd97dc2`；
+  - 面板产物：服务端 index.html 引用新构建 `index-CwFSKpgW.js`，bundle 内含新按钮逻辑（grep 命中 3 处）。
+- **可达性追加**：#45 定级 ①——每一次真实的"用户提交 → 管理员审核"都会踩到（审核者此前无法查看将被执行的内容）。
+- **hook 链补齐（可复用）**：staff 账号走邮件登录门会被设计拒绝（refuse staff）→ owner 会话铸法：`op-login/start`（email=felis-owner@example.com）→ VM 日志 grep `email-otp` 取码（no-Mailer fallback）→ 内部面 `op-login/{id}/approve`（Bearer=felis-service-token；body `approver_uuid`=owner 的 mc_uuid）→ `op-login/finish`（curl -c 存 cookie）。
+
+### 本轮新增真机证据（第二十九批：镜像耐久落地 —— registry 托管 + 回环拉取路径；#46–#49）
+
+- **背景（结论清单第 2 条）**：磁盘压力演练证明 kubelet 会 GC 掉"当前无人使用"的镜像 → ImagePullBackOff，恢复依赖人工重导入。本批让镜像自愈：自建镜像全部托管进内建 registry，节点侧 pull 经回环 hostPort（节点 containerd 到 Service VIP 是死路，实测 "Empty reply"）。
+- **平台侧 `a9b275a`**：registry 容器端口加 `hostPort 127.0.0.1:5000`；同提交修 **#46** —— registry 独立资源模板（1 CPU / 2Gi）：旧模板 256Mi 在实测推 475MB 层时被 OOM kill（dmesg `oom-kill … registry, oom_score_adj=989`，上传中断），2Gi 下同一推送 2 秒完成。
+- **安装器侧 `13d64e0` + 文档 `fa0e8d7`**：自建镜像规范 ref = `registry.felis.svc:5000/felis/{felis,limbo,lobby,paper}:demo`；import 进 containerd 就用该名（首启命中本地，免 registry round-trip），`deploy_bundle` 之后统一 `push_images_to_registry` 入仓（推 `127.0.0.1:5000`；registry 只认主机名之后的路径 —— 推/拉落点一致）。`configure_registry_mirror` 写 `registries.yaml`（`registry.felis.svc:5000 → http://127.0.0.1:5000`），内容不变不重启 k3s；`import_registry_image` 预缓存 registry:2（重跑走跳过分支）。迁移 **0021** 把 recommended 白名单重指到 registry ref（线上实查两行已落）。
+- **真机三次重跑**（`/opt/felis/src` = 765a892 快照；`FELIS_SKIP_FETCH=1 FELIS_INSTALL_MODE=full FELIS_IMAGE=registry.felis.svc:5000/felis/felis:auditfix42`）：
+  - run1 失败 = **#47**：Mac tar 的 `._*` 旁文件混入构建上下文，`._0004_*.sql` 被 //go:embed → 新二进制的 `felis migrate` 报 `non-numeric version "."`，安装器停在 run_migrations。修复 `5fa8b74`（.dockerignore 排除 `._*`/`.DS_Store`）。复验方式：故意在暂存树留 `._zz_probe_junk.sql` → 重建后 `migrations applied`（过滤生效）。
+  - run2 失败 = **#48**：每个镜像一对 `systemctl start/stop docker` 触发 systemd 限流（`Start request repeated too quickly / start-limit-hit`），第 4 个镜像（paper）未入仓。修复 `c7e585e`（整批一次 start/stop，单测锚定）。run3 零 `[fail]`：4 镜像全部入仓（push digest ×4 实收）。
+  - run3 收敛实查：api/operator/reaper = `registry.felis.svc:5000/felis/felis:auditfix42`（registry ref 首滚命中本地 import）；registry 模板 1 CPU/2Gi 生效；reaper CronJob 同步换 ref 且 nodeSelector 保留；login/lobby Running 于 registry ref；迁移 applied。
+- **GC 演练（本批验收本体）**：
+  - A 控制面：`ctr images rm …/felis/felis:auditfix42` + `ctr content prune references` → 本地 ref 消失 → `rollout restart felis-api` → 事件 `Pulling` → `Pulled … Successfully pulled image … in 25ms`，ref 恢复、pod Running。
+  - B 游戏：rm `…/felis/lobby:demo` → 删 `lobby-0` → `Successfully pulled … in 10ms … Image size: 182357358 bytes`，Running。
+  - #46 复验：整轮重建 + 4 推送期间 dmesg `oom-kill` 计数不变（仍 2，历史）。
+- **构建 lane 复验**：`POST /images/build`（context=遗留 `sub-bf7dc18…`；ref `registry.felis.svc:5000/e2e/durable-check2:latest`）→ 202 → succeeded；registry `e2e/durable-check2` tags 落位；Job 事件见 felis/kaniko/trivy 三镜像就绪；白名单条目摘除（204）、Job 清。
+- **#49（同一条升级路径的第二类静默回退）**：`write_felis_toml` 重写整表（[registry] 仅 url+build_namespace、[archive] 仅 store+local_path）→ §8e 的 kaniko/trivy pin、构建上限、uploads 后端、[registry.s3]、reaper 的 retention/warn_before/max_local_bytes 在重跑时全部丢失（S3 安装切回 local、构建回退被 egress 拒绝的上游 executor）。修复 `765a892`+`b8e554d`（沿用 [smtp]/[[auth_source]] 的 carry 模式；url/build_namespace/store/local_path 保持安装器所有）。真机复验：预置 `retention = "30d"`，run3 后 host toml / pod toml / felis-config Secret 三层都在，且构建 lane 直接用 carry 的配置跑通（上条）。
+- **可达性**：#46 ①（用户构建大层或安装器入仓即触发；实测 475MB 层）；#47 ②（Mac 打包树构建路径，真机踩中）；#48 ②（安装器重跑，真机踩中）；#49 ②（升级=重跑安装器，真机踩中）。
+- **口径/遗留**：`demo-up.sh` 未改（dev/demo 路径，本地 tag 导入维持原状）；kaniko/trivy 编译默认值未动（§8e 改为"镜像进 registry"配方）；registry 2Gi 为渲染常量（暂未开 flag）；drill 残留：registry 里 `e2e/*` 小镜像留档，`durable-check2` 白名单条目已摘。
+
+### 本轮新增真机证据（第三十批：缺陷 #50 —— `[smtp]` carry 吞掉下一节的注释块，每次重跑 +1）
+
+- **发现路径**：为 `felis setup` 首装连续走查做前置盘点时读 live 配置——`/etc/felis/felis.host.toml` 已经堆了 **3 份**、`felis.pod.toml` **4 份**重复的 Yggdrasil 注释块（同一段文案逐次叠加）；用脚本自带的提取器实测：一次重跑会把 3 份全部当作 `[smtp]` 内容带走，再叠一份模板注释 → **每次重跑 +1、无上界**（host/pod 增速不同步，现场 3 vs 4 即历史残迹）。
+- **根因**：`persisted_smtp_block` 打印"`[smtp]` 到下一个 section header 之间"的**所有行**；generated 注释块正好落在这段区间里 → 被吞并。同文件的自称"cached on first call"缓存因写方是命令替换（子 shell）从未生效，pod 写实际上重复抽取刚被重写的 host，加剧了两文件的不同步。纯注释膨胀、无功能损失，但属 #49 同族 carry 语义缺陷（把不属于自己的内容也带走了）。
+- **修复 `4d3c85f`**：carry 改为白名单（section header + 键行），与 #49 的 `[registry]`/`[archive]` 提取同型；删掉失效缓存说明。`deploy/bootstrap_test.sh` 新增用例：配置值被携带 / 不吞注释行 / 不越节 / 写回后二次抽取**字节稳定**（幂等）。
+- **真机验证（auditfix43，两次连续全量重跑）**：
+  - run1（`bootstrap-auditfix43.log`）：零 `[fail]`；host 注释块 **3→1**、pod **4→1**；与 run 前快照 diff 恰为 21/31 行（全部是被删掉的重复注释）——值零漂移（kaniko/trivy pin、`trivy_db_repository`、uploads local、`[registry.s3]` 空、`retention="30d"`、auth_source 原样、空 `[smtp]` 节保留；两文件 db host 仍分别为 127.0.0.1 / 10.211.55.6）。
+  - run2（`bootstrap-auditfix43b.log`，紧接再跑）：零 `[fail]`，仍 **1/1** —— 收敛证明（旧代码此处会 1→2 继续增长）。
+  - 部署同步：api/operator = `registry.felis.svc:5000/felis/felis:auditfix43`；reaper CronJob 同 ref 且 nodeSelector 保留；registry `felis/felis` tags = auditfix41/42/43；panel 200；host 二进制已从新镜像提取。
+- **可达性**：#50 ②——配置过 email（存在 `[smtp]` 节）的安装，按文档升级=重跑安装器即触发；危害=配置注释无限膨胀（每次 +1），无功能损失。
+
+### 本轮新增真机证据（第三十一批：首装连续走查 + 缺陷 #51 —— 工作负载 `felis-config` 副本永不刷新）
+
+**一、`felis setup` 首装单次连续走查（队列第 1 项，完成；0 缺陷）**
+
+- **装置**：scratch 库 `felis_scratch`（新建 + `felis migrate up` 21 条迁移）+ scratch 配置（真实 pod toml 副本，仅换库名；root 0600）；hook 直插 `account_link_codes` 一枚绑定码（等价 `/link` 内网端点写入，代替"进服拿码"）；tmux 驱动 `felis setup -config`；k8s 只读复用（登录门 Ready 等待通过）。
+- **连续走查（一条会话走完）**：Preflight（自动：PG✓/迁移 21 applied/面板✓）→ **MC 绑定**（输码 → working →「✓ Owner account is ready.」+ 一次性 setup URL）→ **连接 chooser**（Local）→ **存储 chooser**（Local → working → ~20s 后「✓ Local storage configured.」，含 Secret 重渲染 + API rollout）→ **首装 Summary**（「✓ Setup complete.」+ owner/setup URL/access/storage/panel 卡片 + c/s/e 提示）→ **轨道回顾**（← 依次只读 recap Storage→Connection→Owner→Preflight，→/esc 回到前台）→ Enter 退出 → stdout 汇总（`Owner account … provisioned (passwordless)`、`Recorded as "root"`、setup URL、Admin console）→ **EXIT=0**。
+- **结果**：移动端/文案/切换全部符合设计，**0 新缺陷**；scratch 库侧复核：owner(role=owner) 1 行、绑定码已消费(0)、setup_tokens 1、`local_auth_enabled=true`。
+- **副作用与还原（如实记录）**：存储 apply 会把 `/etc/felis` 两文件重写为 Go encoder 形态（无注释、含空值键如 `[archive.s3]`，**值零漂移**：kaniko/trivy pin、uploads local、retention 30d、auth_source 原样），并重渲染控制面 Secret + 滚 API——这是该向导的既定行为；演练后按 pre 快照整文件还原（sha256 逐字节一致），两 ns Secret 重渲染复核一致，scratch 库/配置/hba 行/tmux 全部清理。
+
+**二、缺陷 #51（工作负载 `felis-config` 副本永不刷新）**
+
+- **发现路径**：上条还原核对时发现 minecraft ns 的 `felis-config` 是**旧形态**（encoder 式）而 felis ns 已是模板式 → 挖出 `ensureSecretReplica` 的「绝不覆盖既有副本」（凭据语义：防冲掉手工轮换值）把 **felis-config 也纳入只建不更**，而 bootstrap 只 apply 控制 ns。后果：改配置后（DB 凭据轮换、[archive] 保留策略调整、root domain 等）backup/restore/fileedit Job 与 reaper 永远读旧副本 → 静默失效（如备份认证失败）。
+- **红证据（auditfix43，真机）**：向 minecraft 副本注入 `# drill-51-stale-marker` → 运行 `felis setup` → 输出 `- config (minecraft ns): skipped (already exists)`；副本 sha `5ec2ff6e…` 保持，控制面 `e4791fe1…` 不同（陈旧坐实）。
+- **修复 `328e570`**：① setup 侧：`ensureSecretReplica` 增 `refreshExisting`（仅 felis-config 传 true）——源缺失降级 skip、内容一致 skip(`already current`)、不同则原地 Update；凭据类保持 create-if-absent；新增 `updated` 结果与「refreshed from the control namespace」文案。② bootstrap 侧：新增 `apply_felis_config_secrets()`，每次运行同时 apply 控制 ns + 工作负载 ns 两份（同渲染自最新 pod toml）。单测：Go 新增 4 例（陈旧刷新/一致跳过/空键补写/源缺失降级）；`bootstrap_test.sh` 新增 4 断言（两 ns apply、同一 pod toml 渲染、恰 2 次 apply）。
+- **绿证据（auditfix44，真机）**：A) 安装器重跑（标记仍在副本中）→ 日志出现 `secret/felis-config configured`（工作负载 ns 被刷新）→ 副本 sha 与控制面一致、标记 0；B) 再注入标记 → 运行**新** `felis setup` → 输出 `- config (minecraft ns): refreshed from the control namespace`、凭据仍 `skipped (already exists)`、副本 sha 一致、`EXIT=0`。部署=auditfix44（api/operator/reaper），panel 200，host 二进制随镜像 `docker cp` 刷新。
+- **可达性**：#51 ②——升级=重跑安装器、或改完配置跑 setup 即触发；旧行为下 backup/reaper 静默使用旧配置（DB 轮换后备份全挂）。
+
+**三、环境修复（非产品）**：现场 pg_hba 缺 `felis_pgint` 规则（按文档走 ssh 隧道跑 pgint 会 ident 失败）——补回 `host felis_pgint felis 127.0.0.1/32 scram-sha-256` 并复测连接成功。
+
+### 本轮新增真机证据（第三十二批：S3 存储向导逐屏走查收尾 + 缺陷 #52 —— 向导 apply 不刷新工作负载 `felis-config` 镜像）
+
+**一、S3 存储向导屏走查（队列第 2 项，完成；0 功能缺陷，走查自身暴露镜像缺口 → #52）**
+
+- **装置**：VM 宿主 MinIO 容器（`quay.io/minio/minio`，`felis`/`felis-drill-9000`，:9000）+ bucket `felis-wizard-uploads`；tmux 驱动 `felis setup` 重跑向导；行动前先留 preS3 快照（两 toml + 两 ns Secret + sha256）。
+- **负例**：错误凭据 → `✗ Could not save storage settings.` + `submit: s3 credentials rejected: The Access Key Id you provided does not exist`（`CheckS3Access` 预检先于一切写入）；**零副作用**（无 Secret、两 toml sha 不变）；`esc` 返回编辑时已填值保留（密钥掩码）✓
+- **正例**：修正凭据 → ~10s working → `✓ Object storage configured.` → Enter → 状态屏 `storage  s3://felis-wizard-uploads · http://10.211.55.6:9000` ✓
+- **落地核对**：`felis-uploads-s3` Secret（access_key_id/secret_access_key）；两 toml `user_uploads_context` + `[registry.s3]` endpoint/refs；API 滚动；启动日志仅既有的 smtp/jwks 警告 ✓
+- **功能链（batch24 同款）**：player.test 邮箱 OTP 登录（hook 取码）→ `POST /api/v1/me/submissions` 201 → context 上传 200 → MinIO 桶实见 `sub-853a4e2ba4ba6443/context.tar.gz` → 内部面取回 200 + tar 内容正确 ✓
+- **UI 回滚（`s` → Local）**：两 toml 归位（`/var/lib/felis/uploads`、`[registry.s3]` 归空）、控制面 Secret 更新、API 滚毕；与 preS3 快照的差异仅「s3 字段回环 + encoder 形态」（注释丢失属该向导既定行为）——**值零漂移** ✓
+- **观察（不计缺陷）**：切回 Local 后 `felis-uploads-s3` Secret 残留（演练按清理流程删除；是否自动清理属产品取舍）。
+
+**二、缺陷 #52（向导内 apply 只刷新控制面，工作负载镜像滞后到下一次运行）**
+
+- **发现路径**：回滚后按计划复核「两 ns 重渲染」——minecraft 副本仍为 S3 内容（`4fb80ed8…`），而控制面与两 toml 已 local（`df206074…`）。
+- **红证据（auditfix44）**：① S3 方向：S3 apply（20:18:44）后副本停在 local 内容（20:22:53 快照 = `e4791fe1…`）达 4 分钟；② Local 方向：回滚 apply（20:23:41）后副本停在 S3 内容。副本 managedFields 两笔写入（12:13:58Z `kubectl-client-side-apply` = 安装器 rerun 的双 ns apply；12:23:06Z manager `felis` = 下一次 setup 启动的 refresh）都不是 apply 时刻——**apply 本身不碰副本**。
+- **根因**：`applyFelisConfigSecret`（storage/connection/edge 三条 apply 的公共出口）只 apply 控制 ns；镜像刷新只存在于 `felis setup` 启动（#51）与安装器。email 路径早有显式镜像刷新，storage/connection 是漏网的两条。
+- **修复 `de7fb2c`**：镜像刷新移入 `applyFelisConfigSecret`（best-effort + stderr 警告；控制面-only 安装无工作负载 ns 时降级不阻塞）；smtp helper 去掉重复块。门禁：`gofmt`/`go vet`/`go test ./...`/`bootstrap_test.sh` 全绿。
+- **绿证据（`v0.0.0+fix52`，宿主二进制 sha `0bd49467…` 已装 `/usr/local/bin/felis`，旧版留 `/root/felis-auditfix44.bin`；真机双向）**：Run1 从 Local `s`→S3：apply 后 `control = mirror = podtoml = 4fb80ed8…`（S3 渲染；旧代码此刻镜像会停在 local）；Run2 `s`→Local：`control = mirror = podtoml = df206074…`；副本 managedFields 写者 = `kubectl-client-side-apply` @ 12:33:40Z / 12:34:59Z（正是 apply 时刻）。Run1 启动块另见 `config (minecraft ns): refreshed from the control namespace`（#51 机制照常先收敛一次旧账）。
+- **可达性**：#52 ②——任何 `s`/`c` 重配置即触发；危害等级低（镜像消费者 backup/restore/fileedit/reaper 当前不读被改动字段——`UserUploadsContext` 仅 `api.go` 消费——但「配置动了、镜像没动」正是 #51 要消灭的静默滞后类，且与 email 路径的既定行为不一致）。
+
+**三、清理与还原**：MinIO 容器/卷/两镜像、`felis-uploads-s3` Secret、DB 行（`image_submissions` `sub-853a4e2ba4ba6443`）、/tmp 残留（player-cookies/drill-ctx/svc-tok 等）全清；docker 停；收尾 `control = mirror = df206074`（ALIGNED）、API 滚毕 Running、panel/healthz 200 ✓。
+
+### 本轮新增真机证据（第三十三批：`/updates` 维护窗口 API+UI 走查全绿；#53 跟随仓库迁移；#54 update 升级指引全假）
+
+**零、仓库迁移（背景）**：remote 已改 `[email protected]:FelisMC/Felis.git`（本批两个修复随 `2c6739a`、`397a400` 直推 main）。带 token 实测：旧 `MliroLirrorsIngenuity/Felis` 路径 301（GitHub rename redirect）、新路径 200——旧坐标当前仍能工作但全靠 redirect。新仓库**尚无 stable release**（`releases/latest` 带 token 也 404）：felis-api 更新检查的 404 属发布流程事实，非代码缺陷。
+
+**一、`/updates` 维护窗口 API 走查（完成；全绿）**
+
+- 读：GET 未设置 → `{"not_before":null,"not_after":null}`。
+- 负例全按预期拒绝：半设 / 倒序 / 相等 / 坏 JSON / 未知字段 → 400；缺 `Content-Type` → 415。
+- 正例：写入 → 回读一致 → **API pod 重启后仍在**（已落 DB，非内存态）。
+- 鉴权：无 cookie → 401；player 会话 → 403（新铸 player 会话，留 `/tmp/player-cookies.txt`）。
+- 收尾：DB 回到 `{null,null}`。
+
+**二、`/updates` 面板 UI 走查（CDP，完成；全绿零 console 错误）**
+
+- 状态流转逐屏：生效中 → 过期 → 计划 → 未设置（截图 `/tmp/updates-{1..5}-*.png`）。
+- 倒序提交 → 校验文案正确；清除后表单清空 + 「未设置」提示。
+- 核账：DB 收尾 `{null,null}`；审计 `updates.window_set` 恰 3 笔（20:46:18 / :20 / :21）。
+- 驱动：`/tmp/cdp-updates2.js`（bun + 原生 WebSocket；须 `Network.setCookie` 注入 owner 会话，否则新标签页 401 跳登录）。
+- 观察（不计缺陷）：窗口自身存取已验证；「窗口被 runner 消费」的端到端链路（Applier/Notifier）仍属 INTEGRATION-ONLY 设计（deferred-seams），不要当缺陷重复修。
+
+**三、#53（旧仓库坐标残留）—— 提交 `2c6739a`**
+
+- 红（`v0.0.0+fix52` 实机）：`felis update` → `github: MliroLirrorsIngenuity/Felis releases/latest returned HTTP 404 — …`。
+- 绿（`v0.0.0+fix53` 实机）：同一命令 → `github: FelisMC/Felis releases/latest returned HTTP 404 — …`。
+- 范围：`internal/updater/topology.go`、`deploy/bootstrap.sh`（默认 `FELIS_REPO_URL` + 两处 UA）、README ×2、两个测试夹具（6 文件 9 处）；门禁四件套全绿。
+- 可达性：②——默认安装 / 每次 `felis update` 都读该坐标；危害在 redirect 退休时兑现（安装与更新一起挂）。
+
+**四、#54（`felis update` 的 apply 指引在既有安装上全是错的）—— 提交 `397a400`**
+
+- 红（`v0.0.0+fix52` 实机文本）：`--panel --force` → `run: sudo felis setup` + “felis setup is idempotent and re-runs the installer…” + felis-api「例外」块。
+- 决定性证据（完成态安装）：`felis setup` 跑 8s 退出，`/opt/felis/velocity/velocity.jar` mtime/hash 不变、零 bootstrap 输出；同刻安装器重跑日志有 `resolving the newest Velocity 3.5.1 build`。代码侧 `shouldRunHostBootstrapBeforeConfig` 仅在 4 个 marker 不全时进 bootstrap——既有安装上 `felis setup` 只开配置 TUI。
+- 绿（`v0.0.0+fix54` 实机；宿主二进制已换 `/usr/local/bin/felis`，fix52 留档 `/root/felis-fix52.bin` sha `0bd49467…`）：
+  - `--panel --force` / `--velocity --force` → `run: curl -fsSL https://raw.githubusercontent.com/FelisMC/Felis/main/deploy/bootstrap.sh | sudo bash` + 单条 trailer（channel 未持久化 caveat、私有仓库 token'd form、`felis setup is not this path`）；
+  - `--mc` 无命令无 trailer；`--all` trailer 恰一次；`--force`/`LatestKnown` 语义不变。
+- 文案同步：`docs/troubleshooting.md §15` 删掉 “or `sudo felis setup`”、补 channel caveat；测试改为 `TestApplyGuidancePointsEveryComponentAtTheInstaller`（钉安装器路径 + 禁 `run: sudo felis setup`）。
+- 可达性：②（文档/指引；同 #38 型）。
+
+### 本轮新增真机证据（第三十四批：`felis nano` 全链走查 60/60；缺陷 #55 —— 坏源静默挡住验证梯子）
+
+**零、装置（全部在 VM，不进产品环境）**：`/srv/nanotest/nano-stub.py`＝可控 Yggdrasil 源，路径即行为——`/ok/<name>`、`/okid/<name>/<id>`、`/props/<name>`、`/badname/<case>`（tooshort/toolong/badchar/section/noname）、`/none`=204、`/boom`=500、`/redir`=302、`/garbage`、`/emptyid`、`/slow/<secs>`；跑在 127.0.0.1:9900，临时单元 `nano-stub`，请求日志 `/srv/nanotest/stub.log`。驱动脚本：`nano-matrix.sh`（HTTP 矩阵 S1–S7）、`nano-service.sh`（systemd 层）、`nano-config.sh`（配置校验）、`nano-fw.sh`（firewalld）；`nano-harness.sh` 从 HEAD 版 `bootstrap.sh` 提取产品函数在 scratch 路径跑（`felis-nano-test` 改名件，不碰产品单元）。
+
+**一、HTTP 矩阵 S1–S7（60 项断言）**
+- 红（修复前 `v0.0.0+audit-nano1`）：`SUMMARY pass=47 fail=13`——13 红全在 #55 语义内：5 个坏名用例「应 503+日志、实得静默 204」（状态+日志 ×2＝10），failover 3（坏源在前登录不落第二源 ×2、无日志 ×1）。留档 `matrix-red.r2.log`。
+- 绿（fix55，同装置重跑）：`SUMMARY pass=60 fail=0`。留档 `matrix-fix55.r2.log` 与 `matrix-fix55/`（更早原跑在 `matrix/`）。
+- 覆盖面：S1 输入校验（缺参/超长→204、POST→405、未知路径→404、带体 GET→400、HEAD→204、400 带 `Connection: close`）、S2 三方登录全链（前缀 + felisAuthNS UUID 确定性、ip/serverId 转发保真）、S3 premium 撞名改名（`LS_`）与免费名不动、S4 失败模式（不可达/500/302/garbage/204 → 跳过、记日志、503）、S5 多源优先级 + 坏源后 failover、S6 日志纪律（长 URI 截断、控制字节转义）、S7 properties 转发。
+
+**二、systemd 层（`nano-service.sh`，真 unit + 真 DynamicUser；重跑落盘 `nano-service.r2.log`，0 FAIL）**：装服务→起服务；重跑语义（配置字节不动、`resolve_nano_listen` 从 unit 读回端点）；drain（stop 期间 in-flight 登录照样答完，实测 stop 等待 5335ms）；SIGKILL 后自动重启；端点迁移 8081→8099 后新址应答、旧址关闭；三种启动文案（loopback「Bound to loopback」/ CIDR「admits」/ 无 CIDR「WARNING」）。
+
+**三、配置层（`nano-config.sh`，LoadNano 16/16，落盘 `nano-config.r2.log`）**：未知键、空/重复/保留 tag、colon、空白、坏 prefix、重复 prefix（大小写不敏感）、非 http scheme、URL 带 query、明文公网 http、缺文件、`[server] listen` 忽略警告、Mojang-only 可服务。
+
+**四、firewalld（`nano-fw.sh`，3/3，落盘 `nano-fw.r2.log`）**：关旧「开全源」8081/tcp、仅对 CIDR 开 rich rule、loopback 不开洞、无 CIDR 警告、快照清理还原。
+
+**五、缺陷 #55（提交 `9dad61f`）**
+- 红（fix54 实机）：配置源回 200 但档案字段不可用（三方源名字进不了 MC 字符集；身份源 UUID 不解析）→ 静默 204、零日志；且筛查在 `handleHasJoined` 直接 `return` → [坏源, 好源] 顺序下 204，好源根本没被询问。
+- 根因：两个 200-后筛查在 handler 层（200 已赢下梯子之后），而同类坏答案（200 无档案/非 200/不可达）在 `resolveHasJoined` 里是「记日志 + skip + failed」——一致性缺口。
+- 修复：筛查移入 `resolveHasJoined`（身份源 `uuid.Parse`、三方源 `mcUsernameRe`）→ 命中记日志（`unusable profile name` / `unparseable profile id`）+ `failed=true` + `continue`；handler 守卫保留为最后防线（注释更新）。单测 +3（含 `log.Writer()` 捕获断言）。
+- 可达性：#55 ②——需要配置了一个「回 200 但字段不可用」的源（马虎的自建/三方 Yggdrasil）；坏源在前时经梯子的登录全部被吞（静默、零日志），安全侧（坏名不落玩家列表）保留。
+
+**六、观察（不计缺陷）**：premium 冷查询 fail-closed 抖动——api.mojang.com 从 VM 首查偶发接近 2s 超时 → 偶按「疑似 premium」给免费名加前缀（`FelisNanoStub1` → `LS_FelisNanoStub`）；符合 `isPremiumName` 既定取舍（误加前缀=外观代价，误放行=抢名），记观察不修。
+
+### 本轮新增真机证据（第三十五批：NetworkPolicy 真机强制矩阵全绿 —— 端到端白名单闭环 + felis-velocity 刷新存活验证）
+
+**零、装置与现场**：三张策略 live 于 `minecraft` ns（`felis-default-deny-ingress` / `felis-allow-rcon-from-control-plane` / `felis-allow-game-from-velocity`；apply 于 09-22T05:37:24Z，与渲染收敛 diff 零漂移）；k3s 参数无 `--disable-network-policy`，**强制执行实测生效**；kube-router 机制实证：per-pod `KUBE-POD-FW-*` 链、未标记流量 `REJECT --reject-with icmp-port-unreachable`、每链首条 `--src-type LOCAL -j ACCEPT`（本节点豁免）、ipBlock 落 ipset `KUBE-SRC-*`（白名单成员可直接查）。探测法：**nsenter 进真实 pod 网络命名空间（源 IP=真实 pod IP）+ netns/veth 合成「转发型外部源」10.99.0.2（模拟 velocity 在另一台机器）**；全程不改产品代码。
+
+**一、pod 源矩阵（源=真实 pod netns）**
+- api（felis，命中 RCON 白名单标签）→ lobby:25575 = **OPEN**；同源 → lobby:25565 = **REFUSED**（端口级区分 ✓）
+- registry（felis，非匹配）→ lobby:25575/25565 = **REFUSED**；同 pod → api:8081 = OPEN（对照：无策略命名空间不受限）
+- coredns（kube-system）→ login:25565 = **REFUSED**
+- 收尾复跑全矩阵与首轮逐行一致（`np-matrix-run1/2.log`）。
+
+**二、host（=velocity 同机侧）**：→ login/lobby 的 ClusterIP 与 podIP :25565 **全 OPEN**（velocity 注册的后端路径实际可用）；→ api-internal:8081 OPEN；→ lobby:25575 亦 OPEN——归因：kube-router 每 pod 链的 `--src-type LOCAL -j ACCEPT`（**本节点流量豁免，kube-router 设计行为**，kubelet 探针等依赖它；netpol 语义无法对节点自身收口，节点 root 本在 TCB 内）。**注记①：node-local 豁免。**
+
+**三、转发型外部源（最严苛模拟）**
+- 基线：netns(10.99.0.2) → 全部服务器端口（podIP 与 ClusterIP、25565/25575）= REFUSED，包级可见 netpol 的 icmp-port-unreachable（`td3.log`）。
+- **白名单闭环**：临时把 10.99.0.0/24 加入 allow-game → ipset 即时生效（members：`10.99.0.0/24` + `10.211.55.6`）→ **lobby-svc:25565 与 login-svc:25565 = OPEN**（ClusterIP=velocity 实际拨号形态）；25575 仍 REFUSED（端口维度不破）→ 回滚 → spec 哈希逐字节一致（md5 `e6b07440…`）、ipset 复原、复测全部 REFUSED。
+- 层间归因：firewalld 规则含 `ct status dnat accept`（**DNAT 后的服务流量放行**；非 DNAT 转发走 forward policy 的 `admin-prohibited` 拒绝）——受支持拓扑（velocity 同机=LOCAL、ClusterIP、Mac→NodePort 面板 200）全部实测可用；「外部未经服务直连 podIP」不属于任何产品流。**注记②：firewalld 只放 DNAT 服务流。**
+- 装置保养：veth 未归区时 firewalld 会拒其转发属装置噪声（已归因）；public 区临时挂载已摘、netns/veth 已删、策略零残留（spec diff 为空）。
+
+**四、felis-velocity 刷新循环存活（顺带验证）**：20:18–20:42 的 `server list refresh failed` 全部落在 #52 演练的 API rollout 窗口（成功不打日志属设计）；tcpdump 450s 窗口抓到 52 条 `GET /api/v1/servers` 载荷行（成对出现=双抓包点看到同一请求，折算约 26 次 ≈ 每 15–17s 一次，与 `REGISTRATION_REFRESH=15s` 常量吻合）及对应 200 响应；"keeping current registrations" 为设计降级。非缺陷。
+
+**五、留档（VM `/srv/npdrill/`）**：`np-matrix.sh`+`np-matrix-run1/2.log`、`np-netns.sh`、`netns-probe.py`、`np-cidr-test.sh/.log`、`spec-before/after.json`、`td3.log`（包级归因）、`tcpdump-8081.log`（刷新存活）、`iptables-save.txt` 与 `nft-rules.txt`（现场快照）。
+
+### 本轮新增真机证据（第三十六批：面板错误文案全量本地化（#60）+ Run4c 管理面写操作复核 0 缺陷）
+
+**零、现场**：三个缺陷批次（#56–#59）已按「一缺陷一 commit」推 main 并部署 `auditfix59`；本批完成 #60 后部署 `auditfix60`（api/operator/reaper 三处 + 宿主 CLI，`felis version` = v0.0.0+fix60）。面板门禁：typecheck 0 错、vitest 117/117、vite build 通过（注意项目测试是 `bun run test`＝vitest run；裸 `bun test` 是 Bun 内置 runner，解析不了 `@/` 别名，41 个 fail 系假象，勿误报）。
+
+**一、Run4c 管理面写操作复核（4 段全真机，最终 0 缺陷；3 处为测试脚本自身误报，均已自纠）**
+1. **镜像添加+删除**：添加 `registry.felis.svc:5000/e2e/probe:1` → 列表出现、API 落库（`added_at=14:50:53Z`）✅。删除初测"未生效"系脚本用 `tr` 找行——该表是 `div` 网格，`tr` 选择器命中 0 → 从未点到按钮。换 `div[class*=grid-cols-12] → button[title]` 重测：confirm（"确定删除此镜像吗？"）自动接受 → 页面 1ms 内刷新、**API 侧同 ref 从 5 条降至 4 条**——删除通道无缺陷。
+2. **构建负例**：外部 registry（`docker.io/...`）+ 不存在 context → 对话框红字拒绝 `build: invalid request: image reference … must target the internal registry "registry.felis.svc:5000"`（截图 `run4c-3`）。脚本"未捕获"系其正则先命中了侧边栏"镜像"二字——误报。
+3. **创建重复用户**：对话框内正确显示 **"该用户名已被使用。"**、对话框保持打开（截图 `run6-1`）。此前 run4c 的"dialog-closed"系脚本点错页面（进到了用户详情页）；且该轮实际是一次**正例**：真管理员的 username 是 UUID 串（`08595879-…`，role=`owner` 是角色名），字面用户名 `owner` 当时空闲、创建确实成功——测试件已删（DELETE→200，列表 7→6）。
+4. **创建重复子域名**：正确拒绝 **"该子域名已被占用。"**（截图 `run4c-5`）✅
+
+**二、缺陷 #60（提交 `c9af548`）：37 个用户可达错误码显示生英文**
+- 发现路径：run4c 复核 `email_taken` 时做了一次系统性对照——后端 `newError` 共 **71 个错误码**，面板 `humanizeError` 只映射 **23 个**；其余走 `default` 分支直接透传 `err.message`（多为 Go 包装嵌套的英文，如 `invalid server name: naming: invalid server name "BadName": must match ^[a-z0-9-]{3,32}$`）。
+- 三个代表真机验证（修复前→修复后）：
+  - `bad_name`：新建服务器"名称"填 `BadName`（前端只查非空）→ 前：对话框直出上述嵌套英文；后：**"服务器名称不合法：需为 3–32 位小写字母、数字或连字符，且不能使用保留名。"**（`run8-1`）
+  - `bad_subdomain`：子域名填 `ab`（前端 RE 允许 1–2 字符，后端要求 ≥3）→ 后：**"子域名不合法：…"**（`run8-2`）
+  - `email_taken`：DB hook 造"他人已验证邮箱" → 账户页真发码（no-mailer 日志取码）→ verify 409 → 后：**"该邮箱已在其他账户上完成验证；请直接用该邮箱登录，或换一个地址。"**（`run9-1`）
+- 修复：api.ts +37 case、zh/en errors.json 各 +37 键（纯映射，无逻辑变更）。**剩 11 个故意不补**：`bad_request`/`conflict`/`internal`/`panic`/`not_found`（通用兜底）、`forbidden`/`not_admin`/`unauthorized`（状态码分支已在 default 覆盖）、`not_ready`（内部面）、`setup_required`（bootstrap 期 SPA 流控信号，403 文案可用）、`unsupported_media_type`（CSRF 门，面板永不可达）。
+- 遗留观察（不计缺陷）：构建负例的英文前缀 `build: invalid request:` 仍会透传——该错误无独立码、message 即最终文案；对管理员可读，记观察。
+
+**三、观察（不计缺陷）**
+- UserDetailPage 对 owner/对自己都显示删除按钮；真的点了会得 403 + 具体 message（"the owner account cannot be deleted from the panel"），但前端 403 分支统一显示"你无权执行此操作"——笼统但不算错，记观察。
+- 字面用户名 `owner` 可注册（用户名无保留名单；角色由服务端管理、无提权路径），记观察。
+
+**四、留档（Mac）**：脚本 `/tmp/cdp-run5.js`（镜像删除复测）、`run6.js`（重复用户复测）、`run7/run8.js`（bad_name/bad_subdomain 前后对照）、`run9b.js`（email_taken，内置 ssh 取码）；截图 `run5-1`/`run6-1`/`run7-1/2`/`run8-1/2`/`run9-1`。VM `/opt/felis/src` = `c9af548` 快照。
+
+### 本轮新增真机证据（第三十七批：管理面交互收尾 0 缺陷 + 缺陷 #61 —— LuckPerms 写操作的过度承诺）
+
+**零、现场**：`auditfix61`（api/operator/reaper 三处 + 宿主 CLI）。面板门禁 typecheck/vitest 117/build 全绿后部署。
+
+**一、管理面交互收尾（3 项全真机，0 缺陷）**
+1. **submissions approve/reject**：player 会话（邮箱 OTP 铸造，no-mailer 日志取码）现场造两条 pending（`sub-489eda47fe4bc12d` / `sub-b1dcc4b4a2cc5f16`，各上传 tar.gz context）→ 面板「通过」→ 状态变「审核通过」且**自动构建 `bld-1790176562566802097` 到 succeeded**（点击到构建完成全链闭环）；「驳回」→ 对话框必填原因 → 状态变「已拒绝」。全程 UI 零报错。
+2. **用户会话撤销**：player.test 铸 2 条新会话（共 4 条）→ UI「单独撤销」首条 → **精确生效**（cookie5 → 401、cookie4 → 200、列表 4→3）；「全部撤销」→ 提示「所有会话已撤销。」、列表清空、cookie4 也 401。
+3. **ServerCard 启动交互**：面板点 test-one「启动」→ `Starting` → `Running/ready`。
+
+**二、缺陷 #61（提交 `b4ef42d`）：LuckPerms 页对写操作"过度承诺"**
+- 发现路径：LP 页写操作交互测试（给 `E2E_Tester` 写权限）→ 历史卡显示绿色 success +「（服务器未返回输出）」→ **落盘取证发现根本没写入**。
+- 真机实验矩阵（test-one，LP 5.5.85 / H2 存储）：
+  | 操作 | RCON 回包 | 落盘（`lp export` 实测） |
+  |---|---|---|
+  | 面板 UI：`E2E_Tester permission set e2e.ui.write.test` | 空 | ❌ |
+  | 面板控制台：`E2E_Tester … set probe.test true` | 空 | ❌ |
+  | 面板控制台：`<UUID> … set probe2.test true` | 空 | ✅ |
+  | **独立 Python RCON 客户端**（绕开 felis）：`E2E_Tester … set probe3.test` | 空 | ❌ |
+  | 独立客户端：`<UUID> … set probe4.test` | "Another command…"（异步提示） | ✅ |
+  - 结论：**felis 只如实转发空回包；名字写入的静默失败是 LuckPerms 自身行为**（独立客户端 1:1 复现）。
+- 根因（LP config.yml 官方注释背书）：`use-server-uuid-cache: false`（LP 默认）→ "commands using a player's username will not work **unless the player has joined since LuckPerms was first installed**"——**未进过服的玩家名永远无法解析**（与有无外网无关）。
+- 面板问题：旧文案承诺"授予与撤销操作仍然会实际生效"（#59 遗留半句）→ 对"给还没来过的玩家预授权"场景是不成立的承诺。
+- 修复（纯文案，双语）：`luckperms_no_reply` → "……按玩家名的操作只对「自 LuckPerms 安装以来进过本服」的玩家可靠，对没进过服的玩家名可能静默不生效"；`luckperms_no_output` → "（服务器未返回输出，无法确认结果）"。真机复验（`run16b`）：读提示与写占位均按新文案显示。
+- 可达性：#61 ①——给未进服的玩家预授权是日常动作，此前面板显示绿色成功构成误导。
+- 局限（记观察）：读提示块只在"读取为空"时渲染；已有部分玩家数据的服上不会露出这段说明（后续增强候选：常驻说明或写前对未解析名字的提示）。
+
+**三、观察（不计缺陷）**
+- LP 命令是**串行异步**执行：快速连发第二条会得到 `§7[§b§lL§3§lP§7]§r §7Another command is being executed, waiting for it to finish...`（带颜色代码原文，面板如实显示）；`lp export` 回包时有时无（响应与执行解耦）。felis 层无责。
+- submissions 徽章（`submissions.json`："审核通过/已拒绝"）与过滤标签（`admin.json`："已通过/已驳回"）两套词，语义均可，记观察。
+- 控制台对空回包命令只显示 echo、无占位提示（终端风格）；LP 页有占位文案。记观察。
+
+**四、留档（Mac）**：`/tmp/cdp-run10.js`（approve/reject）、`run11a/11b.js`（会话撤销）、`run12a.js`（启动交互）、`run13/14/15/15b/16/16b.js`（LP 全链）。VM 独立探针：`/tmp/rcon_probe.py`、`/tmp/rcon_one.py`（+ `/tmp/rcon_pw.txt`）；LP 导出样本在 `/data/plugins/LuckPerms/luckperms-2026-09-23-15-*.json.gz`。测试数据：两条 pending 提交（一 approved 一 rejected，构建 succeeded/failed 各一）。VM `/opt/felis/src` = `b4ef42d` 快照。
+
+### 本轮新增真机证据（第三十八批：#56–#59 证据回填 + 服务器详情/运维面收尾 0 缺陷）
+
+**零、现场**：`auditfix61`（api/operator/reaper 三处 + 宿主 CLI，`felis version` = v0.0.0+fix61）。本批两部分：把 #56–#59 四项修复的真机证据落节（此前只存在于 commit message），并把队列里剩余的服务器详情页/运维面交互复跑做完（0 缺陷）。
+
+**一、缺陷 #56–#59 证据回填（部署 `auditfix59`；四项均 = 修复前真机触发 + 修复后复验）**
+
+1. **#56（`0790f8d`）玩家管理操作把 RCON 回包扔掉、只报 canned 成功。** 触发路径：给"还没进过服"的玩家加白名单/封禁——vanilla 对没见过的名字回 `That player does not exist` 并拒绝，而旧面板无论如何都显示成功。修复后回包逐字上屏：`whitelist add E2E_Bad` → `Added E2E_Bad to the whitelist`（磁盘同步真写入）；负例 `NoSuchPlayerXYZ` → `That player does not exist`（不再伪装成功）；静默服务器回退本地化文案。
+2. **#57（`70c988e`）控制台命令的回复无处显示。** 触发路径：控制台发任何命令——`sendCommand` 拿得到 RCON 回包但被丢弃、pod 日志也不回显命令输出，等于零反馈。修复后 echo + 回包以终端样式渲染在提示符上方（实测 `list`）。
+3. **#58（`4d4cdd6`）无世界盘服务器的文件页 90s 卡死。** 触发路径：对从未启动/已回收的服点"文件"页——旧行为建 Job → Pod `FailedScheduling (pvc not found)` Pending 到 90s 超时 → 误导性 504 `files_timeout`。修复：file 路由补上与 backup/restore 同款 `WorldVolumeExists` 门 → 快速 409 `no_world_volume` + 面板双语文案。本批现场复查：`resolvecheck`（无 world PVC）→ 409 `no_world_volume`，0.03s；`test-one`（有盘）→ 200（~2s）。
+4. **#59（`a2ff2a1`）LuckPerms 页把"读不到"报成"没有"。** 触发路径：打开装了 LP 的服的管理页——LP 5.5.85 的 `lp` 命令 RCON 回包全为空（独立 RCON 客户端 1:1 复现），读投影永远为空，旧页面却断言"没有父组/没有显式节点"（假事实）、写历史伪造 `[RCON]` 行。修复：原始回包随 rosters 同款披露渲染；空回包显式提示、不再假断言；历史占位不再伪造输出。
+
+**二、服务器详情页交互复跑（5 项全真机，0 缺陷）**
+
+1. **ServerCard 停止**：test-one `Stopping` → `Stopped`（启动在第三十七批；收尾复查 `desiredState: Stopped`）。
+2. **ServerFiles 写流程**：编辑 motd → 保存「已保存」→ 重开读回一致（`motd=Felis E2E files drill`）→ 还原默认 `A Minecraft Server`（收尾复查确认）。
+3. **备份/恢复**：立即备份 → Job `succeeded`；恢复 → 确认对话框（破坏性警告原文）→ Job `restore-test-one` `Complete`（5s）→ UI「恢复 27秒钟前 成功」→ 唤醒 `Running/ready`（恢复后的世界可加载）。
+4. **白名单**：加/负例/移除全链复跑（证据见 #56）——终态磁盘只剩 `E2E_Tester`（88 字节）。
+5. **封禁/解封**：封禁（内联确认）→ `Banned E2E_Bad: Banned by an operator.` + 磁盘写入；解封 → `Unbanned E2E_Bad`；`banned-players.json` 终态 `[]`。
+
+**三、运维面复查（3 项，0 缺陷）**
+
+1. **`/admin/updates` 复跑**（队列"深挖"项）：设置窗口 → 「维护窗口更新成功。」+ 状态卡更新；负例 end<start → 前端「结束时间必须在开始时间之后。」、后端 400 `end must be after start`；清除 → 「当前未设置维护窗口」。
+2. **metrics 端点**：内部面 `/metrics` 正常——`felis_*` 样本 18 条 / 3 个指标族（build 失败计数、回收计数、起服时长直方图）。
+3. **CLI 覆盖核销**：`run` 分发表 17 项（15 个用户面命令 + `bootstrap-assets`/`init-forwarding` 两个容器内部入口）在历批演练中均已有真机记录，本批逐项核销无遗漏。
+
+**四、接口语义注记与收尾**
+
+- 文件 API 的 `path` 是**相对路径**：空串 / `.` = 根（正常列出）、`config` 下钻正常；字面 `/` 被路径约束拒绝（400 `bad_path: path escapes from parent`）——面板从不发绝对路径（`joinPath` 只拼相对段），无用户面影响；本批复查脚本初次误用 `/` 时曾见 13s 延迟，属首次 fileedit Job 冷启动，非卡死。
+- 收尾静止态：test-one `Stopped`；三个名单 = whitelist `E2E_Tester` / banned `[]` / ops `[]`；`E2E_Bad` 仅余 latest.log 与 usercache.json（日志与 Mojang 缓存）。
+- 留档（Mac）：脚本 `/tmp/cdp-run17a/17b/17c`（停止、写文件、还原）、`run18/18b/19`（备份、恢复对话框、完成等待）、`run20a/20b`（白名单加/减）、`run21/21b/22`（封禁重试、封禁、解封）、`run23`（维护窗口）＋同名 `-out.json`；截图 `run17b-1`、`run18-1..4`、`run18b-1,2`、`run19-1`、`run20a-1,2`、`run20b-1`、`run21b-1`、`run23-1..3`。
+
+### 本轮新增真机证据（第三十九批：reaper 多节点实机 —— VM 克隆双节点验证）
+
+**零、装置**：`prlctl clone "CentOS Linux 9 Stream" --linked --name felis-node2`（linked 克隆，初始 1.4M）→ node2 改 host 名 `felis-node2`，停用/禁用 `felis-velocity`、`k3s.service`（server 形态）、docker，以 k3s-agent 加入同一集群（`https://10.211.55.6:6443`，复用 node1 node-token）。克隆副作用 = node2 自带 node1 storage 副本（6 项 / 3.1G）——已移开，模拟真实新节点。两节点均 Ready（v1.36.4+k3s1）；为克隆，node1 经历一次正常重启，重启后组件/服务全回归（docker 随自启后又停回 `inactive`）。
+
+**一、pin 正向**：`kubectl create job reaper-b39-ok --from=cronjob/felis-reaper`（live 模板原样，nodeSelector=`localhost.localdomain`；node2 无 taint、是合法调度候选）→ pod 落在 `localhost.localdomain`（nodeSelector 命中；backups PVC 的 affinity 亦指向同节点——两者本就该同节点），4s 跑通：`evaluated=2 reaped=0 warned=0 skipped=0 evicted=0 expired=0`、Job `Complete`。即：渲染出的 pin 在真实双节点集群把 reaper 钉在持盘节点。
+
+**二、错位 pin 反向对照**：同模板把 nodeSelector 改成 `felis-node2` → pod 停在 **Pending**，scheduler 事件原文：`0/2 nodes are available: 1 node(s) didn't match PersistentVolume's node affinity, 1 node(s) didn't match Pod's node affinity/selector`——node1 被错位 selector 拒绝、node2 被 `felis-backups` PVC 的 volume node affinity 拒绝（PV `pvc-0b4fbda6-…`，nodeAffinity=`localhost.localdomain`、hostPath=`/var/lib/rancher/k3s/storage/pvc-0b4fbda6-…_minecraft_felis-backups`）。**结论：本部署形态下 pin 写错是 fail-closed（卡住 + 明确调度事件），不会在无世界节点上静默执行**；真正要防的是首跑顺序（全新多节点安装时，第一次 reaper 运行会把 backups PV 落在其所在节点）——这正是渲染默认要求 `--reaper-node` 并 stderr 警告的原因，维持现状不修。
+
+**三、装置回收**：drill job ×2 删除（minecraft ns 无残留）；node2 关机 → `kubectl delete node felis-node2` → `prlctl delete felis-node2`（VM+克隆件删除）。收尾 `get nodes` = 单节点 `localhost.localdomain`，felis/游戏 pod 全 Running，CronJob `suspend=false` + nodeSelector 原样，docker `inactive`。重启副作用按既有说明处理：Mac 侧 443 入面板依赖的手工 `socat` 中继（本台账开头注记「重启 VM 后需重开」）随重启消失——已按原样重开（`TCP6-LISTEN:443,ipv6only=0,reuseaddr,fork TCP:127.0.0.1:30443`）并复核 op.console 面板 / api-me / player 面板皆 200。
+
+**四、留档**：node2 agent 上线日志（`k3s agent is up and running`、VXLAN subnet event 来自 10.211.55.6）；两向 job 的 pod/调度事件原文（见上）；`/root/node2-storage-copy/`（3.1G，随 VM 删除）。
+
+### 本轮新增真机证据（第四十批：Java 插件层收口 —— #63 demo-up 单起点、CI Java 门禁、插件面真机 E2E）
+
+**零、现场**：本批不改 Go/面板（控制面维持 `auditfix61`）；产出 = `deploy/demo-up.sh` 重写（`f5a76cf`）＋ `plugins/test.sh` 与 CI `plugins` 作业（`c59b387`）。三组真机验证（T1/T2/T3）在 VM 上针对该两文件跑完；CI 首跑即绿。
+
+**一、#62 复核：不成立（立案后剔除，未计为缺陷）**
+- 主张："fresh 非 demo 安装跳过 Velocity 插件构建 → proxy 静默不路由"。复核三条路径，**每条都构建 felis-velocity.jar**：
+  1. 源码臂：`felis-install.log`（首装）L1530 `building felis-velocity.jar`、L1553 `staged`；`bootstrap-auditfix{42,42b,42c,43,43b,44}.log` 每次重跑同两行。
+  2. 嵌入 tar 臂（fresh release/TUI 路径；本 VM 从未走过）：`felis bootstrap-assets game-stack | tar -x` → 37 文件（含 `plugins/velocity`、`plugins/shared`，构建输入齐全）→ `docker run --rm -v /tmp/gs62:/src:z -w /src/plugins/velocity gradle:8.14-jdk21 gradle --no-daemon clean build` → **BUILD SUCCESSFUL in 35s**，产出唯一 `felis-velocity-0.1.0.jar`（71117B，与现装同尺寸）。
+  3. 调用图：`build_velocity_plugin` 唯一调用点在 `build_game_stack` 末尾；`build_game_stack` 在 full 模式主链路（L2868）必经，nano 早返回，不存在可绕开的"demo 分支"。
+- 结论：代码阅读误判，剔除。真正会跳过插件构建的是 demo-up.sh 的旧镜像臂 → 即 #63（本批修复）。
+
+**二、#63（`f5a76cf`）：demo-up.sh 双起源 → 单起点（129 → 63 行）**
+- 红证据（读取 + 真机口径核对）：① 版本分叉——demo-up 硬编码 `PAPER_MC_VERSION:=1.21.8`，bootstrap `resolve_game_jars` 从 Limbo CI 产物名推导（T3 实测当日 Limbo 2026.0.3-ALPHA / MC 26.3；同一登录两跳必须同协议）；② 其镜像臂从不构建 felis-velocity.jar，导入的是本地 tag（`felis-limbo:demo` 等），与 bootstrap 写入的 registry refs 不一致 → 死件、旧基座上可致"起了但无处路由"；③ 起 docker 后从不停（违反 13d64e0 起"构建方以 docker 停收尾"的约定）；wiring 段在现代基座上只是 no-op。
+- 修复后 = bootstrap（未 SKIP 时）＋ 三项硬校验（`felis.host.toml`、`[velocity]` 段、`felis-velocity.jar`）＋ `felis setup` 交棒；镜像逻辑全删。
+- 真机验证：**T1** 健康基座（`SKIP_BOOTSTRAP=1 SKIP_SETUP=1`）→ 通过、exit 0；**T2** 移走 jar → `ERROR: /opt/felis/velocity/plugins/felis-velocity.jar missing — … silently routes nothing …`、exit 1（随即还原 71117B root:root 0644）；**T3** 默认臂全量重跑（`FELIS_SKIP_FETCH=1 FELIS_INSTALL_MODE=full FELIS_IMAGE=…:auditfix61 FELIS_WORLDS_HOST_PATH=… SKIP_SETUP=1`，日志 `/root/demo-up-b40.log`）：L733 插件构建、L752 `staged`、L941 交棒行、`[fail]`=0；收尾态 `felis-velocity active / k3s active / docker inactive`。
+- 附带实录（同一日志窗）：重跑期间 felis-api 滚动，velocity 00:22:25/00:22:40 两次 `refresh failed … keeping current registrations.`（失败保旧——规格 §11 承诺的真机上演）→ 00:22:55 恢复注册 → 00:22:58 服务随 `install_velocity` 重启并重载插件（新 pid）→ 00:22:59 `routing ready` → 00:23:29 收敛到 pod IP。
+
+**三、Java 层 CI 门禁（`c59b387`）：三个手工测试 + 三个装机 jar 首进 CI**
+- 缺口：`plugins/*/test` 三个 main 测试从未被任何 build/CI 运行；velocity/paper/limbo 三个"装机即用"jar 只在 bootstrap/Dockerfile 编译（Go CI 从不碰 Java）。
+- 产出：`plugins/test.sh`（Maven Central 取 adventure 三 jar，pinned+sha256；三测试 javac+java；三生产编译，limbo 按 bootstrap 同源解析当日版本）+ `ci.yml` 新增 `plugins` 作业（temurin 21 + Gradle 8.14）。
+- 首跑两次真跑抓出两处"从未被跑过"的假设错误：① `InviteCardTest` 一行式缺 examination-api（adventure-api 4.26.1 的 `Component` 签名引用 `Examinable`，javac 编译期即需）；② limbo 的 `com.loohp:Limbo:+` **永远不可解析**（LOOHP 仓无 maven-metadata，404 实查）→ 裸 `gradle -p plugins/limbo build`（plugins/README 原文）从来不可行；二者均已修（脚本 + 该测试 javadoc + README）。
+- 验证：VM 容器三测试 `OK (32/36/48 checks)`；velocity 21s / paper 29s / limbo 13s（2026.0.3-ALPHA）全 `BUILD SUCCESSFUL`。GitHub Actions run `35888009965`（c59b387）= go/shell/panel/plugins **4/4 success**（`plugins` 作业首跑即绿）。
+
+**四、插件层真机 E2E（不依赖真实客户端的面）**
+- 自建纯 socket 探针 `/tmp/mcprobe.py`（status = 服务器列表 ping；login = 登录首包），重装前与重装后各跑一遍，结果一致：
+  - 子域 MOTD（§11 只读缓存 + 相位）：`test-one` → `« test-one » 休眠中，加入即唤醒 / sleeping — join to wake`；`lobby`/`login` → `« … » 在线 / online`；`resolvecheck` → 休眠中；`nosuchxyz.<root>` → 回落 `A Felis server`（非 felis 子域不被劫持）。
+  - 登录边界：`LoginStart`（含现代协议 UUID 字段）→ 服务端首包 `0x01 EncryptionRequest` ⇒ 边缘 online-mode 强制成立。
+  - 加载面：`Loaded plugin felis-link 0.2.0`（共 5 插件）。
+- 口径：真实账号进服（limbo 门 → 菜单 → 转服）仍按"进服跳过"决定不演练；以上为不依赖客户端的最大真机面。（首次探测因探针漏发 UUID 字段被静默关闭——探针缺陷，非服务端问题；补齐后一次通过。）
+
+**五、观察（不计缺陷）**
+- ViaVersion 自报有 5.12.0（当前 5.11.0）：bootstrap 的 pin 是 FL-007 实测版本，属刻意，不随提示升级。
+- limbo 插件编译期一条 deprecated API 提示（`FelisLimboPlugin` 用/覆写已弃用 API）：门禁下可见、不阻塞，留观察。
+
+### 本轮新增真机证据（第四十一批：装载器 mod 层收口 —— #64 exec 位、#65 元数据、编译+起服双门禁）
+
+**零、现场**：控制面与面板不动；产出三个 commit：`f6048f2`（#64 修复）、`c2fe6a6`（#65 元数据）、`aa5abfa`（`plugins/test-mods.sh` + CI `mods` 作业 + release 双门禁 + README 记录）——这是最后一块"从未被任何自动化碰过"的模块面（fabric / forge / neoforge 三个装载器 mod）。
+
+**一、#64（`f6048f2`）：三个 vendored `gradlew` 缺可执行位——文档正路第一步即失败**
+- 红证据：`git ls-tree origin/main` 三个 `gradlew` 全为 `100644`（blob `b9bb139f…`）；`plugins/README.md` 第 212–214 行原文教 `plugins/fabric/gradlew -p plugins/fabric build`（forge/neoforge 同式）；VM 全新 clone 实跑 → `bash: line 1: ./gradlew: Permission denied`，exit 126。
+- 为何从未暴露：无 CI、无安装路径，README 是唯一入口——直到本批第一次真跑才现形。
+- 修复：`chmod +x` + `git update-index --chmod=+x`（mode 100644→100755 ×3）。
+- 可达性：①（全新用户照文档走的第一步）。
+
+**二、#65（`c2fe6a6`）：mod 元数据与仓库 LICENSE 相悖**
+- 三处 `license = "MIT"`（`fabric.mod.json`、forge/neoforge 的 `mods.toml`）vs 仓库 `README.md:73` 的 `AGPL-3.0-only` + LICENSE 全文。时间线：mods 2026-06-26 加入（`93f143f`）、LICENSE 2026-07-12 才落地（`037eb24`）——陈旧残留。
+- 另两处 `issueTrackerURL = "https://example.invalid/felis"`（占位域名）→ `https://github.com/FelisMC/Felis/issues`。
+- 用户面：fabric loader 启动打印 license 字段；`mods.toml` 被两个 loader 解析。
+- 可达性：②（低；元数据/合规面，非功能）。
+
+**三、双门禁落地（`aa5abfa`）**
+- `plugins/test-mods.sh`：JDK 17 下依次跑三个模块的 vendored wrapper（`./gradlew --no-daemon build`）；java 大版本非 17 直接 fail-loud（三模块目标是 Java-17 的 Minecraft 线）。
+- `ci.yml` 新增 `mods` 作业（temurin 17 + `gradle/actions/setup-gradle@v4`，版本由各模块 wrapper 自管）；`release.yml` 发布前加两道 Java 门禁（JDK21 `plugins/test.sh` + JDK17 `plugins/test-mods.sh`）。
+- README：Status 更新为 compile + boot 双验证；Building 段补 limbo `-PlimboVersion=<release>` 要求与两个门禁说明。
+- CI：run `35892544563`（`aa5abfa`）——`mods` 作业首跑即绿，五作业全过（shell / mods / panel / go / plugins 5/5）。
+
+**四、三个模块的真机编译 + 起服 E2E（本批核心证据）**
+- 编译（VM 容器 `eclipse-temurin:17-jdk` + 持久 gradle 缓存）：三模块 `BUILD SUCCESSFUL`（首跑 4–5 分钟级/模块），产物 `felis-{fabric,forge,neoforge}-0.1.0.jar` = 29276 / 28855 / 28562 B；留档 `/root/mods-build-b41.log`。
+- 起服（三个真实专用服，控制台直驱 `/link`，随后 `linkx` 未知命令对照、`stop` 正常停服）：
+  - **fabric** 1.20.1 + loader 0.19.5 + fabric-api：`Felis link ready; /link is registered.` → `Done (10.670s)!` → `/link 只能由玩家执行 / /link can only be run by a player.` → `Unknown or incomplete command`（linkx）→ `Stopping server`；`EXIT_fabric_boot=0`。
+  - **forge** 1.20.1-47.3.0：`Felis link ready; /link will be registered.`（modloading-worker）→ `Done (11.751s)!` → 同四段证据；`EXIT_forge_boot=0`。
+  - **neoforge** 20.4.251（1.20.4）：`Done (16.094s)!` → 同四段证据；`EXIT_neoforge_boot=0`。
+- 口径：三服均为"装好 mod 后真启动"的场景，同时验证 loader 真实解析修改后的元数据（license=AGPL、tracker URL）后正常装载；完整取码链（玩家在游戏内 `/link`）按"进服跳过"决定不演练，装载/注册/拒绝面已全部真机成立。
+- 留档：`/root/mods-e2e-b41.log`、`/root/mods-e2e-b41-resume{2}.log`、`/opt/felis/mods-e2e/`（496M，三服目录保留）；收尾态：docker `inactive`、k3s/felis-velocity `active`、磁盘 17G free。
+
+**五、演习装置自身两次修正（非产品缺陷，如实记录）**
+- 控制台驱动脚本 `/root/modserver-drive.sh` 初版把 FIFO 写端先开（`exec 3> console`）→ 自我死锁（内核 `wait_for_partner`、`State: S`），java 从未启动（无 `server.out` 实证）；改 `exec 3<> console`（RDWR 打开不阻塞）后三服全通。
+- forge 安装器首跑：`libraries.minecraft.net` 连接失败 → `com.google.code.findbugs:jsr305:3.0.2` 下载失败、`There was an error during installation`、无 `run.sh`；原样重试即 `The server installed successfully`（VM 网络抖动，非 forge/产品问题）。
+
+**六、观察（不计缺陷）**
+- 无新增。
+
+### 本轮新增真机证据（第四十二批：覆盖面对账 + 文档/工具收尾 —— #66 README_EN、§28 图对齐、sync.sh 本地修复）
+
+**零、现场**：本批不碰产品代码；对"所有模块均已演练"的结论做**独立对账**（枚举仓库全部模块/交付物 × 台账覆盖），只挖到文档与本地工具层面的残留。commits：`62c5a8a`（#66 README_EN）、`62e8c87`（§28 序列图对齐）。
+
+**一、覆盖面对账（本批核心，逐项）**
+- 仓库卫生：`git ls-files` 全量扫描无 `.DS_Store`/`*.log`/`node_modules`/构建产物误入库；顶层 `felis` 二进制（83MB）被 `.gitignore` 正确忽略；旧仓库名 `MliroLirrorsIngenuity` 残留仅存在于台账（历史记录）与本地未跟踪二进制。
+- 未竟工作标记：`TODO|FIXME|XXX|HACK` 全仓（go/ts/sh/md/yml）仅 1 处命中——`internal/api/api_test.go:263` 的 `context.TODO()`（合法测试写法），无未完成实现标记。
+- CLI 派发面：`cmd/felis/run.go` 的 usage ↔ dispatch 表有双向测试强约束（`run_test.go`，含显式 `undocumentedCommands` 允许表：`bootstrap-assets`/`init-forwarding` 等 in-Pod 入口），机制已核实成立。
+- CI 覆盖：`shell` 作业按 shebang 对**全部 tracked `*.sh`** 做语法检查（`git ls-files '*.sh'` 逐文件 `bash -n`/`sh -n`）并运行 `deploy/bootstrap_test.sh`；`plugins`/`mods` 各自作业。无"存在但从无入口运行"的脚本。
+- `internal/*` 22 个包逐一对账：0 提及者（`internal/apis`）为 CRD 类型包（被 40+ 站点引用、随 operator/manifests 演练间接全覆盖）；`deploy/{limbo,lobby,paper}` 为三张游戏镜像定义，随每次起服演练执行。真盲区＝无。
+- `panel` 全部 17 条页面路由 × 既有批次（CDP 巡检、深挖、写流程）覆盖；`scripts/` 三脚本为**本地私有**（`.git/info/exclude`，不入库）。
+- `docs/` 四件：troubleshooting（多批引用）、openapi（含 parity 强制测试）、deferred-seams（内容为现行状态，含 CLOSED/verified-live 注记，抽查与代码一致）、sequence-diagrams（发现漂移 → 修复，见第三节）。
+
+**二、#66（`62c5a8a`）：README_EN 与中文 README 漂移**
+- 红证据：`README.md`（中文）"使用方式"含**必需**的私有仓库 workaround（仓库私有 → 裸 raw 命令 404；token 经 `curl --config -` 由 stdin 下发、不进 argv；`sudo -E` 传给安装器）与"重跑=升级 felis-api、通道不继承（跟 main 需 `FELIS_VERSION_BOOTSTRAP=dev`）"两段；`README_EN.md` 是 `README.md:6` 链接的英文入口，这两段**完全没有** → 英文读者照文档第一步就 404 且无指引。
+- 修复：两段译入 EN（保留原命令与语义；`grep` 校验 token/config/dev 三关键串在位）。可达性：①（轻）——英文首装路径第一步。
+
+**三、§28 序列图对齐（`62e8c87`）**
+- 漂移点（两处，与修复后代码实读对照）：① Claim Transaction 图注仍是 audit #4 前形态（`SELECT EXISTS` + 事务外 pre-check）→ 更新为现行实现：`pg_advisory_xact_lock(user_id)` + 行 `FOR UPDATE` + **事务内**四维配额门（图上 `QuotaAvailable` 只是 fast path）；② Link Flow 更新为：取码 `FOR UPDATE`、同用户重验幂等、退役（soft-deleted）账户链接被当场接管——409 仅对"其他活跃用户"。均以 `internal/api/pgrepo.go` 实读 + handler 映射核对。
+- 惯例依据：该文件有维护史（`676407d docs(diagrams): align §28 sequence diagrams with implemented routes`）。
+
+**四、本地工具修复（私有脚本，不入库）：`scripts/sync.sh` 排除 `plugins/` 与 `go:embed` 冲突**
+- 红证据（本机复现）：以 sync.sh 原排除清单 rsync 一个 fresh 目标 → `go build` 立即失败：`bootstrap_asset.go:30:12: pattern plugins/limbo/build.gradle: no matching files found`（`bootstrap_asset.go` embed 了 plugins/{limbo,paper,velocity,shared} 源码；排除清单成文于 embed 引入之前）。存量目标靠 rsync 对排除路径的"保护"而看似正常，fresh 目标必炸。
+- 修复：排除项从 `plugins/` 改为 `plugins/*/build|.gradle|bin/`（与 `.gitignore` 同义）；`watch.sh` 同步去掉 plugins 的忽略。验证：新清单 rsync fresh 目标 → `go build` 绿（83.7MB 二进制产出）；两脚本 `bash -n` 通过。因脚本被 `.git/info/exclude` 私有，修复留在工作机、无 commit。
+
+**五、release.yml 静态复核（该流水线从未执行过——仓库当前无任何 tag/release）**
+- 三处输入/输出契约实核：① 发布资产名 `felis-linux-amd64/arm64` ↔ `deploy/bootstrap.sh` 的 `asset="felis-linux-${arch}"` 及收敛检查 `felis ${FELIS_REF}`（版本首行契约两侧一致）；② `out/linux_amd64/usr/local/bin/felis` ↔ Dockerfile 最终层 `COPY /out/felis /usr/local/bin/felis`；③ 版本断言 `felis ${GITHUB_REF_NAME}` ↔ `cmdVersion` 的 `felis %s` 首行。
+- 唯一无法在本机完全复刻的 `file(1)` 机器类型断言：用交叉编译的 linux/arm64 二进制实测输出 `ELF 64-bit LSB executable, ARM aarch64, …`，断言串 `ARM aarch64` 匹配（BSD/Ubuntu file 同源 magic）。
+- 结论：整条流水线未跑属"尚未切 tag"的发布流程事实（第四十批已注记）；切首个 tag 前无待修项。
+
+**六、观察（不计缺陷）**
+- `internal/api/handlers_account.go:169` 的 reclaim 选择仍为 Java/Velocity 侧 CODE-ONLY（deferred-seams 已记 accepting）——维持。
+
+### 本轮新增真机证据（第四十三批：构建 Job 收尸 #67 + #68 上下文拉取重试 + 生命周期/故障注入压测）
+
+**零、现场**：控制面从 auditfix61 升到 **auditfix62**（含 #67；自建镜像推入内建 registry，api/operator/reaper 三处滚动，`felis version` = `v0.0.0+fix62`）。本批证据 = 6 轮起停循环 + 3 处故障注入 + PG 断连语义 + 并发风暴 + 持续轮询。
+
+**一、#67（`2755e41`）：build Job 从不回收——每构建一个、完成 Pod 永久堆积**
+- 红证据（真机）：`felis-build` 里 5 个完成 Job/Pod 最长 26h 无人回收；全仓 TTL 对照——fileedit 2m / backup 10m / restore 10m / reaper CronJob 3+3 历史，**唯独 build 没有 `ttlSecondsAfterFinished`**；且 `build.go:69` 的注释写着 "(e.g. GC'd); treated as failed"——预期的 GC 从未存在。危害：完成 Pod 计入节点 pod 预算（stock k3s 110），构建量一上来先把节点塞满，新构建全部 Pending；etcd/磁盘同步膨胀。
+- 修复：`buildJobTTL = 7 * 24h`（终态后计时；日志路由是失败分诊面故取长窗；`Sync` 对终态幂等、`JobUnknown` 只影响非终态 → 删除后的唯一代价是日志 404）。单测 `TestBuildJobIsReapedAfterCompletion`。
+- 真机验证（auditfix62）：① 真实路径触发构建（owner 会话 → `POST /api/v1/images/build`，借遗留 blob 作 context）→ 新 Job `ttlSecondsAfterFinished=604800`、构建 25s `succeeded`；② 对旧 Job 打 `ttl=30s` → 40s 内 Job+Pod 被 TTL 控制器收走（机制实证）；③ 旧堆积 5 个里 1 个已收、4 个留存对照。
+
+**二、生命周期 ×6 + 三处故障注入（operator / api / postgresql）**
+- 每轮：CR `desiredState` Running → 等 Running → Stopped → 等 pod 消失。结果：**6/6 全收敛**——Running 23–29s（注入轮与无注入轮无差）、Stopped 3s、pod 每轮如期消失；收尾 conditions 无 Failed 残留（Ready=False / RconReached=False = Stopped 正常；Provisioned=True）。
+- 注入 1（cycle 2，删 operator pod @t+2s）：Running 仍 23s 达成（新 operator 立即接管 reconcile）。
+- 注入 2（cycle 4，`systemctl restart postgresql`）：CR 路径无感；另做定点验证：PG 停机窗口 `/me` = **503**（非 401，#11 语义保持）、`/healthz` = 200（存活探针独立）、内部 status = 200（纯集群读）；PG 恢复后 `/me` = 200。
+- 注入 3（cycle 5，删 api pod @t+2s）：内部轮询出现 2 次 `http=000`（约 6s 窗口）后自愈；Running 28s 达成。
+- 观察（不计缺陷）：高频翻杆期间 operator 报 9 条 `Operation cannot be fulfilled ... object has been modified`（乐观锁冲突 → 重排队自愈），目标时间无差；controller-runtime 正常重试语义，生产低频操作下更罕见。
+
+**三、并发风暴 + 拒绝面 + 持续轮询**
+- 20 路并发 internal status + 10 路 `/me`：30/30 全 200。
+- 无主 + ownerOnly 服 wake：403 `forbidden`（正确拒绝面，非 5xx）。
+- 持续轮询（status + healthz + me，5s 一轮 × 300 轮 ≈ 25 分钟，900 样本）：收盘 `LONG POLL DONE fails=27`——27 个失败样本**全部**落在三个自导演练窗口内（02:24:15/20 红证据杀 api〔6〕、02:29:48/53 auditfix63 滚动升级〔6〕、02:30:53–02:31:13 #68 确定性复现 scale api→0〔15〕），窗口外 873 样本全 200、零自发失败；每个窗口在动作结束后 ≤1 个探测周期（~5s）内恢复。
+
+**四、留档与收尾**
+- 留档：`/root/soak43.sh` + `soak43.log`、`/root/poll43-long.sh` + `poll43-long.log`、`/root/mint-owner-43.sh`（owner 会话重铸）、`/root/felis-image-build43.log`（镜像构建）、`/tmp/owner-jar43.txt`（owner cookie）；镜像 `10.43.182.43:5000/felis/felis:auditfix62` 已入 registry（`e2e/ttl-probe:latest` = 本批 drill 镜像，保留）。
+- VM 状态：docker 用毕已停（inactive）、k3s/velocity active；控制面三处 = auditfix62。
+
+**五、#68（`b76d0ac`）：构建 context 拉取对控制面重启零容忍——一次拒连即终态失败**
+- 红证据（真机）：删 api pod 后触发构建 `bld-1790187851749257160` → Job `Failed`；pod 时间线 `18:24:12Z` 起、context-fetch `18:24:13Z` exit 1，日志原文 `dial tcp 10.43.237.249:8081: connect: connection refused`；Kaniko 从未启动（PodInitializing）；新 api pod 11s 后就绪、同 blob 前后各一次构建均 25s succeeded——纯粹"单次尝试"造成的无谓失败。
+- 修复（`b76d0ac`）：`fetchContextWithRetry`——传输错误/5xx 重试至 45s 窗口（3s 间隔），4xx 快速失败（是答案不是抖动）；URL 校验提前为 usage error(2)；保留原错误文案前缀。窗口/间隔为 vars（测试可缩窗）；4 例单测：5xx 后恢复、拒连后恢复、404 不重试、窗口耗尽（全绿，含 `-race`）。
+- 真机验证（auditfix63，确定性优先，不复刻竞态）：`scale api→0` → 以复刻 Job 直建（由红证据 Job 派生：换名 `build-retry68`、剥 controller 标签与 selector、镜像改 auditfix63，复用 `felis-service-token` Secret）→ fetch 连续 8 次 `connection refused; retrying`（实捕日志）→ `scale api→1` → fetch 恢复、Kaniko 构建推送、trivy 干净 → **Job Complete**（`18:30:52Z→18:31:24Z`，32s）；同形场景对照旧版 = 终态 Failed。正常路径回归：api 在线直构 `bld-1790188306964162960` → `succeeded`（10s），fetch 日志**零重试行**（重试不引入正常路径开销）。
+- 附带发现（运维）：本批升级沿用 `set image` 后 `FELIS_IMAGE` env 仍停 `auditfix61`（两代落后）——构建 Job 的 fetch 容器实际一直在用旧镜像；已把 api/operator 的 `FELIS_IMAGE` 与 api/operator/reaper 三处镜像全部对齐 `auditfix63`。真实安装器路径随 manifests 重渲注入该 env，手工升级须成对改（升级清单事项）。
+
+### 本轮新增真机证据（第四十四批：modpack 规模上下文全链 + 并发构建）
+
+**零、现场**：控制面 = auditfix63（含 #67/#68；api/operator `FELIS_IMAGE` 同版本），VM docker 停。本批目标 = 把"真实模组包大小"这条链压到生产尺度：此前所有提交/构建演练的上下文都是 KB 级。
+
+**一、200 MiB 上下文全链（零缺陷）**
+- 构造：5×40 MiB `/dev/urandom` 载荷 + Dockerfile（随机数据不可压缩 = 最坏情形），tar.gz = 209,750,275 B。
+- 上传（真实 app 路由）：`POST /api/v1/me/submissions/{id}/context`（owner 会话）→ 200，**0.44s / ~481 MB/s**；宿主核对 PVC 落盘 `sub-cc160b3ad19080f9/context.tar.gz`。服务端读写超时设计（ReadTimeout/WriteTimeout 有意不设、仅 ReadHeaderTimeout）与流式落盘（io.Copy → cappedReader）均按预期，无内存尖峰。
+- 批准 → 构建（`bld-1790189227767686191`）：fetch init **1s**（200MB 内部面拉取 + 解压，无重试行）、kaniko **4s**（unpack/COPY/snapshot/push gzip 层 **209,738,853 B**）、trivy **4s**（0 findings）；Job Complete，全链 13s（18:47:07Z→18:47:20Z）。registry manifest 实查层大小一致（推送非虚）。
+- 覆盖点随验：提交构建 Job `ttlSecondsAfterFinished=604800`（#67 对真实提交路同样生效）；`/me/submissions` `build_status=succeeded`（#36 面在 200MB 规模下仍成立）。
+- 资源采样（2s 粒度，构建太快可能漏尖峰，标注局限）：无 OOM——api 17.9MB / kaniko 33.4MB / registry 34.1MB / trivy 35.9MB；emptyDir+快照为临时占用，随 Job TTL 回收。
+- 1 GiB 帽的单测已有覆盖（`internal/submit/submit_test.go` oversize→ErrInvalid）；帽下真机 = 本条（帽上真机不划算，不做）。
+
+**二、并发构建 ×3（零缺陷）**
+- 三路同时 `POST /api/v1/images/build`（各自 ref）→ 202×3；三个 Job 均 **11s Complete**，无串扰：每个 Job destination 与 tag 一一对应（conc-b44-1/2/3 各落位）、`ttl=604800` 各自在。
+
+**三、磁盘运维注记（非缺陷）**
+- 连续镜像构建（docker build）把 VM 根盘从 11G free 压到 4.3G；`docker builder prune -af` 一键回收 **6.5G**（回 11G free）。即每轮全量重建约需 5–7G 构建缓存空间；生产主机根盘建议 ≥40G 并定期 prune（如需可在部署文档补一行，本批未改码）。
+- 留档：`/root/bigctx.tar.gz`（200MiB 上下文原件）、`/root/bigctx/`、`/root/stats44.log`、`/root/stats44-build.log`；提交 `sub-cc160b3ad19080f9`（blob 与 `user-uploads` 镜像保留作尺寸样本）；镜像 `e2e/conc-b44-{1,2,3}:latest`。
+
+### 本轮新增真机证据（第四十五批：模组构建链三连修 #70/#71/#72 + 提交→装服全链闭环 + 文档/测试修 #69/#73/#74）
+
+**零、现场**：控制面 auditfix63→66（每修一版真机重演）；本批目标 = 真实用户会写的那种模组包 Dockerfile（`FROM <平台底座>` + 内容层）。此前所有构建演练都是 `FROM scratch`——用户真实形态的第一步从未跑过，本批连撞三个缺陷。
+
+**一、#70（`ac403b9`）：kaniko 拉底座缺 pull 侧 insecure 标志**
+- 红证据（真机）：提交 `FROM registry.felis.svc:5000/felis/paper:demo`（`bld-1790189685537480076`）→ kaniko `Retrieving image manifest …` 后即败：`Get "https://registry.felis.svc:5000/v2/": http: server gave HTTP response to HTTPS client`。根因：Job 只给了 `--insecure --skip-tls-verify`——kaniko v1.24 `--help` 实测二者均只覆盖 **push**，拉取默认 HTTPS，撞上明文 registry。
+- 修复：补 `--insecure-pull` / `--skip-tls-verify-pull`（对称补齐；构建命名空间 egress 本就只许 DNS/内建 registry，不扩大可达面）。单测断言两 flag 在参。
+- 真机复验：下一次构建 `Retrieving image manifest` 不再报错，进入解包阶段（随即暴露 #71）。
+
+**二、#71（`6e47730`）：kaniko 解包底座层被 drop-ALL 卡死**
+- 红证据（真机）：`error building image: error building stage: failed to get filesystem from image: chown /etc/gshadow: operation not permitted`——kaniko 以 root 解包 tar 层要把文件 chown 到层里记录的属主（root:shadow 等），drop-ALL 后 CAP_CHOWN/CAP_FOWNER/DAC_OVERRIDE 全无。`FROM scratch` 全用 COPY（属主= kaniko 自己）所以从未暴露；**任何真实底座**必触。
+- 修复：仅 kaniko 容器补回最小能力集 `CHOWN + DAC_OVERRIDE + FOWNER`（fetch/trivy 保持 drop-ALL 基线；单测断言"恰好三枚"防漂移）。
+- 真机复验：解包通过、COPY、推送成功（进入 trivy 阶段，撞上 #72）。
+
+**三、#72（`b14bacb`）：trivy 扫 jar 必拉 Java DB——egress 锁下必败**
+- 红证据（真机）：kaniko 全绿后 trivy `FATAL … Unable to initialize the Java DB … failed to download artifact from mirror.gcr.io/aquasec/trivy-java-db:1: connection refused`。Java DB 按需下载：镜像一有 jar 就触发——模组包 = jar 集合，故所有真实用户构建都会倒在扫描门；旧构建全 scratch（无 jar）从未触发。
+- 修复：新增 `[registry] trivy_java_db_repository`（→ `--java-db-repository`，与 vuln-DB 旋钮对称）；installer carry 白名单收录；§8e 配方补 Java DB 镜像步骤；deferred-seams 更新。
+- 真机复验（第 4 次尝试，auditfix66）：Job spec 实测含 `--db-repository registry.felis.svc:5000/mirror/trivy-db:2 --java-db-repository registry.felis.svc:5000/mirror/trivy-java-db:1`；扫描 `ubuntu 26.04` + `paper/paper.jar` + `pebble` 全 0 漏洞；**Job Complete**（全链 27s，`bld-1790190788310114143`）。
+
+**四、capstone：提交 → 装服 → 起服，全链闭环（零缺陷）**
+- 第 4 次尝试产物 `user-uploads/sub-c006cbd633317ccb:latest`（= `felis/paper:demo` + 用户 marker，经完整提交管道构建）：
+  - 白名单自动收录（`source=built`）→ 产品路由 `PATCH /api/v1/servers/test-one {"image": …}` **200**（`image_not_whitelisted` 门通过）；
+  - `POST /servers/test-one/wake` 202 → pod `test-one-0` Running 1/1（25s 内）；
+  - `kubectl -n minecraft exec test-one-0 -- cat /felis-probe-marker.txt` → **`felis-probe44-ok`**（用户构建上下文的内容确在运行中的服务器里）；
+  - 服务器日志 `Done (4.678s)!`（paper 完整启动）；RCON 线程应答平台就绪探针；
+  - 收尾：stop 202、pod 消失、镜像回滚 `felis/paper:demo`、desiredState=Stopped。
+- 口径：这是"玩家上传模组包 → 服主审批 → 自动构建 → 选用为该服镜像 → 起服"的机制全链（真实客户端进服仍按决定跳过）。
+
+**五、#69（`5cbfa89`）：README 过度承诺"自动部署"**
+- 复核（读全）：提交数据模型无目标服务器字段；approve→build→白名单是唯一自动化；部署 = 服务器编辑框选镜像（产品路由与 UI 均在）；面板文案本身只承诺"自动触发安全构建"。
+- 修复：README 中英两行改为"自动构建；产物进入镜像白名单，可直接选用为服务器镜像完成部署"。可达性 ①（轻）。
+
+**六、#73（`2961beb`）：bootstrap 测试在跑 k3s 的主机上必假失败（工具级）**
+- 红证据：VM（真实 k3s 主机）跑 `bootstrap_test.sh`（**原版同现**）→ `FAIL a missing worlds root is warned about`：用例拿真实路径 `/var/lib/rancher/k3s/storage` 期望 WARN，而该目录在"安装器真正要服务的机器"上必然存在 → 假红；CI 从未见到（runner 无此路径）。
+- 修复：warn 情形改用保证不存在的路径；并把 `trivy_db_repository` / `trivy_java_db_repository` carry 断言补齐。VM 复跑 **140 PASS / 0 FAIL**。
+
+**七、#74（`96aa817`）：CI flaky——console 断连测试读写竞态（工具级）**
+- 红证据：run `35907662213`（纯 README 提交）go 作业红：`TestServerConsoleDisconnectTeardown: expected the first event before disconnect, got ""`；`gh run rerun --failed` 即绿 → 竞态确认。
+- 根因：测试只等"源读到第一行"就 cancel，"中继把事件写入响应"尚未发生；cancel 落进缝里时 body 为空。
+- 修复：加 `firstDataWriter` 信号（写完成后再 cancel，写→读有 happens-before）；本地 `-count=60` 与 `-race ×5` 全绿；CI 复跑 `5cbfa89` = success。
+
+**八、运维注记（非缺陷）**
+- 升级滚动撞 kubelet **ephemeral-storage 压力**：auditfix64 滚动时新 api/operator pod `Pending 4m46s`（事件 `untolerated taint(s)`），02:58:12 kubelet eviction manager 回收后自动调度成功。诱因 = 构建把根盘压到 89%（docker 构建缓存 + kaniko/emptyDir 临时层）；`docker builder prune -af` 两清共回收 ~11G，余 11G free。生产清单：根盘 ≥40G + 升级前清构建缓存（与批 44 注记合并）。
+- 中间失败构建留下的 registry 镜像（`user-uploads/sub-83f6…`、`sub-171e…` = 已推未过门；`sub-c006…` = capstone 产物）留档；测试服 `test-one` 已回 `felis/paper:demo` + Stopped；docker 停。
+
+### 本轮新增真机证据（第四十六批：上传面 #75/#76 真机闭环 + tracker #8/#1 收口 #77/#78 + 文案 #79 + 发布链 rc smoke）
+
+**零、现场**：代码三条 commit 先落（`43699b4` #77 面板跳转、`c57daaf` #78 converge、`b9ebc87` #79 INERT 文案），随后按 `b9ebc87` 重建镜像 = `registry.felis.svc:5000/felis/felis:auditfix77`（`v0.0.0+fix77`；宿主源 `10.43.182.43:5000`，docker build → push），控制面 api/operator + minecraft ns `felis-reaper` CronJob + api/operator 的 `FELIS_IMAGE` 四处成对齐；`/opt/felis/src` = b9ebc87 快照（上一版 `src.bak46`）；宿主 drill 二进制 `/root/felis-fix77.bin`（Mac 侧 `GOOS=linux GOARCH=arm64` 交叉编译，`-X main.version=v0.0.0+fix77`）。
+
+**一、#75 真机闭环（配额/限流，owner 会话，auditfix77）**——留档 `/root/probe77/create-lane.log`、`lane2.log`：
+- **create 失败不烧窗口**：坏 JSON → 400，**同一秒**合法 create → 201（`sub-811427aa7dcf6807`）——`release` 路径生效；
+- **create 冷却**：同用户 30s 内第二条 → 429 `submission_cooldown`；
+- **pending 上限**：攒到 5 条 pending → 第 6 条 → 403 `submission_quota_exceeded`；
+- **upload 失败不烧窗口**：对已审核行上传 → 409 `already_reviewed`，**同一秒**对 pending 行上传 → 200；
+- **upload 冷却**：15s 内第二次 → 429 `submission_cooldown`；等 16s → 200；
+- **存储预算**：把 S2 的 blob 稀疏 `truncate` 到让 owner 已存字节 = 2GiB−100B（`blocks=8`，不占盘）→ 上传 → 403 `submission_quota_exceeded`；管理员删除 S2（行 + 目录双清）后 → 同一上传 → 200（预算即时释放）。
+- 口径：预算按 blob 的实际占用聚合（`Blobs.Size` 遍历汇总，正是"用久了就超"的同一条读取路径）；稀疏垫付只是把"已经存了 2GiB 的用户"这一状态合成出来。
+
+**二、#76 真机闭环（撤回 + 管理员删除）**——留档 `/root/probe77/lane3.log`：
+- 撤回自己 pending → 200；DB 行 = 0、uploads 目录 = gone；重复撤回 → 404；对已审核 → 409 `already_reviewed`；对他人 id → 404 `not_found`（owner 判定先于状态判定，不泄露他人行状态）；
+- 管理员删除（pending）→ 200；重复 → 404；行与 blob 双清（lane2 的 S2 同证：`s2 rows=0 dir=gone`）；
+- 面板侧 CDP：**撤回两步确认**（展开行 → 「撤回提交」→「确认撤回」→ 行消失）与 **admin 队列每行两步删除**（trash →「确认删除」→ 行消失、计数回落）双双走过；截图 `/tmp/withdraw-{1,2}-*.png`、`/tmp/admdelete-{1,2}-*.png`（Mac），脚本 `/tmp/cdp-withdraw76.js`、`/tmp/cdp-admdelete76.js`。
+- 真机注记（非缺陷，分级 ZT 的对照实证）：admin 面只在 `op.console.<root>` 主机可用——同一 owner 会话在玩家面板主机上 `is_admin=false`（admin 路由 403），在 `op.console.<root>` 上 `is_admin=true`（200）；面板按要求显示「无权访问」而不是假装能点。
+
+**三、#77（tracker #8，`43699b4`）真机闭环：锁定会话 → /setup**
+- 铸一个未完成引导会话（SQL：`sha256('drill-locked-77')` → `sessions`，用户 `3f2c1b0a-…`，`email_verified=f`、无 passkey）→ `GET /me` 200、`GET /me/submissions` → 403 `setup_required`（后端本就正确）；
+- CDP 用该 cookie 打开 `/submissions`：页面请求 `/me/submissions` 收 403（code=`setup_required`）→ **自动跳转 `/setup`**（`location.href` 实测）→ 向导渲染「初始化你的账户 / 第一步 · 填写邮箱」（截图 `/tmp/setup8-redirect.png`；网络事件 = 200/403(/setup/status 200) 序列在脚本输出里）。
+- 顺带确认：Dashboard 首屏三请求（`/me`、`/me/servers`、`account/link/start`）都是 `SetupAllowed`——所以补丁前用户是在"点受保护操作"时才撞到那句误报「无权执行此操作」。
+
+**四、#78（tracker #1，`c57daaf`）真机演练：`felis converge`**
+- 现场先剥后补：`kubectl patch` 清空 lobby `spec.rcon` 与 login `spec.startup.healthHTTPPort`（复刻老装机 CR 缺后加字段的状态）→ `/root/felis-fix77.bin converge` → 两条全部填回（`rcon{enabled:true,secretRef:lobby-rcon/password}`、`healthHTTPPort:8080`）；**第二次运行 → 双双 `already converged`**（幂等）；operator 侧滚动收尾，login/lobby 回 Running/Ready（日志 `converge-{1,2}.log`、`crs-before.yaml`）。
+- 语义边界（单测）：非零值一律不碰（运维自换的 secretRef 生还）、缺席 CR 只提示（创建仍归 setup）、占名非系统角色 CR 拒绝收敛、未配镜像跳过。
+
+**五、#79（`b9ebc87`，文档）**：troubleshooting 的 `[INERT]` 图例仍写「§12 列着唯一仍适用的字段」，而唯一候选 `spec.storage.retainOnDelete` 早已移除、§12 自述「每个字段都有 controller 读」。改为「今天没有字段处于该状态」并指向 §13 的移除记录。
+
+**六、发布链 rc smoke（tag `v0.1.0-rc1` → release run `35949621233`）**：
+**六、发布链 rc smoke（tag `v0.1.0-rc1` → release run `35949621233`）：首次全绿**
+- run 步骤实况：go vet/test ✅ → `plugins/test.sh` ✅ → `plugins/test-mods.sh` ✅（**release.yml 史上第一次真正跑这三关**）→ buildx 双架构构建 ✅ → **stamp 校验** ✅（`felis v0.1.0-rc1` + arm64 ELF 断言）→ publish ✅；
+- 产物释出：`felis-linux-amd64`（61,378,722 B）与 `felis-linux-arm64`（57,344,162 B）双资产，release 标记 `prerelease=true`；
+- **prerelease 语义复核**：`GET /repos/FelisMC/Felis/releases/latest` → **404**（RC 没有顶掉 latest——正是 release.yml 注释里防的那件事）；bootstrap 默认通道在无 stable 时按设计给出显式指引后 die、`felis update` 的 404 报错可读——两个行为本批均实测；
+- 产物级复验（目标架构实机）：arm64 资产 scp 到 VM 执行 → `felis v0.1.0-rc1`（`go1.26.8 linux/arm64`），且直接可用：`/root/felis-rc1.bin converge` 打现集群 → 双服 `already converged`；
+- 留档：`/root/felis-rc1.bin`；asset 副本 Mac `/tmp/rc1b/felis-linux-arm64`。
+- 剩余（产物决定，未代拍板）：仓库仍无 **stable** release → 新装走默认 release 通道会以指引性报错 die（提示改 dev 或等 stable）；切首个稳定 tag（如 `v0.1.0`）即让默认通道与 `felis update` 真正上线，建议维护者择时执行。
+
+**七、运维注记（非缺陷）**
+- pg_hba 的 `host felis_pgint felis 127.0.0.1/32 scram-sha-256` 行**再次丢失**（批 31 补过一次）——补回并 reload 后 pgint 才能连。重装/动过 PG 后先查这条（已写进速查）。
+- 升级域注记：#75/#76 的 pgint 新断言（`CountPendingSubmissionsBy`、`DeletePendingSubmission`/`DeleteSubmission` CAS）首次上真 PG：**17/17 全绿**。
+- 根盘：镜像构建后 83% → `docker builder prune -af` 回收 3.5G，docker 停回 inactive。
+- CI：`43699b4` success；`c57daaf` 被后一 commit 的并发策略取消（同分支 cancel-in-progress），其树被 `b9ebc87` 的 success 完整覆盖；`b9ebc87` success。
+- tracker 收编：**关** #10/#20/#21/#22（#20→`d9246dd`、#21→`f5a76cf`、#22→`3f2b28d`、#10→`23792d6`，均附证据评论）+ **关** #1/#8（本轮实现并真机验证）；**注记**（保持打开作老装机待办）#2/#3/#4；**留存** #9/#12/#13/#15（真增强，超出生产可用主干，未动）。
+
+## 结论：离"生产可用"还差什么（按优先级）
+
+1. ~~构建链路的上下文通道~~ ✅ **已修**（`f79e5eb`/`02fd2de`，真机全链路含拉回校验；Trivy DB 需按 §8e 镜像一次）。
+2. ~~镜像耐久~~ ✅ **已完成**（第二十九批：`a9b275a`/`13d64e0`/`fa0e8d7` + `c7e585e`；三次真机重跑 + GC 两演练——控制面与游戏镜像被 GC 后自动回拉；同批修复并复验 #46/#47/#48/#49）。
+3. ~~PG 级契约测试~~ ✅ **已落地**（`2a55a0d`）：`internal/pgint`（`-tags pgint`，需 `FELIS_TEST_PG_URL` 指向名字含 `pgint` 的库，harness 会 drop schema + 重放真实迁移）已覆盖会话/OTP/op-login/绑定码/submission/build/owner 角色；**首跑即抓到 #20**（索引与 ErrEmailTaken 从未存在）。运行方式见 CONTRIBUTING.md。
+4. ~~面板把 /jobs 显示出来~~ ✅ **已完成**（`97a64c8`，备份页「最近操作」卡，正是它把 #35 暴露出来的）。
+5. ~~多节点回收~~ ✅ **已修**（`daf7602`：`--reaper-node` → CronJob pod `kubernetes.io/hostname` nodeSelector，真机 render/dry-run/收敛 diff 三连；单节点部署不传即维持原状）。附带核销缺陷 #38（渲染提示里过期的 uid-1000/setfacl 指导）。
+6. ~~告警~~ ✅ **已完成**（`94f71ee` + `43df08b` + 第二十七批实弹演练：真实构建失败 → pending → 08:33:14Z firing；规则随 `deploy/alerts/` 交付）。
+7. ~~玩家可见的构建结果~~ ✅ **已修**（`72c4aa3`，缺陷 #36：列表路由附 `build_status`/`build_error`，面板「我的提交」展开行呈现，真机双例验证）。
+8. ~~面板文件编辑器入口~~ ✅ **已补**（`0a36b3f`，缺口补齐，真机 CDP 全链）。
+9. ~~fleet 系统服务死操作~~ ✅ **已修**（`2f90851`，缺陷 #37）。
+10. ~~S3 上传通道演练 + 存储向导逐屏~~ ✅ **已演练**（第二十四批：上传通道全链、0 缺陷；第三十二批：向导屏逐屏 + 上传取件 + UI 回滚，0 功能缺陷并修 #52）。
+11. ~~breakGlass 控制台逐屏~~ ✅ **已演练**（第二十五批：菜单 4 操作 + bootstrap 分支，缺陷 #39–#43 全修全验）。~~`felis setup` 全屏向导逐屏~~ ✅ **重跑侧已逐屏**（第二十六批），**首装侧连续走查也已完成**（第三十一批：scratch 库单次连续走完全部屏幕 + 轨道回顾，0 缺陷）。
+12. ~~`felis nano` 全链~~ ✅ **已走查**（第三十四批：HTTP 矩阵红 13→绿 60/60；systemd/配置/firewalld 三层重跑全绿；同批修复 #55——坏源静默挡住验证梯子）。
+13. ~~NetworkPolicy 真机强制矩阵~~ ✅ **已演练**（第三十五批：pod 源端口级矩阵全绿、转发型外部源「拒→放→拒」白名单闭环、host/ClusterIP/NodePort 路径全通；两条已归因注记——node-local 流量豁免、firewalld 仅放 DNAT 服务流）。
+14. ~~面板错误文案全量本地化~~ ✅ **已修**（第三十六批：#60，37 个用户可达码补齐双语映射，三类真机验证；管理面写操作 Run4c 复核 0 缺陷）。
+15. ~~管理面交互级收尾（submissions / 会话撤销 / 启动）~~ ✅ **已演练**（第三十七批：点击→构建→succeeded 全链、单条/全部撤销精确验证；同批修 #61——LP 写承诺按 LP 官方语义收窄，真机复验）。
+16. ~~缺陷 #56–#59（玩家管理回包 / 控制台回显 / 文件页 fail-fast / LP 诚实化）~~ ✅ **已修并归档**（第三十八批回填真机证据：修复前触发 + 修复后复验，四项全绿）。
+17. ~~服务器详情与运维面收尾~~ ✅ **已演练**（第三十八批：ServerFiles 写流程、备份/恢复全链、白名单/封禁闭环、ServerCard 停止、`/admin/updates` 复跑、metrics、CLI 核销——0 缺陷）。
+18. ~~reaper 多节点实机~~ ✅ **已演练**（第三十九批：VM 克隆双节点 k3s——pin 命中持盘节点并跑通、错位 pin fail-closed Pending、装置全回收；`felis-backups` PVC 的 volume node affinity 构成第二层保险）。
+19. ~~Java 插件层自动化缺口~~ ✅ **已补**（第四十批：`plugins/test.sh` + CI `plugins` 作业——三个手工测试与三个装机 jar 全进门禁，CI 4/4 绿；首跑抓出两处"没跑过"的错误：InviteCardTest 编译类路径缺 examination-api、limbo `+` 版本不可解析。同批关闭 demo-up 单起点化 #63、并复核剔除 #62）。
+20. ~~装载器 mod 层（fabric/forge/neoforge）~~ ✅ **已收口**（第四十一批：#64 三个 `gradlew` 补可执行位、#65 元数据对齐 AGPL-3.0-only；`plugins/test-mods.sh` + CI `mods` 作业 + release 双门禁；三真实专用服起服 E2E——mod 装载、`/link` 注册、控制台拒绝全绿。CI run `35892544563` = 5/5，`mods` 作业首跑即绿）。
+21. ~~全仓覆盖面对账 + 文档收尾~~ ✅ **已完成**（第四十二批：22 个 internal 包 × 面板 17 条页面路由 × 全部 tracked 脚本 × 仓库卫生逐项对账，无盲区；#66 README_EN 修复；§28 两张序列图对齐现行实现；本地 `scripts/sync.sh` 排除清单与 `go:embed` 冲突——修复+复现验证〔脚本私有，不入库〕）。
+22. ~~构建层资源回收~~ ✅ **已修**（第四十三批：#67 build Job TTL——完成 Job/Pod 不再无限堆积；真机：新构建 Job `ttl=604800` + 旧 Job 打 30s TTL 实测被 TTL 控制器收走。同批：6 轮生命周期循环 + operator/api/pg 三处注入全绿、PG 503 语义复验、30 路并发全 200）。
+23. ~~构建上下文拉取的短暂断连~~ ✅ **已修**（第四十三批追加：#68——fetch-context 有界重试〔45s 窗口 / 3s 间隔；4xx 不重试〕；真机：api 停机期连续 8 次拒连全部重试、恢复后构建 32s Complete；正常路径无重试开销。附：演练升级暴露的 `FELIS_IMAGE` 两代错位已对齐 auditfix63）。
+24. ~~modpack 规模上下文全链 + 并发构建~~ ✅ **已演练**（第四十四批：200MiB 上下文上传 0.44s → 批准 → 构建全链 13s、三路并发构建全绿；TTL 与玩家可见面在规模下复验，零缺陷。见该批节）。
+25. ~~内建底座构建（`FROM registry.felis.svc:5000/…`）~~ ✅ **已修并真机闭环**（第四十五批三连修：#70 拉取缺 `--insecure-pull`、#71 drop-ALL 卡解包、#72 trivy Java DB 未镜像——"真实模组包"形态的三个必踩点；修复后 paper 底座构建 27s Complete、`paper.jar` 扫描 0 漏洞）。
+26. ~~提交 → 装服 → 起服 闭环~~ ✅ **已演练**（第四十五批 capstone：构建产物 PATCH 为服务器镜像〔白名单门通过〕→ wake → Running 1/1 → 用户 marker 可读 + `Done (4.678s)!`；test-one 已复原为 paper:demo/Stopped）。
+27. ~~文档与测试工具面~~ ✅ **已修**（第四十五批：#69 README 部署承诺对齐实现；#73 bootstrap 测试密闭化〔VM 140 PASS〕；#74 console 断连测试消抖；CI 全绿）。
+28. ~~提交上传面的上限与生命周期~~ ✅ **已修并真机闭环**（第四十六批：#75 per-user create/upload 冷却〔429〕、pending ≤5〔403〕、2GiB 存储预算〔403〕；#76 撤回〔owner+pending CAS〕与管理员删除〔行+blob 双清〕；面板两步确认 CDP 全绿）。
+29. ~~未完成引导的导航（tracker #8）~~ ✅ **已修并真机闭环**（第四十六批 #77：`403 setup_required` → 自动 `/setup`；CDP 复验）。
+30. ~~已装机系统服的新增 CR 字段（tracker #1）~~ ✅ **已修并真机闭环**（第四十六批 #78：`sudo felis converge`——零值才填、非零不覆写；剥字段→填回→幂等三连真机过）。
+31. ~~troubleshooting `[INERT]` 图例~~ ✅ **已修**（第四十六批 #79，文档级）。
+
+## 剩余待演练队列（截至第四十三批）
+
+- ~~reaper 多节点实机~~ ✅ 已完结（第三十九批：临时克隆第二节点组双节点 k3s 实机——pin 命中持盘节点并跑通、错位 pin fail-closed Pending；装置已回收）。
+- ~~面板交互级收尾~~ ✅ 已完结（第三十六–三十八批：管理面写操作、submissions/会话、服务器详情页全链，0 缺陷）。
+- ~~`demo-up.sh`~~ ✅ 已完结（第四十批：单起点化 #63——bootstrap 装机 + 三项硬校验 + `felis setup` 交棒；T1/T2/T3 真机全绿）。
+- ~~Java 插件层自动化~~ ✅ 已补（第四十批：`plugins/test.sh` + CI `plugins` 作业；三个手工测试与三个装机 jar 首进 CI）。
+- ~~装载器 mod 层（fabric/forge/neoforge）~~ ✅ 已收口（第四十一批：#64/#65 修复 + `plugins/test-mods.sh`/CI `mods`/release 双门禁 + 三真实专用服起服 E2E）。
+- ~~全仓覆盖面对账~~ ✅ 已复核（第四十二批：模块×台账逐项对账无盲区；批次产出为文档/本地工具层面的修复，见该批节）。
+- ~~长稳/故障注入（首轮）~~ ✅ 已完成（第四十三批：6 轮起停循环 + 3 处注入 + PG 断连 503 + 30 路并发 + 25 分钟持续轮询；见该批节）。
+- ~~modpack 规模上下文 + 并发构建~~ ✅ 已完成（第四十四批：200MiB 上下文全链 13s + 并发 ×3 全绿；见该批节）。
+- ~~提交 → 装服 全链（含平台底座 FROM）~~ ✅ 已完成（第四十五批：三连修 #70/#71/#72 + capstone 提交→装服→起服；见该批节）。
+
+**队列现已清空。**（真实游戏客户端进服已按决定跳过，用内部 mint/approve hook 链替代；不依赖客户端的最大真机面见第四十批节 + 第四十一批节〔装载器三服起服〕；第四十二批为覆盖面对账复核。）
 
 ## 系统性观察
 
@@ -54,8 +954,27 @@ IPv6-only 接入（`ssh -6 -i ~/.ssh/id_ed25519 root@fdb2:2c26:f4e4:0:21c:42ff:f
 
 ## 复现入口速查
 
-- 面板会话 cookie：`/tmp/felis-cookies.json`；API 助手：`/tmp/fcurl.sh`
+- 面板会话 cookie：`/tmp/owner-cookies2.txt`（2026-09-23 铸；auditfix42 部署后仍有效；staff 邮件门被设计拒绝 → 走第二十八批的 op-login + 内部面 approve hook 链）；API 助手：`/tmp/fcurl.sh`
 - 测试服：`test-one`（minecraft ns，stopped）；合法备份 `bk-47ee2e7e96e5a4ca9d0e51b805518bac`
 - CDP 调试口：Mac `127.0.0.1:9333`（独立 Chrome，profile `/tmp/felis-chrome2`）；WebAuthn 虚拟认证器需在**同一 CDP 会话**内完成仪式，且先 `Page.bringToFront`（否则 NotAllowedError: page does not have focus）
-- 分支已部署到 VM：`felis-api` 镜像 = `felis:auditfix7`（含全部修复）
+- 已部署到 VM：控制面（api/operator/reaper + `FELIS_IMAGE`）= `registry.felis.svc:5000/felis/felis:auditfix61`（含至 #61 的全部修复；镜像托管于内建 registry）；系统服/推荐 ref = `registry.felis.svc:5000/felis/{limbo,lobby,paper}:demo`；源码快照 `/opt/felis/src`（`b4ef42d`；上一版在 `/opt/felis/src.bak`，更早 `src.old43`）；host 二进制 = `/usr/local/bin/felis` = `v0.0.0+fix61`（含 #53–#61；旧件留档 `/root/felis-fix54.bin`（sha `a0b29153…`）、`/root/felis-fix52.bin`、`/root/felis-auditfix44.bin`）；**docker 守护进程在第三十七批构建后已停（用前 `sudo systemctl start docker`）**；**world 执行器以 root+DAC_OVERRIDE 运行**；reaper CronJob（minecraft ns）`suspend=false / …auditfix61 / nodeSelector=localhost.localdomain`；迁移 `schema_migrations` max=21；SMTP 密码 env 待「configure email」刷新时落地
-- VM 内部面：`k3s kubectl -n felis port-forward svc/felis-api-internal 18081:8081`（Pod 重建后转发会悬死，需重启）；reaper CronJob（minecraft ns）已应用但 `suspend=true`
+- 安装器重跑配方（第三十批复用；先同步 `/opt/felis/src` 再跑）：`cd /root && FELIS_SKIP_FETCH=1 FELIS_INSTALL_MODE=full FELIS_IMAGE=registry.felis.svc:5000/felis/felis:<tag> FELIS_WORLDS_HOST_PATH=/var/lib/rancher/k3s/storage nohup bash /opt/felis/src/deploy/bootstrap.sh > /root/bootstrap-<tag>.log 2>&1 &`；完成后 `grep -c '\[fail\]'` 应为 0
+- 第三十一批 drill 留档：`/root/preTUI43/`（首装走查前快照：两 toml + 两 ns Secret + sha256、走查后 Secret 快照）、`/root/pre51-replica.toml`/`post51-replica.toml`/`post51b-replica.toml`（#51 红/绿副本证据）；走查残留（scratch 库、scratch 配置、hba 行、tmux 会话）均已清理
+- 第三十二批 drill 留档：`/root/preS3/`（S3 走查前快照：两 toml + 两 ns Secret + sha256 + deploys.txt）、`/root/s3wiz/`（posts3 快照、postroll-sha256、fix52-red.txt、fix52-green.txt）、`/root/felis-fix52`（含 #52 的宿主二进制）；MinIO 容器+卷+两镜像、`felis-uploads-s3` Secret、DB 行（`image_submissions` `sub-853a4e2ba4ba6443`）、/tmp 残留均已清理，docker 已停
+- 第三十三批 drill 留档（VM）：`/tmp/felis-fix53`、`/tmp/felis-fix54`（宿主二进制；sha `9afd141d…` / `a0b29153…`）、`/root/felis-fix52.bin`；`/tmp/player-cookies.txt`（player 会话）；面板 CDP 驱动脚本 `/tmp/cdp-updates2.js`（Mac 侧）
+- 第三十四批 drill 留档（VM）：`/srv/nanotest/`（nano 装置：`nano-stub.py` + 5 驱动脚本 + `matrix/` 红原跑 / `matrix-fix55/` 绿原跑）与四层重跑落盘 `matrix-red.r2.log`（`pass=47 fail=13`）/ `matrix-fix55.r2.log`（`pass=60 fail=0`）/ `nano-service.r2.log` / `nano-config.r2.log` / `nano-fw.r2.log`；`nano-stub` 临时单元仍在跑（收尾 `systemctl stop nano-stub`）；宿主二进制 `/usr/local/bin/felis-nano-test` = fix55 改名件（sha `2c8c34fd…`，与 `/usr/local/bin/felis` 同物）
+- 第三十五批 drill 留档（VM）：`/srv/npdrill/`（netpol 装置全套：`np-matrix.sh` + `np-matrix-run1/2.log`、`np-netns.sh`、`netns-probe.py`、`np-cidr-test.sh/.log`、`spec-before/after.json`、`td3.log`（包级归因）、`tcpdump-8081.log`（刷新存活）、`iptables-save.txt` 与 `nft-rules.txt`）；netns/veth、firewalld 临时挂载、策略改动全部已清理/复原（spec md5 `e6b07440…` 前后一致）
+- pgint 正确跑法（走 VM 的 PG，勿在本机起容器）：Mac 侧短隧道 `ssh -6 -i ~/.ssh/id_ed25519 -N -L 15433:127.0.0.1:5432 root@…` → `FELIS_TEST_PG_URL='postgres://felis:<pw>@localhost:15433/felis_pgint?sslmode=disable' go test -tags pgint ./internal/pgint/ -count=1`（pg_hba 需 `host felis_pgint felis 127.0.0.1/32 scram-sha-256`；第三十一批把现场丢失的这条补回；隧道用完即 kill）
+- breakGlass TUI 驱动法：VM tmux `new-session -d -s bg -x 160 -y 45 "/root/felis-auditfixNN breakGlass"` + `set-window-option -t bg remain-on-exit on`（否则退出摘要读不到）；send-keys/capture-pane 驱动；退出码 1 = 操作失败（卡上会显示原因）
+- 构建链路 drill 现成条件：`felis-build` 里有 `felis-service-token`（Secret 复制）；`felis-config` 的 kaniko/trivy pin 指向内建 registry 的 mirror 路径（`registry.felis.svc:5000/mirror/...`，见 §8e）——本批后安装器重跑会 **carry** 这些 pin（不再被写盘冲掉；#49 复验点）；直构一行：`POST /api/v1/images/build`（context=`http://felis-api-internal.felis.svc.cluster.local:8081/api/v1/internal/submissions/sub-bf7dc18e9dd97dc2/context`）
+- registry 工具：`curl -s http://127.0.0.1:5000/v2/_catalog`、`/v2/<repo>/tags/list`；**取 manifest 必须带 `Accept:`（OCI index/manifest list；缺 Accept 的 GET 会 404，别误判没推上）**；GC 演练配方：`k3s ctr images rm <ref>`（要清到 blob 级再加 `k3s ctr content prune references`）→ 删 pod / `rollout restart` → `kubectl describe` 看 `Pulled … in …ms`
+- setup 重跑向导（第二十六批）：状态屏 `c/s/e` 三流已走查；**reconfigure 非只读**——storage 选 Local / 提交 S3 即 apply + 滚 API（幂等），email 表单 esc 无副作用；从 S3 表单 esc 退回会重置为 Local 预选；存储屏（第三十二批补充）：chooser 预选当前值（↑↓ 切换），Local 直接 enter 提交、S3 五字段 `enter next`/末字段 `enter submit`，提交后 10–40s（含 API rollout status 180s 上限）
+- VM 内部面：`k3s kubectl -n felis port-forward svc/felis-api-internal 18081:8081`（Pod 重建后转发会悬死，需重启）；reaper CronJob（minecraft ns）已应用且已收敛（`suspend=false`）
+- 第三十八批收尾复查：docker `inactive`；控制面三处 = `auditfix61`；`/opt/felis/src` = `b4ef42d` tarball 展开（无 `.git`）；reaper `suspend=false` + `nodeSelector=localhost.localdomain`；test-one `Stopped`；`resolvecheck` 文件接口 409 `no_world_volume`（0.03s）
+- 第三十九批 drill（已回收）：临时 VM `felis-node2`（`prlctl clone --linked` 克隆 node1；node2 停用 felis-velocity/k3s-server/docker 后以 agent 加入）——流程与两向对照见该批节；node1 为克隆经历过一次正常重启（组件全回归、docker 停回 `inactive`）；收尾 job/节点对象/VM 全部删除，集群回到单节点
+- 第四十批工具与留档：VM `/tmp/mcprobe.py`（纯 socket MC 探针：`python3 /tmp/mcprobe.py status <sub>.<root>`；`login <sub>.<root> <name>` 打登录首包）；插件门禁跑法 = `docker run --rm -v /opt/felis/src:/src:z -w /src gradle:8.14-jdk21 bash plugins/test.sh`（或宿主 `bash plugins/test.sh`，需 JDK 21 + Gradle 8.14；日志 `/root/plugins-test-b40.log`）；T3（demo-up 默认臂全量重跑）日志 `/root/demo-up-b40.log`；`/opt/felis/src` = `c59b387` 快照（上一版 `src.bak40`；此后全量重跑用 `sudo bash /opt/felis/src/deploy/demo-up.sh` 即可，SKIP 开关见脚本头）；CI run `35888009965`（含新 `plugins` 作业首跑绿）
+- 第四十三批工具与留档：控制面三处 = `registry.felis.svc:5000/felis/felis:auditfix62`（含 #67；镜像源 `10.43.182.43:5000/felis/felis:auditfix62`，宿主 docker build 后 `docker push 10.43.182.43:5000/...`）；`/opt/felis/src` = `2755e41` 快照（上一版 `src.bak42`）；owner 会话重铸 = `bash /root/mint-owner-43.sh`（cookie 落 `/tmp/owner-jar43.txt`）；直构触发（含 TTL 检查）= `curl -sk -b /tmp/owner-jar43.txt -X POST https://127.0.0.1:30443/api/v1/images/build -H 'Content-Type: application/json' -d '{"image_ref":"registry.felis.svc:5000/e2e/ttl-probe:latest","dockerfile":"FROM scratch\nLABEL felis=x\n","context_ref":"http://felis-api-internal.felis.svc.cluster.local:8081/api/v1/internal/submissions/sub-bf7dc18e9dd97dc2/context"}'` → `kubectl -n felis-build get job build-bld-<id> -o jsonpath='{.spec.ttlSecondsAfterFinished}'` 应为 `604800`；压测编排 `/root/soak43.sh`（日志 `soak43.log`，含 6 轮循环 + 3 注入）、长轮询 `/root/poll43-long.sh`（日志 `poll43-long.log`）；镜像构建日志 `/root/felis-image-build43.log`。注意：`lobby`/`login` 是保留名，internal per-server 路由（status/wake 等）对它们**按设计**返回 400，勿作缺陷误报。
+- 第四十三批追加（#68）留档：控制面 api/operator/reaper 镜像 + api/operator `FELIS_IMAGE` = `registry.felis.svc:5000/felis/felis:auditfix63`（`felis version` = `v0.0.0+fix63`；升级时 `set image` 与 `FELIS_IMAGE` env 必须成对改，否则构建 Job 的 fetch 容器会悄悄用旧镜像）；`/opt/felis/src` = `b76d0ac` 快照（上一版 `src.bak43` = `2755e41`）；复刻 Job `/root/job68-new.json`（由红证据 Job `build-bld-1790187851749257160` 派生：换名 `build-retry68`、剥 controller uid 标签与 selector、镜像改 auditfix63、复用 `felis-service-token` Secret）；重演验证法：`kubectl -n felis scale deploy/felis-api --replicas=0` → `kubectl -n felis-build create -f /root/job68-new.json` → `kubectl -n felis-build logs <pod> -c context-fetch`（应见 `connection refused; retrying`）→ `scale --replicas=1` → Job 收敛 `Complete`；对照：api 在线直构成功且 fetch 日志无重试行。
+- 第四十四批工具与留档：200MiB 上下文构造 = 5×40MiB `head -c 41943040 /dev/urandom` + `Dockerfile`（`FROM scratch` + `COPY payload /payload`），`tar -czf /root/bigctx.tar.gz Dockerfile payload`；上传 = `curl -sk -b /tmp/owner-jar43.txt -X POST -H "Content-Type: application/gzip" --data-binary @/root/bigctx.tar.gz https://127.0.0.1:30443/api/v1/me/submissions/<id>/context`；采样 = `k3s crictl stats`（此版 **无 `--no-trunc`**，列解析取 `$2=NAME $4=MEM`）；docker 缓存清理 = `systemctl start docker && docker builder prune -af && systemctl stop docker`；留档 `/root/stats44.log`、`/root/stats44-build.log`、提交 `sub-cc160b3ad19080f9`、镜像 `e2e/conc-b44-{1,2,3}`。
+- 第四十五批工具与留档：从平台底座构建的配方 = 提交上下文含 `Dockerfile`（`FROM registry.felis.svc:5000/felis/paper:demo` + `COPY marker.txt /felis-probe-marker.txt`）；装服验证 = `PATCH /api/v1/servers/test-one {"image":"registry.felis.svc:5000/user-uploads/<sub>:latest"}` → `POST /servers/test-one/wake` → `kubectl -n minecraft exec test-one-0 -- cat /felis-probe-marker.txt`；trivy 双 DB 镜像 = `mirror/trivy-db:2` + `mirror/trivy-java-db:1`（Java DB digest `5766dfbb…`；两键在 `/etc/felis/felis.{host,pod}.toml` 与 felis-config Secret 双副本）；`bootstrap_test.sh` 需在 Linux 跑（VM 留档 `/root/btest72/`，应 140 PASS）；本批提交样本 `sub-{b45d416f4af811ef(无镜像),83f677e41acd80c3,171e8f967f0de3bc,c006cbd633317ccb,cc160b3ad19080f9}`；留档 `/root/probe44/`、`/root/probe44.tar.gz`、`/root/probe44*.sid`。
+- 第四十六批工具与留档：控制面三处 = `registry.felis.svc:5000/felis/felis:auditfix77`（宿主 docker 源 `10.43.182.43:5000/felis/felis:auditfix77`，`v0.0.0+fix77`；api/operator + minecraft ns `felis-reaper` CronJob + api/operator `FELIS_IMAGE` 四处成对改）；`/opt/felis/src` = `b9ebc87` 快照（上一版 `src.bak46`）；宿主 drill 二进制 `/root/felis-fix77.bin`（Mac 侧 `CGO_ENABLED=0 GOOS=linux GOARCH=arm64 go build -ldflags="-s -w -X main.version=v0.0.0+fix77" ./cmd/felis`；scp 到 IPv6 主机时主机位必须写 `root@[fdb2:…]` 方括号）；#75/#76/#77/#78 留档 `/root/probe77/`（`create-lane.log`、`lane2.log`、`lane3.log`〔含玩家会话矩阵 + 分级 ZT 对照 + curl 7.76 在 `-H Host:` 下**不发 jar cookie**的坑——host 路由 drill 用显式 `-H "Cookie: felis_session=…"`〕、`converge-1.log`/`converge-2.log`、`crs-before.yaml`、各步请求/响应 json）；面板 CDP 截图（Mac）：`/tmp/setup8-redirect.png`、`/tmp/withdraw-{1,2}-*.png`、`/tmp/admdelete-{1,2}-*.png`，脚本 `/tmp/cdp-setup8.js`、`/tmp/cdp-withdraw76.js`、`/tmp/cdp-admdelete76.js`；锁定会话铸法 = `sha256('drill-locked-77')` 直插 `sessions`（用户 `3f2c1b0a-…`）；玩家会话铸法 = `[email protected]` 邮件 OTP（码在 felis-api 日志 `no Mailer configured` 行；`edge-dis2` 是软删死账号，登录门按设计排除）；pgint 前置：`pg_hba` 需 `host felis_pgint felis 127.0.0.1/32 scram-sha-256`（本批第二次丢失并补回）；发布链 smoke = tag `v0.1.0-rc1`（release run `35949621233`），prerelease 不移动 `/releases/latest`——无 stable 时 bootstrap 默认通道给出显式指引后 die、`felis update` 404 报错可读；首个稳定 tag 是产物决定（未代拍板）。

CONTRIBUTING.md

@@ -67,6 +67,20 @@ go test ./internal/api
 go test ./cmd/felis
 ```
 
+The hermetic suites run against in-memory fakes; the business stores' SQL is
+verified separately against a real Postgres, on a throwaway database whose name
+must contain `pgint` (the harness drops and recreates its schema and replays the
+embedded migrations):
+
+```bash
+FELIS_TEST_PG_URL='postgres://felis:***@127.0.0.1:5432/felis_pgint?sslmode=disable' \
+  go test -tags pgint ./internal/pgint/ -v
+```
+
+Run it after touching anything under `internal/api/pgrepo.go`, `internal/submit`,
+or `internal/build` that speaks SQL: the fakes encode the contract, and this
+suite exists to catch the drift between the fakes and the real queries.
+
 Build the CLI:
 
 ```bash

README.md

@@ -17,10 +17,10 @@ A Kubernetes-driven Minecraft server hosting platform — one command to deploy,
 
 - **即开即玩**：玩家尝试连接时自动唤醒服务器，空闲后自动休眠，像游戏主机一样省资源。
 - **Web 控制面板**：浏览器中查看服务器状态、在线玩家与资源用量，管理备份与恢复。
-- **自动备份与恢复**：定时将世界打包存档，支持从任意备份点一键回滚。
+- **备份与恢复**：一键把整服数据（世界、配置、插件/模组，即整个 /data 卷）打包进集群内的归档库，支持从任意备份点回滚；默认安装就已启用（归档 PVC 与路径由安装器一并生成）。
-- **智慧回收**：超过 15 天无人游玩的世界自动备份后删除，释放磁盘空间。
+- **智慧回收（可选开启）**：超过 15 天无人游玩的世界自动备份后删除，释放磁盘空间；安装时设置 `FELIS_WORLDS_HOST_PATH`（k3s 默认 `/var/lib/rancher/k3s/storage`）即启用每日回收，不设置则不删任何世界。
 - **多核心支持**：兼容 Paper、Fabric、Forge、NeoForge，经由 Velocity 代理统一入口。
-- **模组自助提交**：玩家自行上传模组包，服主审批通过后自动构建并部署。
+- **模组自助提交**：玩家自行上传模组包，服主审批通过后自动构建；构建产物进入镜像白名单，可直接选用为服务器镜像完成部署。
 - **Passkey 登录**：支持指纹、面容、硬件密钥等无密码认证方式。
 - **零信任安全**：面板流量由 Cloudflare Access 保护，集群内 API 不暴露到公网。
 
@@ -29,7 +29,7 @@ A Kubernetes-driven Minecraft server hosting platform — one command to deploy,
 在准备好的 Linux 主机上执行：
 
 ```bash
-curl -fsSL https://raw.githubusercontent.com/MliroLirrorsIngenuity/Felis/main/deploy/bootstrap.sh | sudo bash
+curl -fsSL https://raw.githubusercontent.com/FelisMC/Felis/main/deploy/bootstrap.sh | sudo bash
 ```
 
 脚本将自动安装 K3s、部署控制平面并启动设置向导。完成后浏览器访问已配置的域名进入控制面板即可使用。
@@ -41,7 +41,7 @@ curl -fsSL https://raw.githubusercontent.com/MliroLirrorsIngenuity/Felis/main/de
 > export FELIS_GITHUB_TOKEN=<对本仓库有读权限的 token>
 > printf 'header = "Authorization: Bearer %s"\n' "$FELIS_GITHUB_TOKEN" \
 >   | curl -fsSL --config - -H "Accept: application/vnd.github.raw" \
->       https://api.github.com/repos/MliroLirrorsIngenuity/Felis/contents/deploy/bootstrap.sh \
+>       https://api.github.com/repos/FelisMC/Felis/contents/deploy/bootstrap.sh \
 >   | sudo -E bash
 > ```
 >

README_EN.md

@@ -17,10 +17,10 @@ Table of Contents
 
 - **Wake on Join**: Servers start automatically when a player connects, and stop when idle — like hibernate for your server.
 - **Web Dashboard**: Monitor server status, online players, and resource usage from your browser, with backup and restore management.
-- **Auto Backup & Restore**: Scheduled world backups with one-click rollback from any backup point.
+- **Backup & Restore**: One-click snapshots of a server's whole data volume (worlds, config, plugins/mods — the entire /data volume) into the cluster's archive store, with rollback from any backup point — enabled by default (the installer renders the archive PVC and its path).
-- **World Reaper**: Worlds idle for more than 15 days are automatically backed up and removed to free disk space.
+- **World Reaper** (opt in): Worlds idle for more than 15 days are automatically backed up and removed to free disk space. Enable it by setting `FELIS_WORLDS_HOST_PATH` at install time (on k3s: `/var/lib/rancher/k3s/storage`); without it, no world is ever deleted.
 - **Multi-core Support**: Compatible with Paper, Fabric, Forge, and NeoForge, federated behind a Velocity proxy.
-- **Modpack Submission**: Players submit custom modpacks; admin approval triggers automatic build and deployment.
+- **Modpack Submission**: Players submit custom modpacks; admin approval triggers an automatic build, and the result is whitelisted as a server image you can select to deploy.
 - **Passkey Login**: Passwordless authentication via fingerprint, face recognition, or hardware security keys.
 - **Zero Trust Security**: Panel traffic protected by Cloudflare Access; the internal API is never exposed to the internet.
 
@@ -29,11 +29,32 @@ Table of Contents
 On a prepared Linux host, run:
 
 ```bash
-curl -fsSL https://raw.githubusercontent.com/MliroLirrorsIngenuity/Felis/main/deploy/bootstrap.sh | sudo bash
+curl -fsSL https://raw.githubusercontent.com/FelisMC/Felis/main/deploy/bootstrap.sh | sudo bash
 ```
 
 The script installs K3s, deploys the control plane, and launches a setup wizard. Once done, open your browser at the configured domain.
 
+> **This repository is currently private**, so the command above returns 404. Use the
+> credentialed form instead; the installer itself needs the same token to resolve and
+> download the release, so pass it through with `sudo -E`:
+>
+> ```bash
+> export FELIS_GITHUB_TOKEN=<a token with read access to this repository>
+> printf 'header = "Authorization: Bearer %s"\n' "$FELIS_GITHUB_TOKEN" \
+>   | curl -fsSL --config - -H "Accept: application/vnd.github.raw" \
+>       https://api.github.com/repos/FelisMC/Felis/contents/deploy/bootstrap.sh \
+>   | sudo -E bash
+> ```
+>
+> The token reaches `curl --config -` over stdin instead of the command line: argv is
+> readable by any local user via `/proc`, and that is exactly why the installer's
+> internal `github_api` uses the same form.
+
+Rerunning this command is also how you upgrade felis-api to a newer version (`felis setup`
+cannot — it uses the binary already installed on the host). The rerun keeps the installed
+root domain but **not** the channel: if this host follows main, also
+`export FELIS_VERSION_BOOTSTRAP=dev`.
+
 ## Build from Source
 
 Felis is built with Go and Node.js:

cmd/felis/api.go

@@ -18,6 +18,7 @@ import (
 	"felis.lolicon.best/internal/config"
 	"felis.lolicon.best/internal/fileedit"
 	"felis.lolicon.best/internal/mail"
+	"felis.lolicon.best/internal/naming"
 	"felis.lolicon.best/internal/panel"
 	"felis.lolicon.best/internal/passkey"
 	"felis.lolicon.best/internal/platform"
@@ -142,10 +143,14 @@ func cmdAPI(args []string, stdout, stderr io.Writer) int {
 	// build namespace and pushes to the internal registry. The build Pod never
 	// holds DB credentials — felis-api owns the PG store and admits scanned
 	// images, so the Builder is constructed here with both bindings.
+	buildCfg := buildConfig(cfg)
+	// The fetch initContainer runs THIS image's fetch-context entrypoint, so the
+	// build config carries the api's own image (the platform sets FELIS_IMAGE).
+	buildCfg.FelisImage = os.Getenv("FELIS_IMAGE")
 	builder := &build.Builder{
 		Store:  build.NewPGStore(drv.DB()),
-		Jobs:   build.NewK8sJobs(cl, buildConfig(cfg)),
+		Jobs:   build.NewK8sJobs(cl, buildCfg),
-		Config: buildConfig(cfg),
+		Config: buildCfg,
 	}
 
 	// User-modpack approval lane (user-directed extension over §16; see
@@ -159,14 +164,19 @@ func cmdAPI(args []string, stdout, stderr io.Writer) int {
 	// The blob upload transport is selected by the shape of user_uploads_context —
 	// the two backends the setup wizard chooses between. A local path wires
 	// LocalContextStore (the mounted uploads PVC); an s3:// base wires
-	// S3ContextStore when its credentials resolve. Either way the store's target is
+	// S3ContextStore when its credentials resolve. Anything else — or an s3:// base
-	// derived from the SAME config field the context ref uses, so the blob lands
+	// with no credentials configured — leaves Blobs nil so POST
-	// exactly where Kaniko's --context points. Anything else — or an s3:// base with
-	// no credentials configured — leaves Blobs nil so POST
 	// /me/submissions/{id}/context returns 503, honest like the restore executor
-	// when its PVC is not supplied. (Letting the sandboxed Kaniko build Pod READ the
+	// when its PVC is not supplied.
-	// context — PVC mount for local, creds+egress for S3 — is a separate deployment
+	//
-	// integration.)
+	// Reading the blob back is the API's job, not Kaniko's: the build Pod runs in
+	// another namespace and can neither mount the uploads PVC (a PVC does not cross
+	// namespaces) nor hold object-store credentials, so ContextBaseURL makes the
+	// derived context ref an internal-face URL that the build Job's fetch
+	// initContainer streams (cmd/felis fetch-context). The platform renders this
+	// address into the api Deployment (felis API base URL env); the fallback keeps
+	// a hand-rolled deployment working under the platform's default control
+	// namespace.
 	contextBase := cfg.Registry.UserUploadsContext
 	var blobs submit.Blobs
 	switch {
@@ -186,11 +196,12 @@ func cmdAPI(args []string, stdout, stderr io.Writer) int {
 		fmt.Fprintf(stderr, "felis api: user-uploads context %q is neither a local path nor an s3:// base — modpack upload transport disabled (POST /api/v1/me/submissions/{id}/context returns 503)\n", contextBase)
 	}
 	submissions := &submit.Manager{
-		Store:        submit.NewPGStore(drv.DB()),
+		Store:          submit.NewPGStore(drv.DB()),
-		Builds:       builder,
+		Builds:         builder,
-		Registry:     cfg.Registry.URL,
+		Registry:       cfg.Registry.URL,
-		ContextStore: contextBase,
+		ContextStore:   contextBase,
-		Blobs:        blobs,
+		ContextBaseURL: internalAPIBaseURL(),
+		Blobs:          blobs,
 	}
 
 	// Restore subsystem (spec §7): the weak-SA restore Job mounts the target
@@ -259,6 +270,7 @@ func cmdAPI(args []string, stdout, stderr io.Writer) int {
 		Builder:     builder,
 		Restorer:    restorer,
 		Backuper:    backuper,
+		JobStatus:   api.NewK8sJobStatus(cl, cfg.K8s.Namespace),
 		Files:       files,
 		Submissions: submissions,
 		Mailer:      mailer,
@@ -279,6 +291,11 @@ func cmdAPI(args []string, stdout, stderr io.Writer) int {
 		AdminHostname: cfg.Auth.AdminHostname,
 		PanelHostname: cfg.Auth.PanelHostname,
 		WakeCooldown:  30 * time.Second,
+		// The user-modpack lane's per-user throttles: a create spaces out
+		// review-queue rows, an upload spaces out (up to 1 GiB) context streams.
+		// Separate keys, so the normal create→upload sequence stays immediate.
+		SubmitCreateCooldown: 30 * time.Second,
+		SubmitUploadCooldown: 15 * time.Second,
 		// Bound concurrent console/build-log SSE streams per principal. Generous enough
 		// for legitimate multi-tab / multi-server watching, while capping how many
 		// upstream follow connections a single caller can tie up if their streams stall.
@@ -401,9 +418,32 @@ func buildConfig(cfg *config.Config) build.Config {
 	return build.Config{
 		Namespace:   cfg.Registry.BuildNamespace,
 		RegistryURL: cfg.Registry.URL,
+		// Empty overrides fall back to the build package's defaults, so an
+		// install that has not imported kaniko/trivy keeps the compiled-in refs
+		// (and fails loudly on pull rather than silently building with the wrong
+		// image).
+		KanikoImage: cfg.Registry.KanikoImage,
+		TrivyImage:  cfg.Registry.TrivyImage,
+		CPULimit:    cfg.Registry.BuildCPULimit,
+		MemLimit:    cfg.Registry.BuildMemLimit,
+		// Empty keeps Trivy's own default; an install with builds points this at
+		// the internal DB mirror (see config.RegistryConfig.TrivyDBRepository).
+		TrivyDBRepository:     cfg.Registry.TrivyDBRepository,
+		TrivyJavaDBRepository: cfg.Registry.TrivyJavaDBRepository,
 	}
 }
 
+// internalAPIBaseURL resolves the platform's internal-face base URL: the address
+// the platform rendered into this pod (felis API base URL env), or — for a
+// hand-rolled deployment that set none — the platform default control namespace,
+// the same fallback setup.go uses to hand the login gate its address.
+func internalAPIBaseURL() string {
+	if base := os.Getenv(naming.EnvAPIBaseURL); base != "" {
+		return base
+	}
+	return platform.InternalAPIBaseURL(platform.DefaultControlNamespace)
+}
+
 // uploadsSchemeRE matches a leading URL scheme like "s3://" or "gs://".
 var uploadsSchemeRE = regexp.MustCompile(`^[a-zA-Z][a-zA-Z0-9+.-]*://`)
 

cmd/felis/api_test.go

@@ -44,6 +44,32 @@ func TestAuthSourcesFromConfig(t *testing.T) {
 	}
 }
 
+// TestBuildConfig_ProjectsOverrides pins the [registry] overrides reaching the
+// build subsystem: unset fields must stay EMPTY (the build package's compiled-in
+// defaults apply there, not here), and set fields must pass through verbatim —
+// an air-gapped install points these at its imported mirrors.
+func TestBuildConfig_ProjectsOverrides(t *testing.T) {
+	empty := buildConfig(&config.Config{})
+	if empty.KanikoImage != "" || empty.TrivyImage != "" || empty.CPULimit != "" || empty.MemLimit != "" {
+		t.Errorf("empty registry config must project empty overrides (defaults live in internal/build), got %+v", empty)
+	}
+	full := buildConfig(&config.Config{Registry: config.RegistryConfig{
+		URL:            "registry.felis.svc:5000",
+		BuildNamespace: "felis-build",
+		KanikoImage:    "reg/kaniko:v1",
+		TrivyImage:     "reg/trivy:v1",
+		BuildCPULimit:  "1",
+		BuildMemLimit:  "2Gi",
+	}})
+	if full.KanikoImage != "reg/kaniko:v1" || full.TrivyImage != "reg/trivy:v1" ||
+		full.CPULimit != "1" || full.MemLimit != "2Gi" {
+		t.Errorf("registry overrides did not reach build.Config: %+v", full)
+	}
+	if full.Namespace != "felis-build" || full.RegistryURL != "registry.felis.svc:5000" {
+		t.Errorf("namespace/registry url must keep projecting: %+v", full)
+	}
+}
+
 // TestNewAPIServerSetsHardenedTimeouts pins the gosec-G112 hardening on every
 // felis-api listener: the shared factory must bound the header and idle phases
 // (Slowloris + idle-connection exhaustion) while leaving WriteTimeout UNSET, because

cmd/felis/backupnow.go

@@ -86,23 +86,31 @@ func requestBackup(ctx context.Context, hc *http.Client, baseURL, token, name, o
 
 // backupErrorFromResponse turns a non-202 into a human message. The well-known codes get
 // an operator-facing explanation; anything else falls back to the API's
-// {"error":{message}} body, then the bare status code.
+// {"error":{code,message}} body, then the bare status code.
 func backupErrorFromResponse(resp *http.Response) error {
-	switch resp.StatusCode {
-	case http.StatusConflict: // not_stopped
-		return fmt.Errorf("the server must be stopped before its world can be backed up — halt it first")
-	case http.StatusServiceUnavailable: // backup_unavailable
-		return fmt.Errorf("the backup subsystem is not configured on felis-api (FELIS_IMAGE / FELIS_BACKUP_PVC unset)")
-	case http.StatusNotFound:
-		return fmt.Errorf("no such server")
-	}
 	var e struct {
 		Error struct {
+			Code    string `json:"code"`
 			Message string `json:"message"`
 		} `json:"error"`
 	}
 	raw, _ := io.ReadAll(io.LimitReader(resp.Body, 1<<16))
 	_ = json.Unmarshal(raw, &e)
+
+	switch resp.StatusCode {
+	case http.StatusConflict:
+		// Two refusals share 409: the stopped gate and the missing-world-volume
+		// gate. The body's code distinguishes them; a code-less body reads as the
+		// stopped gate (the only 409 before the volume gate existed), and any other
+		// coded 409 falls through to the API's own operator text.
+		if e.Error.Code == "" || e.Error.Code == "not_stopped" {
+			return fmt.Errorf("the server must be stopped before its world can be backed up — halt it first")
+		}
+	case http.StatusServiceUnavailable: // backup_unavailable
+		return fmt.Errorf("the backup subsystem is not configured on felis-api (FELIS_IMAGE / FELIS_BACKUP_PVC unset)")
+	case http.StatusNotFound:
+		return fmt.Errorf("no such server")
+	}
 	if e.Error.Message != "" {
 		return fmt.Errorf("felis-api: %s", e.Error.Message)
 	}
@@ -120,3 +128,19 @@ func performBackupNow(ctx context.Context, cl client.Client, controlNamespace, n
 	hc := &http.Client{Timeout: 10 * time.Second}
 	return requestBackup(ctx, hc, baseURL, token, name, osUser)
 }
+
+// backupPickable narrows the backup picker to servers the backup API can accept.
+// System servers (login/lobby) are excluded: they have no row in the servers
+// table and carry reserved names, so every attempt dies in name validation —
+// offering them would be a dead pick. The halt picker keeps them on purpose
+// (break-glass retains full power over system servers); only the API-backed
+// backup op cannot reach them.
+func backupPickable(servers []haltableServer) []haltableServer {
+	out := make([]haltableServer, 0, len(servers))
+	for _, s := range servers {
+		if !s.system {
+			out = append(out, s)
+		}
+	}
+	return out
+}

cmd/felis/backupnow_test.go

@@ -114,16 +114,23 @@ func TestRequestBackup(t *testing.T) {
 	cases := []struct {
 		name   string
 		code   int
+		body   string // optional JSON error body
 		expect string
 	}{
-		{"409 not_stopped", http.StatusConflict, "must be stopped"},
+		{"409 not_stopped", http.StatusConflict, "", "must be stopped"},
-		{"503 backup_unavailable", http.StatusServiceUnavailable, "not configured"},
+		{"409 no_world_volume surfaces the API's own text", http.StatusConflict,
-		{"404 not found", http.StatusNotFound, "no such server"},
+			`{"error":{"code":"no_world_volume","message":"this server has no world volume yet — start it once to create it, then retry"}}`,
+			"no world volume yet"},
+		{"503 backup_unavailable", http.StatusServiceUnavailable, "", "not configured"},
+		{"404 not found", http.StatusNotFound, "", "no such server"},
 	}
 	for _, tc := range cases {
 		t.Run(tc.name, func(t *testing.T) {
 			srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
 				w.WriteHeader(tc.code)
+				if tc.body != "" {
+					_, _ = io.WriteString(w, tc.body)
+				}
 			}))
 			defer srv.Close()
 			_, err := requestBackup(context.Background(), hc, srv.URL, "tok", "survival", "alice")
@@ -142,3 +149,25 @@ func TestRequestBackup(t *testing.T) {
 		}
 	})
 }
+
+// The Sync picker must not offer system servers: the backup API validates names
+// and resolves a servers-table row, so a lobby/login pick can only die in
+// validation — a dead choice in an emergency console.
+func TestBackupPickable(t *testing.T) {
+	got := backupPickable([]haltableServer{
+		{name: "lobby", phase: "Running", system: true},
+		{name: "login", phase: "Running", system: true},
+		{name: "test-one", phase: "Stopped"},
+	})
+	if len(got) != 1 || got[0].name != "test-one" || got[0].system {
+		t.Fatalf("backupPickable = %+v, want only the user server", got)
+	}
+
+	// Survivors keep their input order (the picker's cursor math depends on it).
+	got = backupPickable([]haltableServer{
+		{name: "alpha"}, {name: "login", system: true}, {name: "beta"},
+	})
+	if len(got) != 2 || got[0].name != "alpha" || got[1].name != "beta" {
+		t.Fatalf("backupPickable order = %+v, want [alpha beta]", got)
+	}
+}

cmd/felis/breakglass.go

@@ -76,12 +76,19 @@ type ownerStore interface {
 	AdminExists(ctx context.Context) (bool, error)
 	// UserByUsername loads a staff login projection.
 	UserByUsername(ctx context.Context, username string) (*api.StaffUser, error)
+	// OwnerUsername names the single active Owner seat, or "" when none exists.
+	// provisionOwner refuses to re-target anything but this username: with the
+	// seat occupied, a fresh name would mint a second owner row (UpsertOwner's
+	// insert arm) while the existing — possibly compromised — seat stays live,
+	// and no supported path can delete an owner row afterwards.
+	OwnerUsername(ctx context.Context) (string, error)
 	UpsertOwner(ctx context.Context, id, username, email string) error
 	// InsertOperator mints a NEW Operator staff account. Unlike UpsertOwner it is
 	// insert-only: a username already taken is a conflict (api.ErrConflict), never a
 	// silent reset, so adding an Operator can never clobber the Owner or an existing
-	// Operator. The row is role=admin, identical in shape to the Owner — Felis has no
+	// Operator. The row is role=admin — an Operator is staff BELOW the single
-	// separate operator DB role (migration 0003: staff = role=admin).
+	// role=owner identity (migration 0011 adds that role); the two are the only
+	// staff roles.
 	InsertOperator(ctx context.Context, id, username, email string) error
 	// CompleteOwnerSetup atomically consumes the in-game link code, creates or
 	// promotes the bound Owner, enables local auth, and stores the one-time setup
@@ -266,20 +273,31 @@ func authenticateAdmin(ctx context.Context, s ownerStore, username string) (matc
 	if err != nil {
 		return "", false, err
 	}
-	if u.Role != "admin" {
+	// Staff means admin OR owner: recovery attribution must accept the Owner (the
+	// primary break-glass identity), not just plain admins.
+	if u.Role != "admin" && u.Role != "owner" {
 		return "", false, nil
 	}
 	return u.Username, true, nil
 }
 
 // provisionOwner mints or resets the single Owner account direct-to-Postgres,
-// passwordless. The account is role=admin with no password — the Owner completes
+// passwordless. The account is role=owner with no password — the Owner completes
 // passwordless login setup via the web setup-token flow after `felis setup`.
+// With a seat already occupied the reset must name that seat (ownerSeatTakenError
+// otherwise): the upsert's insert arm would silently mint a SECOND owner, and
+// every owner row is undeletable through the panel, so the tier could never
+// converge back to one.
 func provisionOwner(ctx context.Context, s ownerStore, username, email string) error {
 	username = strings.TrimSpace(username)
 	if username == "" {
 		return errors.New("owner username is required")
 	}
+	if seat, err := s.OwnerUsername(ctx); err != nil {
+		return fmt.Errorf("check the owner seat: %w", err)
+	} else if seat != "" && seat != username {
+		return &ownerSeatTakenError{seat: seat}
+	}
 	id := newOwnerID()
 	if id == "" {
 		return errors.New("generate owner id: entropy source failed")
@@ -290,13 +308,26 @@ func provisionOwner(ctx context.Context, s ownerStore, username, email string) e
 	return nil
 }
 
-// provisionOperator mints a NEW Operator staff account direct-to-Postgres. Like the
+// ownerSeatTakenError refuses an Owner reset that names anything but the
-// Owner it is role=admin and passwordless — Felis has no separate operator DB role,
+// occupied seat, naming it so the operator can retype. Is reports
-// so an Operator is simply an additional staff admin (migration 0003). UNLIKE
+// api.ErrConflict so the TUI's recoverable-error branch (shared with the
-// provisionOwner, which upserts the single Owner and resets it on a username
+// operator path's taken-name clash) routes back to the form instead of ending
-// conflict, this is insert-only: a username already taken returns api.ErrConflict
+// the console.
-// rather than overwriting a live account, so adding an Operator can never silently
+type ownerSeatTakenError struct{ seat string }
-// clobber the Owner's or another Operator's account.
+
+func (e *ownerSeatTakenError) Error() string {
+	return fmt.Sprintf("an Owner already exists as %q — enter that username to reset the Owner", e.seat)
+}
+
+func (e *ownerSeatTakenError) Is(target error) bool { return target == api.ErrConflict }
+
+// provisionOperator mints a NEW Operator staff account direct-to-Postgres. It is
+// role=admin and passwordless — an additional staff admin below the single
+// role=owner identity (migrations 0003 + 0011). UNLIKE provisionOwner, which
+// upserts the single Owner and resets it on a username conflict, this is
+// insert-only: a username already taken returns api.ErrConflict rather than
+// overwriting a live account, so adding an Operator can never silently clobber
+// the Owner's or another Operator's account.
 func provisionOperator(ctx context.Context, s ownerStore, username, email string) error {
 	username = strings.TrimSpace(username)
 	if username == "" {
@@ -387,7 +418,7 @@ func newSetupToken() (raw, hash string, err error) {
 
 // performSetupMCBind is the `felis setup` Owner-establishment path: the operator
 // binds their Minecraft account via a one-time link code the login gate handed
-// them in-game, the bound user is promoted to role='admin' (passwordless Owner),
+// them in-game, the bound user is promoted to role='owner' (passwordless Owner),
 // local auth is enabled, and a one-time setup URL is minted for the first web
 // login where the Owner verifies email / enrolls a passkey. adminHostname is the
 // operator-console host the URL points at (op.console.<root>): the Owner is staff,

cmd/felis/breakglass_test.go

@@ -19,14 +19,15 @@ import (
 // terminal. The design is passwordless: accounts carry no credential, and the
 // Owner completes first-login through the setup-token web flow.
 type fakeOwnerStore struct {
-	upserts  []upsertCall
+	upserts   []upsertCall
-	inserts  []upsertCall
+	inserts   []upsertCall
-	settings map[string][]byte
+	settings  map[string][]byte
-	audits   []api.AuditEntry
+	audits    []api.AuditEntry
-	tokens   []setupTokenCall
+	tokens    []setupTokenCall
-	redeems  []redeemCall
+	redeems   []redeemCall
-	users    map[string]*api.StaffUser // keyed by username
+	users     map[string]*api.StaffUser // keyed by username
-	admins   bool                      // AdminExists answer
+	admins    bool                      // AdminExists answer
+	ownerSeat string                    // OwnerUsername answer: the occupied seat, "" when none
 
 	// CompleteOwnerSetup's success result. redeemUserID defaults to the fresh id
 	// the caller passes (the unlinked-UUID case) when left empty.
@@ -40,6 +41,7 @@ type fakeOwnerStore struct {
 	auditErr       error
 	userErr        error // non-not-found error from UserByUsername
 	adminErr       error
+	seatErr        error
 	redeemErr      error
 	createTokenErr error
 }
@@ -80,6 +82,15 @@ func (f *fakeOwnerStore) UserByUsername(_ context.Context, username string) (*ap
 	return nil, api.ErrNotFound
 }
 
+// OwnerUsername reports the single active Owner seat. Tests set ownerSeat; the
+// zero value models a fresh install where bootstrap is free to mint.
+func (f *fakeOwnerStore) OwnerUsername(_ context.Context) (string, error) {
+	if f.seatErr != nil {
+		return "", f.seatErr
+	}
+	return f.ownerSeat, nil
+}
+
 func (f *fakeOwnerStore) UpsertOwner(_ context.Context, id, username, email string) error {
 	if f.upsertErr != nil {
 		return f.upsertErr
@@ -201,6 +212,34 @@ func TestProvisionOwner(t *testing.T) {
 		}
 	})
 
+	t.Run("an occupied seat refuses any other username", func(t *testing.T) {
+		// The seat is the single owner row: upserting a fresh name would take the
+		// insert arm and mint a SECOND owner, while the existing seat — possibly the
+		// compromised account this reset was meant to replace — stays live, and no
+		// supported path can delete an owner row.
+		f := &fakeOwnerStore{ownerSeat: "seat-holder"}
+		err := provisionOwner(ctx, f, "someone-else", "")
+		if !errors.Is(err, api.ErrConflict) {
+			t.Fatalf("error = %v, want it to wrap api.ErrConflict so the TUI routes back to the form", err)
+		}
+		if !strings.Contains(err.Error(), `"seat-holder"`) {
+			t.Errorf("error = %q, want it to name the occupied seat", err)
+		}
+		if len(f.upserts) != 0 {
+			t.Errorf("want no write against an occupied seat, got %d", len(f.upserts))
+		}
+	})
+
+	t.Run("the occupied seat's own username still resets", func(t *testing.T) {
+		f := &fakeOwnerStore{ownerSeat: "seat-holder"}
+		if err := provisionOwner(ctx, f, "seat-holder", "[email protected]"); err != nil {
+			t.Fatalf("provisionOwner(reset): %v", err)
+		}
+		if len(f.upserts) != 1 || f.upserts[0].username != "seat-holder" || f.upserts[0].email != "[email protected]" {
+			t.Fatalf("want 1 reset upsert for the seat, got %+v", f.upserts)
+		}
+	})
+
 	t.Run("propagates a store error", func(t *testing.T) {
 		f := &fakeOwnerStore{upsertErr: errors.New("boom")}
 		if err := provisionOwner(ctx, f, "owner", ""); err == nil {
@@ -264,6 +303,16 @@ func TestAuthenticateAdmin(t *testing.T) {
 		}
 	})
 
+	t.Run("the owner role attributes like an admin", func(t *testing.T) {
+		owner := mkAdmin("root")
+		owner.Role = "owner" // the platform owner is staff too (migration 0011)
+		f := &fakeOwnerStore{users: map[string]*api.StaffUser{"root": owner}}
+		matched, ok, err := authenticateAdmin(ctx, f, "root")
+		if err != nil || !ok || matched != "root" {
+			t.Fatalf("authenticateAdmin(owner) = (%q, %v, %v), want (root, true, nil)", matched, ok, err)
+		}
+	})
+
 	t.Run("an unknown user is a non-match, not an error", func(t *testing.T) {
 		f := &fakeOwnerStore{}
 		_, ok, err := authenticateAdmin(ctx, f, "nobody")

cmd/felis/converge.go

@@ -0,0 +1,73 @@
+package main
+
+import (
+	"context"
+	"errors"
+	"flag"
+	"fmt"
+	"io"
+	"os"
+	"strings"
+
+	"felis.lolicon.best/internal/config"
+	"felis.lolicon.best/internal/platform"
+)
+
+// cmdConverge is the explicit convergence pass over already-installed system
+// servers (#1). Provisioning is create-if-absent, so a field the desired spec
+// gained after an install (spec.rcon, spec.startup.healthHTTPPort, a derived env
+// key) never reaches the existing CR — and nothing says so. This command fills
+// exactly those zero-value fields; see convergeSystemServers for the full contract
+// and why it is a separate, operator-timed step rather than part of setup.
+//
+// It reads the same host config as setup (the control plane's felis.toml) and
+// talks to the cluster with the local kubeconfig, so it must run as root on the
+// control-plane host.
+func cmdConverge(args []string, stdout, stderr io.Writer) int {
+	fs := flag.NewFlagSet("converge", flag.ContinueOnError)
+	fs.SetOutput(stderr)
+	cfgPath := fs.String("config", defaultSetupConfigPath, "path to felis.toml")
+	if err := fs.Parse(args); err != nil {
+		if errors.Is(err, flag.ErrHelp) {
+			return 0
+		}
+		return 2
+	}
+	if os.Geteuid() != 0 {
+		fmt.Fprintln(stderr, "felis converge: refused — converging needs the cluster credentials, so it must run as root (try: sudo felis converge)")
+		return 1
+	}
+
+	cfg, err := config.Load(*cfgPath)
+	if err != nil {
+		fmt.Fprintf(stderr, "felis converge: %v\n", err)
+		fmt.Fprintln(stderr, "If this host was never installed, run `sudo felis setup` first.")
+		return 1
+	}
+	cl, err := buildSystemServerClient()
+	if err != nil {
+		fmt.Fprintf(stderr, "felis converge: %v\n", err)
+		return 1
+	}
+
+	controlNS := platform.DefaultControlNamespace
+	outcomes := convergeSystemServers(context.Background(), cl, cfg.K8s.Namespace,
+		cfg.Velocity.LoginImage, cfg.Velocity.LobbyImage,
+		platform.InternalAPIBaseURL(controlNS), cfg.Server.RootDomain,
+		defaultPanelHostname(cfg.Server.RootDomain, cfg.Auth.PanelHostname))
+
+	fmt.Fprintln(stdout, "felis converge: filling fields an installed system server predates (operator-set values are never overwritten):")
+	exit := 0
+	for _, o := range outcomes {
+		switch {
+		case o.err != nil:
+			fmt.Fprintf(stdout, "  - %s: ERROR %v\n", o.name, o.err)
+			exit = 1
+		case len(o.changes) > 0:
+			fmt.Fprintf(stdout, "  - %s: updated (%s)\n", o.name, strings.Join(o.changes, ", "))
+		default:
+			fmt.Fprintf(stdout, "  - %s: %s\n", o.name, o.skipped)
+		}
+	}
+	return exit
+}

cmd/felis/converge_test.go

@@ -0,0 +1,185 @@
+package main
+
+import (
+	"context"
+	"slices"
+	"strings"
+	"testing"
+
+	"felis.lolicon.best/internal/apis/felis/v1alpha1"
+	"felis.lolicon.best/internal/naming"
+	"sigs.k8s.io/controller-runtime/pkg/client"
+	"sigs.k8s.io/controller-runtime/pkg/client/fake"
+)
+
+// converge is the explicit pass over an installed system server whose CR predates
+// a field the desired spec has since gained (#1). It must fill exactly the
+// zero-valued whitelist fields and the derived env, and must not touch anything a
+// non-zero value already occupies — that is the operator's.
+func TestConvergeSystemServersFillsPredatedFields(t *testing.T) {
+	scheme := newSystemServerScheme(t)
+	ctx := context.Background()
+
+	// An old install: the lobby CR was created before the desired spec began
+	// rendering spec.rcon, and the login CR before the HTTP readiness gate existed.
+	// One derived env key is absent entirely (as if it were added later), and one
+	// hand-added env var plus a non-whitelisted spec field must survive.
+	lobby, err := lobbySystemServer("reg/lobby:1", "minecraft")
+	if err != nil {
+		t.Fatalf("build lobby: %v", err)
+	}
+	lobby.Spec.Rcon = v1alpha1.RconSpec{}
+	lobby.Spec.JavaMemory = "999Mi"
+
+	login, err := loginSystemServer("reg/limbo:1", "minecraft",
+		"http://felis-api.felis.svc.cluster.local:8081", "mc.example.net", "console.mc.example.net")
+	if err != nil {
+		t.Fatalf("build login: %v", err)
+	}
+	login.Spec.Startup.HealthHTTPPort = 0
+	kept := login.Spec.Env
+	login.Spec.Env = nil
+	for _, e := range kept {
+		if e.Name != envPanelHostname {
+			login.Spec.Env = append(login.Spec.Env, e)
+		}
+	}
+	login.Spec.Env = append(login.Spec.Env, v1alpha1.EnvVar{Name: "OPERATOR_TUNING", Value: "keep-me"})
+
+	cl := fake.NewClientBuilder().WithScheme(scheme).WithObjects(lobby, login).Build()
+	outcomes := convergeSystemServers(ctx, cl, "minecraft", "reg/limbo:1", "reg/lobby:1",
+		"http://felis-api.felis.svc.cluster.local:8081", "mc.example.net", "console.mc.example.net")
+
+	byName := map[string]systemServerOutcome{}
+	for _, o := range outcomes {
+		if o.err != nil {
+			t.Fatalf("%s: unexpected error: %v", o.name, o.err)
+		}
+		byName[o.name] = o
+	}
+	lobbyOut := byName[naming.SystemLobbyServer]
+	if len(lobbyOut.changes) != 1 || lobbyOut.changes[0] != "spec.rcon" {
+		t.Errorf("lobby changes = %v, want [spec.rcon] (only the zero-valued field)", lobbyOut.changes)
+	}
+	loginOut := byName[naming.SystemLoginServer]
+	if !slices.Contains(loginOut.changes, "spec.startup.healthHTTPPort") || !slices.Contains(loginOut.changes, "env "+envPanelHostname) {
+		t.Errorf("login changes = %v, want the health port plus the missing derived env key", loginOut.changes)
+	}
+
+	var gotLobby v1alpha1.MinecraftServer
+	if err := cl.Get(ctx, client.ObjectKey{Namespace: "minecraft", Name: naming.SystemLobbyServer}, &gotLobby); err != nil {
+		t.Fatalf("get lobby: %v", err)
+	}
+	if !gotLobby.Spec.Rcon.Enabled ||
+		gotLobby.Spec.Rcon.SecretRef.Name != naming.RconSecretName(naming.SystemLobbyServer) ||
+		gotLobby.Spec.Rcon.SecretRef.Key != naming.RconSecretKey {
+		t.Errorf("lobby rcon = %+v, want the desired block with the %s secret",
+			gotLobby.Spec.Rcon, naming.RconSecretName(naming.SystemLobbyServer))
+	}
+	if gotLobby.Spec.JavaMemory != "999Mi" {
+		t.Errorf("lobby javaMemory = %q, want 999Mi — converge fills new fields, it does not rewrite the spec", gotLobby.Spec.JavaMemory)
+	}
+
+	var gotLogin v1alpha1.MinecraftServer
+	if err := cl.Get(ctx, client.ObjectKey{Namespace: "minecraft", Name: naming.SystemLoginServer}, &gotLogin); err != nil {
+		t.Fatalf("get login: %v", err)
+	}
+	if gotLogin.Spec.Startup.HealthHTTPPort != felisLimboHealthPort {
+		t.Errorf("login healthHTTPPort = %d, want %d", gotLogin.Spec.Startup.HealthHTTPPort, felisLimboHealthPort)
+	}
+	env := map[string]string{}
+	for _, e := range gotLogin.Spec.Env {
+		env[e.Name] = e.Value
+	}
+	if env[envPanelHostname] != "console.mc.example.net" {
+		t.Errorf("%s was not added back: %q", envPanelHostname, env[envPanelHostname])
+	}
+	if env["OPERATOR_TUNING"] != "keep-me" {
+		t.Error("a hand-added env var was dropped; converge only touches config-derived names")
+	}
+}
+
+// A field already holding a non-zero value belongs to the operator: converge must
+// report "already converged" and write nothing.
+func TestConvergeSystemServersLeavesNonZeroFieldsAlone(t *testing.T) {
+	scheme := newSystemServerScheme(t)
+	ctx := context.Background()
+
+	lobby, err := lobbySystemServer("reg/lobby:1", "minecraft")
+	if err != nil {
+		t.Fatalf("build lobby: %v", err)
+	}
+	lobby.Spec.Rcon = v1alpha1.RconSpec{
+		Enabled:   true,
+		SecretRef: v1alpha1.SecretKeyRef{Name: "operator-rotated", Key: "password"},
+	}
+	login, err := loginSystemServer("reg/limbo:1", "minecraft",
+		"http://felis-api.felis.svc.cluster.local:8081", "mc.example.net", "console.mc.example.net")
+	if err != nil {
+		t.Fatalf("build login: %v", err)
+	}
+
+	cl := fake.NewClientBuilder().WithScheme(scheme).WithObjects(lobby, login).Build()
+	for _, o := range convergeSystemServers(ctx, cl, "minecraft", "reg/limbo:1", "reg/lobby:1",
+		"http://felis-api.felis.svc.cluster.local:8081", "mc.example.net", "console.mc.example.net") {
+		if o.err != nil {
+			t.Fatalf("%s: unexpected error: %v", o.name, o.err)
+		}
+		if len(o.changes) != 0 || o.skipped != "already converged" {
+			t.Errorf("%s outcome = %+v, want already converged with no writes", o.name, o)
+		}
+	}
+	var got v1alpha1.MinecraftServer
+	if err := cl.Get(ctx, client.ObjectKey{Namespace: "minecraft", Name: naming.SystemLobbyServer}, &got); err != nil {
+		t.Fatalf("get lobby: %v", err)
+	}
+	if got.Spec.Rcon.SecretRef.Name != "operator-rotated" {
+		t.Errorf("lobby rcon secretRef = %q — converge overwrote a field the operator had already set",
+			got.Spec.Rcon.SecretRef.Name)
+	}
+}
+
+// Guards: an absent CR is reported (creation is setup's job), a foreign CR is
+// refused rather than adopted, and an unset image skips like the provisioner does.
+func TestConvergeSystemServersGuards(t *testing.T) {
+	scheme := newSystemServerScheme(t)
+	ctx := context.Background()
+	run := func(cl client.Client, loginImage, lobbyImage string) []systemServerOutcome {
+		return convergeSystemServers(ctx, cl, "minecraft", loginImage, lobbyImage,
+			"http://felis-api.felis.svc.cluster.local:8081", "mc.example.net", "console.mc.example.net")
+	}
+
+	t.Run("absent CRs are reported, not created", func(t *testing.T) {
+		cl := fake.NewClientBuilder().WithScheme(scheme).Build()
+		for _, o := range run(cl, "reg/limbo:1", "reg/lobby:1") {
+			if o.err != nil {
+				t.Fatalf("%s: %v", o.name, o.err)
+			}
+			if o.created || !strings.Contains(o.skipped, "not present") {
+				t.Errorf("%s outcome = %+v, want a not-present skip", o.name, o)
+			}
+		}
+	})
+
+	t.Run("foreign CR is refused", func(t *testing.T) {
+		foreign := &v1alpha1.MinecraftServer{}
+		foreign.Name = naming.SystemLoginServer
+		foreign.Namespace = "minecraft"
+		cl := fake.NewClientBuilder().WithScheme(scheme).WithObjects(foreign).Build()
+		out := run(cl, "reg/limbo:1", "")
+		if len(out) != 2 {
+			t.Fatalf("outcomes = %d, want 2", len(out))
+		}
+		if out[0].err == nil || !strings.Contains(out[0].err.Error(), "not marked") {
+			t.Fatalf("login error = %v, want an unmarked-name refusal", out[0].err)
+		}
+	})
+
+	t.Run("unset image skips", func(t *testing.T) {
+		cl := fake.NewClientBuilder().WithScheme(scheme).Build()
+		out := run(cl, "", "reg/lobby:1")
+		if out[0].skipped != "image not configured" {
+			t.Errorf("login skipped = %q, want %q", out[0].skipped, "image not configured")
+		}
+	})
+}

cmd/felis/fetchcontext.go

@@ -0,0 +1,202 @@
+package main
+
+import (
+	"archive/tar"
+	"compress/gzip"
+	"context"
+	"errors"
+	"flag"
+	"fmt"
+	"io"
+	"net/http"
+	"os"
+	"os/signal"
+	"path/filepath"
+	"strings"
+	"syscall"
+	"time"
+)
+
+// cmdFetchContext is the in-Pod entrypoint the build Job's context-fetch
+// initContainer runs. It reads the blob the platform stored for a submission
+// from the felis-api INTERNAL face (with a bounded retry — see
+// fetchContextWithRetry) and extracts it into the shared emptyDir the Kaniko
+// container then builds from.
+//
+// Why this exists: the build Pod runs in the build namespace, where it can neither
+// mount the control-plane uploads PVC (a PVC does not cross namespaces) nor hold
+// object-store credentials, so the API that WROTE the blob is the transport. The
+// route is service-token-gated; the token arrives through a namespace-local Secret
+// mounted only into this initContainer, never into Kaniko's — so the untrusted
+// Dockerfile's build steps have no credential to read (their containers share no
+// environment, no PID namespace, and Kaniko itself mounts the context read-only).
+//
+// The extraction is deliberately paranoid: the tarball is attacker-controlled
+// input, so absolute paths, ".." escapes, links, and special files are refused
+// rather than sanitized. Kaniko treats the extracted tree as hostile regardless
+// (spec §16), but the pod's own filesystem still must not be written outside the
+// context directory it was given.
+func cmdFetchContext(args []string, _, stderr io.Writer) int {
+	fs := flag.NewFlagSet("fetch-context", flag.ContinueOnError)
+	fs.SetOutput(stderr)
+	url := fs.String("url", "", "internal-face URL of the submission's build-context tarball")
+	out := fs.String("out", "/context", "directory to extract the build context into")
+	if err := fs.Parse(args); err != nil {
+		return 2
+	}
+	if *url == "" {
+		fmt.Fprintln(stderr, "felis fetch-context: --url is required")
+		return 2
+	}
+	token := os.Getenv("FELIS_SERVICE_TOKEN")
+	if token == "" {
+		fmt.Fprintln(stderr, "felis fetch-context: FELIS_SERVICE_TOKEN is empty — the internal face rejects anonymous reads")
+		return 2
+	}
+
+	ctx, stop := signal.NotifyContext(context.Background(), os.Interrupt, syscall.SIGTERM)
+	defer stop()
+
+	// Validate the URL once up front: a bad one is a usage error (2), not
+	// something to sit in the retry loop.
+	if _, err := http.NewRequest(http.MethodGet, *url, nil); err != nil {
+		fmt.Fprintf(stderr, "felis fetch-context: bad --url: %v\n", err)
+		return 2
+	}
+	// No overall client timeout: a legitimate modpack context can be large and the
+	// Job's activeDeadlineSeconds is the real bound. The header timeout catches a
+	// wedged endpoint without capping a healthy download.
+	client := &http.Client{Transport: &http.Transport{ResponseHeaderTimeout: time.Minute}}
+	resp, err := fetchContextWithRetry(ctx, client, *url, token, stderr)
+	if err != nil {
+		fmt.Fprintf(stderr, "felis fetch-context: %v\n", err)
+		return 1
+	}
+	defer resp.Body.Close()
+
+	if err := extractTarGz(resp.Body, *out); err != nil {
+		fmt.Fprintf(stderr, "felis fetch-context: %v\n", err)
+		return 1
+	}
+	return 0
+}
+
+// fetchRetryInterval/fetchRetryWindow bound how long the fetch waits out a
+// control-plane blip before giving up. The api pod being replaced is a normal
+// event (rollout, eviction, a chaos drill), and without a retry one refused
+// dial turns it into a failed build: BackoffLimit=0 gives the Job no second
+// Pod, so the terminal verdict costs a manual re-approval — the live drill hit
+// exactly this (context-fetch exit 1 on `connect: connection refused` while
+// the api pod rolled; the new pod was serving 11 seconds later and the same
+// 198-byte blob). The window is tiny next to the Job's 30-minute
+// activeDeadline; a 4xx (missing blob, rejected token) still fails fast.
+//
+// Vars, not consts, so tests can shrink the window.
+var (
+	fetchRetryInterval = 3 * time.Second
+	fetchRetryWindow   = 45 * time.Second
+)
+
+// fetchContextWithRetry GETs the context tarball, retrying transport failures
+// and 5xx responses until fetchRetryWindow runs out. A 4xx is an answer, not a
+// blip — retrying it only delays the honest error.
+func fetchContextWithRetry(ctx context.Context, client *http.Client, url, token string, stderr io.Writer) (*http.Response, error) {
+	deadline := time.Now().Add(fetchRetryWindow)
+	for {
+		req, err := http.NewRequestWithContext(ctx, http.MethodGet, url, nil)
+		if err != nil {
+			return nil, fmt.Errorf("bad --url: %w", err)
+		}
+		req.Header.Set("Authorization", "Bearer "+token)
+
+		resp, err := client.Do(req)
+		if err == nil && resp.StatusCode == http.StatusOK {
+			return resp, nil
+		}
+		if err == nil {
+			status := resp.Status
+			_ = resp.Body.Close()
+			err = fmt.Errorf("GET returned %s", status)
+			if resp.StatusCode < 500 {
+				return nil, err
+			}
+		}
+		if ctx.Err() != nil {
+			return nil, fmt.Errorf("GET failed: %w", err)
+		}
+		if time.Now().After(deadline) {
+			return nil, fmt.Errorf("GET failed (retried for %s): %w", fetchRetryWindow, err)
+		}
+		fmt.Fprintf(stderr, "felis fetch-context: %v; retrying (the internal face may be restarting)\n", err)
+		select {
+		case <-ctx.Done():
+			return nil, fmt.Errorf("GET failed: %w", err)
+		case <-time.After(fetchRetryInterval):
+		}
+	}
+}
+
+// extractTarGz streams a gzip'd tarball into root, creating directories as
+// needed. Every entry is vetted BEFORE anything is written: a path that is
+// absolute or escapes root (via ".."), a link (symlink or hardlink), or any
+// special file kind aborts the whole extraction. Refusing rather than skipping is
+// deliberate — a context that needs one of those constructs is not a context this
+// transport carries, and silently dropping entries would build from a corpus the
+// submitter did not upload.
+func extractTarGz(r io.Reader, root string) error {
+	if err := os.MkdirAll(root, 0o755); err != nil {
+		return fmt.Errorf("create context dir: %w", err)
+	}
+	zr, err := gzip.NewReader(r)
+	if err != nil {
+		return fmt.Errorf("context is not a valid gzip tarball: %w", err)
+	}
+	defer zr.Close()
+	tr := tar.NewReader(zr)
+	for {
+		hdr, err := tr.Next()
+		if errors.Is(err, io.EOF) {
+			return nil
+		}
+		if err != nil {
+			return fmt.Errorf("read context tarball: %w", err)
+		}
+		name := filepath.Clean(hdr.Name)
+		if name == "." {
+			continue
+		}
+		// The zip-slip guard: reject, never rewrite. filepath.Clean collapses any
+		// "a/../../b", so these two checks are sufficient once Clean has run.
+		if filepath.IsAbs(name) || name == ".." || strings.HasPrefix(name, ".."+string(filepath.Separator)) {
+			return fmt.Errorf("context entry %q escapes the context directory", hdr.Name)
+		}
+		target := filepath.Join(root, name)
+		switch hdr.Typeflag {
+		case tar.TypeDir:
+			if err := os.MkdirAll(target, 0o755); err != nil {
+				return fmt.Errorf("create %q: %w", name, err)
+			}
+		case tar.TypeReg, tar.TypeRegA:
+			if err := os.MkdirAll(filepath.Dir(target), 0o755); err != nil {
+				return fmt.Errorf("create parent of %q: %w", name, err)
+			}
+			mode := os.FileMode(0o644)
+			if hdr.FileInfo().Mode()&0o111 != 0 {
+				mode = 0o755 // preserve executability (entrypoint scripts), nothing else
+			}
+			f, err := os.OpenFile(target, os.O_CREATE|os.O_WRONLY|os.O_TRUNC, mode)
+			if err != nil {
+				return fmt.Errorf("create %q: %w", name, err)
+			}
+			if _, err := io.Copy(f, tr); err != nil {
+				_ = f.Close()
+				return fmt.Errorf("write %q: %w", name, err)
+			}
+			if err := f.Close(); err != nil {
+				return fmt.Errorf("close %q: %w", name, err)
+			}
+		default:
+			return fmt.Errorf("context entry %q has unsupported type %q (links and special files are refused)", hdr.Name, string(hdr.Typeflag))
+		}
+	}
+}

cmd/felis/fetchcontext_test.go

@@ -0,0 +1,315 @@
+package main
+
+import (
+	"archive/tar"
+	"bytes"
+	"compress/gzip"
+	"io"
+	"net"
+	"net/http"
+	"net/http/httptest"
+	"os"
+	"path/filepath"
+	"strings"
+	"sync/atomic"
+	"testing"
+	"time"
+)
+
+type tarEntry struct {
+	name     string
+	body     string
+	mode     int64
+	typ      byte
+	linkname string
+}
+
+// tgzBody builds an in-memory .tar.gz from entries, preserving each entry's type
+// and mode so the tests can exercise the guards with exactly the bytes an
+// attacker could upload.
+func tgzBody(t *testing.T, entries ...tarEntry) []byte {
+	t.Helper()
+	var buf bytes.Buffer
+	zw := gzip.NewWriter(&buf)
+	tw := tar.NewWriter(zw)
+	for _, e := range entries {
+		typ := e.typ
+		if typ == 0 {
+			typ = tar.TypeReg
+		}
+		mode := e.mode
+		if mode == 0 {
+			mode = 0o644
+		}
+		hdr := &tar.Header{Name: e.name, Typeflag: typ, Mode: mode, Size: int64(len(e.body))}
+		if typ == tar.TypeSymlink {
+			hdr.Linkname = e.linkname
+			hdr.Size = 0
+		}
+		if err := tw.WriteHeader(hdr); err != nil {
+			t.Fatalf("write header %q: %v", e.name, err)
+		}
+		if hdr.Size > 0 {
+			if _, err := tw.Write([]byte(e.body)); err != nil {
+				t.Fatalf("write body %q: %v", e.name, err)
+			}
+		}
+	}
+	if err := tw.Close(); err != nil {
+		t.Fatalf("close tar: %v", err)
+	}
+	if err := zw.Close(); err != nil {
+		t.Fatalf("close gzip: %v", err)
+	}
+	return buf.Bytes()
+}
+
+// A normal context extracts with its tree intact, and the executable bit that
+// modpack entrypoints rely on survives.
+func TestExtractTarGzRoundTrip(t *testing.T) {
+	dir := t.TempDir()
+	body := tgzBody(t,
+		tarEntry{name: "Dockerfile", body: "FROM scratch\n"},
+		tarEntry{name: "mods/example.jar", body: "jar-bytes"},
+		tarEntry{name: "start.sh", body: "#!/bin/sh\n", mode: 0o755},
+		tarEntry{name: "mods/", typ: tar.TypeDir, mode: 0o755},
+	)
+	if err := extractTarGz(bytes.NewReader(body), dir); err != nil {
+		t.Fatalf("extract: %v", err)
+	}
+	for name, want := range map[string]string{
+		"Dockerfile":       "FROM scratch\n",
+		"mods/example.jar": "jar-bytes",
+	} {
+		got, err := os.ReadFile(filepath.Join(dir, name))
+		if err != nil || string(got) != want {
+			t.Fatalf("%s = (%q, %v), want %q", name, got, err, want)
+		}
+	}
+	fi, err := os.Stat(filepath.Join(dir, "start.sh"))
+	if err != nil || fi.Mode()&0o111 == 0 {
+		t.Fatalf("entrypoint script lost its exec bit: %v (%v)", fi, err)
+	}
+}
+
+// The guards: "..", absolute paths, symlinks, and special files are refused whole
+// — nothing escapes, and nothing is silently skipped.
+func TestExtractTarGzRefusesEscapes(t *testing.T) {
+	cases := []struct {
+		name    string
+		entries []tarEntry
+	}{
+		{"dotdot", []tarEntry{{name: "../outside", body: "x"}}},
+		{"nested dotdot", []tarEntry{{name: "a/../../outside", body: "x"}}},
+		{"absolute", []tarEntry{{name: "/etc/outside", body: "x"}}},
+		{"symlink", []tarEntry{{name: "link", typ: tar.TypeSymlink, linkname: "/etc"}}},
+		{"hardlink", []tarEntry{{name: "hard", typ: tar.TypeLink, linkname: "somewhere"}}},
+		{"device", []tarEntry{{name: "dev", typ: tar.TypeChar}}},
+	}
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			dir := t.TempDir()
+			if err := extractTarGz(bytes.NewReader(tgzBody(t, tc.entries...)), dir); err == nil {
+				t.Fatal("extract accepted a hostile entry, want an error")
+			}
+			// Nothing may have been written outside the target (or at all).
+			entries, _ := os.ReadDir(dir)
+			if len(entries) != 0 {
+				t.Fatalf("hostile archive left %d entries behind", len(entries))
+			}
+		})
+	}
+}
+
+// The command end to end: it dials the URL with the bearer token from the
+// environment, and refuses to run without it (the internal face would 401
+// anyway; failing at parse time is the honest earlier error).
+func TestCmdFetchContextFetchAndExtract(t *testing.T) {
+	body := tgzBody(t, tarEntry{name: "Dockerfile", body: "FROM scratch\n"})
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if r.Header.Get("Authorization") != "Bearer test-token" {
+			w.WriteHeader(http.StatusUnauthorized)
+			return
+		}
+		w.Header().Set("Content-Type", "application/gzip")
+		_, _ = w.Write(body)
+	}))
+	defer srv.Close()
+
+	dir := t.TempDir()
+	t.Setenv("FELIS_SERVICE_TOKEN", "test-token")
+	if code := cmdFetchContext([]string{"--url=" + srv.URL + "/sub-1/context", "--out=" + dir}, io.Discard, io.Discard); code != 0 {
+		t.Fatalf("cmdFetchContext exit = %d, want 0", code)
+	}
+	if got, err := os.ReadFile(filepath.Join(dir, "Dockerfile")); err != nil || string(got) != "FROM scratch\n" {
+		t.Fatalf("extracted Dockerfile = (%q, %v)", got, err)
+	}
+
+	// No token: refuse before dialing.
+	t.Setenv("FELIS_SERVICE_TOKEN", "")
+	var stderr bytes.Buffer
+	if code := cmdFetchContext([]string{"--url=" + srv.URL + "/sub-1/context", "--out=" + t.TempDir()}, io.Discard, &stderr); code != 2 {
+		t.Fatalf("missing token exit = %d, want 2 (stderr %q)", code, stderr.String())
+	}
+
+	// A non-200 answer (e.g. the route's 404 for a never-uploaded context) fails.
+	srv404 := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
+		w.WriteHeader(http.StatusNotFound)
+	}))
+	defer srv404.Close()
+	t.Setenv("FELIS_SERVICE_TOKEN", "test-token")
+	if code := cmdFetchContext([]string{"--url=" + srv404.URL + "/sub-1/context", "--out=" + t.TempDir()}, io.Discard, io.Discard); code != 1 {
+		t.Fatalf("404 exit = %d, want 1", code)
+	}
+}
+
+// A body that is not a gzip tarball must fail the extraction rather than produce
+// an empty (or partial) context Kaniko would then try to build.
+func TestExtractTarGzRejectsNonGzip(t *testing.T) {
+	dir := t.TempDir()
+	err := extractTarGz(strings.NewReader("not a tarball"), dir)
+	if err == nil || !strings.Contains(err.Error(), "gzip") {
+		t.Fatalf("err = %v, want a gzip complaint", err)
+	}
+}
+
+// shrinkFetchWindow swaps the retry knobs for a faster test and restores them
+// afterwards, so no test leaks a tiny window into another.
+func shrinkFetchWindow(t *testing.T, interval, window time.Duration) {
+	t.Helper()
+	oldInterval, oldWindow := fetchRetryInterval, fetchRetryWindow
+	fetchRetryInterval, fetchRetryWindow = interval, window
+	t.Cleanup(func() { fetchRetryInterval, fetchRetryWindow = oldInterval, oldWindow })
+}
+
+// A control-plane blip mid-fetch is survived: a 5xx on the first attempt is
+// retried and the second attempt's tarball extracts. This walks back the live
+// drill's failure, where the api pod rolled mid-fetch and the single attempt
+// died, failing the build Job.
+func TestFetchContextRetriesThroughBlip(t *testing.T) {
+	shrinkFetchWindow(t, 10*time.Millisecond, time.Second)
+	body := tgzBody(t, tarEntry{name: "Dockerfile", body: "FROM scratch\n"})
+	var calls int32
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
+		if atomic.AddInt32(&calls, 1) == 1 {
+			w.WriteHeader(http.StatusBadGateway) // the port is up, the API is not
+			return
+		}
+		_, _ = w.Write(body)
+	}))
+	defer srv.Close()
+
+	dir := t.TempDir()
+	t.Setenv("FELIS_SERVICE_TOKEN", "test-token")
+	var stderr bytes.Buffer
+	if code := cmdFetchContext([]string{"--url=" + srv.URL + "/sub-1/context", "--out=" + dir}, io.Discard, &stderr); code != 0 {
+		t.Fatalf("exit = %d, want 0 (stderr %q)", code, stderr.String())
+	}
+	if got, err := os.ReadFile(filepath.Join(dir, "Dockerfile")); err != nil || string(got) != "FROM scratch\n" {
+		t.Fatalf("extracted Dockerfile = (%q, %v)", got, err)
+	}
+	if !strings.Contains(stderr.String(), "retrying") {
+		t.Fatalf("stderr %q does not mention the retry", stderr.String())
+	}
+}
+
+// The live drill's exact shape: the dial itself is refused (the api pod is
+// gone and no endpoint answers). A refused dial is retried like any other
+// transport failure, and once the face is back the fetch completes.
+func TestFetchContextRetriesRefusedDial(t *testing.T) {
+	shrinkFetchWindow(t, 10*time.Millisecond, 5*time.Second)
+	body := tgzBody(t, tarEntry{name: "Dockerfile", body: "FROM scratch\n"})
+
+	// Borrow a listen address, then close it: the first attempts dial into a
+	// refused connection, exactly like a restarting control plane.
+	probe := httptest.NewServer(http.HandlerFunc(func(http.ResponseWriter, *http.Request) {}))
+	addr := strings.TrimPrefix(probe.URL, "http://")
+	probe.Close()
+
+	dir := t.TempDir()
+	t.Setenv("FELIS_SERVICE_TOKEN", "test-token")
+	var stderr bytes.Buffer
+	// Start the fetch; while the retry loop burns refused dials, bring the same
+	// address back.
+	result := make(chan int, 1)
+	go func() {
+		result <- cmdFetchContext([]string{"--url=http://" + addr + "/sub-1/context", "--out=" + dir}, io.Discard, &stderr)
+	}()
+	time.Sleep(100 * time.Millisecond) // let a handful of dials be refused
+	ln, err := net.Listen("tcp", addr)
+	if err != nil {
+		t.Fatalf("rebind %s: %v", addr, err)
+	}
+	back := &http.Server{Handler: http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if r.Header.Get("Authorization") != "Bearer test-token" {
+			w.WriteHeader(http.StatusUnauthorized)
+			return
+		}
+		_, _ = w.Write(body)
+	})}
+	defer back.Close()
+	go func() { _ = back.Serve(ln) }()
+
+	code := <-result
+	if code != 0 {
+		t.Fatalf("exit = %d, want 0 (stderr %q)", code, stderr.String())
+	}
+	if got, err := os.ReadFile(filepath.Join(dir, "Dockerfile")); err != nil || string(got) != "FROM scratch\n" {
+		t.Fatalf("extracted Dockerfile = (%q, %v)", got, err)
+	}
+	if !strings.Contains(stderr.String(), "retrying") {
+		t.Fatalf("stderr %q does not mention the retry", stderr.String())
+	}
+}
+
+// A 4xx is an answer, not a blip: a missing/never-uploaded context fails
+// immediately — no retry loop burns the build's deadline on a terminal error.
+func TestFetchContextDoesNotRetry4xx(t *testing.T) {
+	shrinkFetchWindow(t, 5*time.Millisecond, 200*time.Millisecond)
+	var calls int32
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
+		atomic.AddInt32(&calls, 1)
+		w.WriteHeader(http.StatusNotFound)
+	}))
+	defer srv.Close()
+
+	t.Setenv("FELIS_SERVICE_TOKEN", "test-token")
+	var stderr bytes.Buffer
+	if code := cmdFetchContext([]string{"--url=" + srv.URL + "/sub-1/context", "--out=" + t.TempDir()}, io.Discard, &stderr); code != 1 {
+		t.Fatalf("exit = %d, want 1 (stderr %q)", code, stderr.String())
+	}
+	if got := atomic.LoadInt32(&calls); got != 1 {
+		t.Fatalf("server saw %d attempts, want exactly 1", got)
+	}
+	if strings.Contains(stderr.String(), "retrying") {
+		t.Fatalf("stderr %q mentions a retry for a terminal 4xx", stderr.String())
+	}
+}
+
+// The retry is bounded: an internal face that stays down does not hang the
+// build pod; the window runs out and the fetch reports the exhausted retries.
+func TestFetchContextGivesUpAfterWindow(t *testing.T) {
+	shrinkFetchWindow(t, 5*time.Millisecond, 60*time.Millisecond)
+	var calls int32
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
+		atomic.AddInt32(&calls, 1)
+		w.WriteHeader(http.StatusServiceUnavailable)
+	}))
+	defer srv.Close() // the face is up but never healthy: 503 forever
+
+	t.Setenv("FELIS_SERVICE_TOKEN", "test-token")
+	var stderr bytes.Buffer
+	start := time.Now()
+	if code := cmdFetchContext([]string{"--url=" + srv.URL + "/sub-1/context", "--out=" + t.TempDir()}, io.Discard, &stderr); code != 1 {
+		t.Fatalf("exit = %d, want 1 (stderr %q)", code, stderr.String())
+	}
+	if elapsed := time.Since(start); elapsed > 5*time.Second {
+		t.Fatalf("gave up after %v; the window is supposed to bound it", elapsed)
+	}
+	if got := atomic.LoadInt32(&calls); got < 2 {
+		t.Fatalf("server saw %d attempts, want at least one retry", got)
+	}
+	if !strings.Contains(stderr.String(), "retried for") {
+		t.Fatalf("stderr %q does not report the exhausted retry window", stderr.String())
+	}
+}

cmd/felis/manifests.go

@@ -26,8 +26,10 @@ func (m *multiFlag) Set(v string) error {
 // felis-reaper identity only when the retention reaper is enabled, gated with
 // its CronJob), the weak build/restore Job SAs, the build/minecraft
 // NetworkPolicies, and the running control-plane workloads (felis-api/operator
-// Deployments + the in-cluster registry Deployment/Service/PVC) — as a single
+// Deployments + the in-cluster registry Deployment/Service/PVC + the
-// multi-document YAML stream on stdout, ready for `kubectl apply -f -`.
+// world-archive PVC that backs backup/restore, unless --backup-pvc is emptied)
+// — as a single multi-document YAML stream on stdout, ready for
+// `kubectl apply -f -`.
 //
 // It is a pure renderer: it never contacts a cluster and holds no credentials.
 // --velocity-cidr records the proxy host addresses allowed by the game NetworkPolicy.
@@ -44,9 +46,10 @@ func cmdManifests(args []string, stdout, stderr io.Writer) int {
 	panelNodePort := fs.Int("panel-node-port", int(platform.DefaultPanelNodePort), "NodePort that exposes the built-in HTTPS panel/API origin")
 	felisImage := fs.String("felis-image", "", "container image the felis-api/operator Deployments run, also passed through as FELIS_IMAGE (REQUIRED)")
 	registryImage := fs.String("registry-image", "", "in-cluster registry image (default: registry:2)")
-	backupPVC := fs.String("backup-pvc", "", "name of the backup PVC advertised to the restore executor via FELIS_BACKUP_PVC (default none = restore endpoint returns 503)")
+	backupPVC := fs.String("backup-pvc", "felis-backups", "name of the world-archive PVC this bundle renders in the Minecraft namespace and advertises to the backup/restore executors via FELIS_BACKUP_PVC (default: felis-backups; pass an empty value to render none, leaving backup/restore answering 503)")
-	worldsHostPath := fs.String("worlds-host-path", "", "node directory under which each world PVC is visible as <path>/<pvc>; enables the reaper CronJob (requires --backup-pvc and --archive-local-path)")
+	worldsHostPath := fs.String("worlds-host-path", "", "node directory the reaper reads worlds from: each world PVC resolves as <path>/<pvc>, or as the stock local-path directory <path>/<pv-name>_<ns>_<pvc-name> (k3s storage root: /var/lib/rancher/k3s/storage); enables the reaper CronJob (requires --archive-local-path and a non-empty --backup-pvc)")
 	archiveLocalPath := fs.String("archive-local-path", "", "path the backup PVC is mounted at in the reaper CronJob; MUST equal felis.toml [archive] local_path")
+	reaperNode := fs.String("reaper-node", "", "node that holds --worlds-host-path: pins the reaper CronJob's pod there via nodeSelector kubernetes.io/hostname (multi-node clusters need this, or the reaper may schedule where the hostPath is empty)")
 	var velocityCIDRs multiFlag
 	fs.Var(&velocityCIDRs, "velocity-cidr", "CIDR of a Velocity proxy host allowed to reach game port 25565 (repeatable, REQUIRED)")
 	var packageCIDRs multiFlag
@@ -81,36 +84,53 @@ func cmdManifests(args []string, stdout, stderr io.Writer) int {
 		fmt.Fprintf(stderr, "felis manifests: --panel-node-port must be in Kubernetes NodePort range 30000-32767 (got %d)\n", *panelNodePort)
 		return 2
 	}
+	// The node pin exists only for the reaper's hostPath: naming a node without the
+	// worlds root would be silently dropped (no CronJob renders), so fail loud like
+	// the storage-trio check below.
+	if *reaperNode != "" && *worldsHostPath == "" {
+		fmt.Fprintln(stderr, "felis manifests: --reaper-node requires --worlds-host-path "+
+			"(it pins the reaper CronJob, which renders only with the retention storage trio)")
+		return 2
+	}
 
-	// Retention/reaper rendering is opt-in and needs all three storage coordinates
+	// Retention/reaper rendering is opt-in and needs a storage topology together:
-	// together: where worlds live (to read+archive them), the backup PVC (to write
+	// where worlds live (to read+archive them), a backup PVC (to write archives
-	// archives into), and the path it is mounted at (which MUST equal felis.toml
+	// into — rendered from --backup-pvc), and the path it is mounted at (which MUST
-	// [archive] local_path so tarLocal's absolute archive refs resolve). A partial
+	// equal felis.toml [archive] local_path so tarLocal's absolute archive refs
-	// configuration is almost certainly an operator mistake, so fail loud rather than
+	// resolve). A partial configuration is almost certainly an operator mistake, so
-	// silently drop retention. Asking for it without the other two is rejected; an
+	// fail loud rather than silently drop retention or render a reaper with nowhere
-	// empty trio renders the bundle WITHOUT the reaper and says so.
+	// to write. The backup PVC itself defaults to felis-backups (it is what makes a
+	// default install's backup endpoint work at all); retention additionally needs
+	// --worlds-host-path.
 	if *worldsHostPath != "" {
 		if *backupPVC == "" || *archiveLocalPath == "" {
 			fmt.Fprintln(stderr, "felis manifests: --worlds-host-path enables the reaper CronJob and requires "+
-				"--backup-pvc and --archive-local-path too (--archive-local-path must equal felis.toml [archive] local_path)")
+				"--archive-local-path (must equal felis.toml [archive] local_path) and a non-empty --backup-pvc "+
+				"(the archive store; default felis-backups)")
 			return 2
 		}
-		// The reaper WILL render. Two deployment preconditions this generator cannot
+		// The reaper WILL render. Two deployment facts this generator cannot check
-		// check would SILENTLY turn retention into a no-op if unmet — surface them as
+		// would silently turn retention into a no-op if unmet — surface them as
 		// loudly as the fail-closed cases above, so an operator is never left with a
-		// reaper that reaps an empty directory. (Both are also in the WorldsHostPath
+		// reaper that reaps nothing. (Both are also in the WorldsHostPath flag/field
-		// flag/field docs, but nobody deploying from stdout reads those.)
+		// docs, but nobody deploying from stdout reads those.)
+		pin := "the CronJob sets NO nodeSelector: a single-node starter pins it to the worlds implicitly, but on a " +
+			"multi-node cluster you MUST pass --reaper-node <name> (or add a nodeSelector) for the node holding the " +
+			"worlds, or the reaper may schedule where the hostPath is empty"
+		if *reaperNode != "" {
+			pin = fmt.Sprintf("the CronJob is pinned to node %q via kubernetes.io/hostname — keep this pointed at the "+
+				"node that actually holds the world volumes", *reaperNode)
+		}
 		fmt.Fprintf(stderr, "felis manifests: note: rendering the retention reaper CronJob (worlds hostPath %q). "+
-			"Two preconditions are NOT verified here:\n"+
+			"These points are NOT verified here:\n"+
-			"  - each world PVC must be visible at %s/<pvc> on the node: a stock local-path-provisioner lays "+
+			"  - the node's world volumes must actually live below %s: the reaper resolves a world as "+
-			"volumes under PV-name paths (.../pvc-<uuid>_<ns>_<pvc>/), so unless the worlds StorageClass is "+
+			"%s/<pvc>, then as the stock local-path directory <path>/<pv-name>_<ns>_<pvc-name> (what k3s "+
-			"arranged to expose <path>/<pvc>, the reaper tars an empty directory;\n"+
+			"writes under /var/lib/rancher/k3s/storage). Any other provisioner needs its volumes exposed as "+
-			"  - the CronJob sets NO nodeSelector: a single-node starter pins it to the worlds implicitly, but "+
+			"<path>/<pvc>, or each candidate's archive fails and the world is preserved;\n"+
-			"on a multi-node cluster you MUST add a nodeSelector for the node holding the worlds, or the reaper "+
+			"  - %s.\n", *worldsHostPath, *worldsHostPath, *worldsHostPath, pin)
-			"may schedule where the hostPath is empty.\n", *worldsHostPath, *worldsHostPath)
 	} else {
 		fmt.Fprintln(stderr, "felis manifests: note: retention reaper CronJob not rendered "+
-			"(pass --worlds-host-path, --backup-pvc and --archive-local-path to enable it)")
+			"(pass --worlds-host-path and --archive-local-path — the archive PVC defaults to felis-backups — to enable it)")
 	}
 
 	out, err := platform.RenderYAML(platform.Params{
@@ -124,6 +144,7 @@ func cmdManifests(args []string, stdout, stderr io.Writer) int {
 		RegistryImage:      *registryImage,
 		BackupPVC:          *backupPVC,
 		WorldsHostPath:     *worldsHostPath,
+		ReaperNode:         *reaperNode,
 		ArchiveLocalPath:   *archiveLocalPath,
 		VelocityCIDRs:      []string(velocityCIDRs),
 		PackageSourceCIDRs: []string(packageCIDRs),

cmd/felis/manifests_test.go

@@ -77,6 +77,10 @@ func TestManifestsRendersBundle(t *testing.T) {
 		"10.0.0.5/32",
 		// The felis image flows through to the Deployments.
 		"registry.felis.svc:5000/felis:v1",
+		// Backup works out of the box: the archive PVC renders and the api gets
+		// the env that wires the backup/restore executors to it.
+		"name: felis-backups",
+		"name: FELIS_BACKUP_PVC",
 	} {
 		if !strings.Contains(text, want) {
 			t.Errorf("rendered bundle missing %q", want)
@@ -95,15 +99,19 @@ func TestManifestsRendersBundle(t *testing.T) {
 	}
 }
 
-// TestManifestsReaperRequiresTrio proves --worlds-host-path is a fail-loud opt-in:
+// TestManifestsReaperRequiresStorage proves --worlds-host-path is a fail-loud
-// asking for the reaper without the backup PVC and its mount path (which must equal
+// opt-in: asking for the reaper without a writable archive store (the backup PVC,
-// [archive] local_path) is rejected rather than silently dropping retention.
+// which defaults to felis-backups but can be emptied) and its mount path (which
-func TestManifestsReaperRequiresTrio(t *testing.T) {
+// must equal [archive] local_path) is rejected rather than silently dropping
+// retention or deleting worlds it could not archive first.
+func TestManifestsReaperRequiresStorage(t *testing.T) {
 	base := []string{"manifests", "--felis-image", "reg/felis:test", "--velocity-cidr", "10.0.0.5/32", "--worlds-host-path", "/var/lib/felis/worlds"}
 	for _, extra := range [][]string{
-		{},                                   // neither backup-pvc nor archive-local-path
+		{},                        // missing archive-local-path (backup-pvc defaults)
-		{"--backup-pvc", "felis-backups"},    // missing archive-local-path
+		{"--backup-pvc", "other"}, // still missing archive-local-path
-		{"--archive-local-path", "/backups"}, // missing backup-pvc
+		// A reaper with no archive store would have nowhere to write the archive
+		// it must verify before deleting a world; emptying the PVC is rejected.
+		{"--archive-local-path", "/backups", "--backup-pvc="},
 	} {
 		var out, errBuf bytes.Buffer
 		code := run(append(append([]string{}, base...), extra...), &out, &errBuf)
@@ -119,6 +127,23 @@ func TestManifestsReaperRequiresTrio(t *testing.T) {
 	}
 }
 
+// TestManifestsBackupPVCOptOut proves --backup-pvc= renders a bundle with no
+// archive store at all: no PVC and no FELIS_BACKUP_PVC env, so backup/restore
+// answer 503 instead of pointing Jobs at a claim nobody provisions.
+func TestManifestsBackupPVCOptOut(t *testing.T) {
+	var out, errBuf bytes.Buffer
+	code := run([]string{"manifests", "--felis-image", "reg/felis:test",
+		"--velocity-cidr", "10.0.0.5/32", "--backup-pvc="}, &out, &errBuf)
+	if code != 0 {
+		t.Fatalf("exit code = %d, want 0; stderr=%q", code, errBuf.String())
+	}
+	for _, absent := range []string{"felis-backups", "FELIS_BACKUP_PVC"} {
+		if strings.Contains(out.String(), absent) {
+			t.Errorf("--backup-pvc= bundle must not contain %q", absent)
+		}
+	}
+}
+
 // TestManifestsRendersReaper proves the happy path with the full retention trio:
 // a batch/v1 CronJob is emitted, named felis-reaper, mounting the backup PVC at the
 // supplied archive path.
@@ -150,9 +175,47 @@ func TestManifestsRendersReaper(t *testing.T) {
 	// this generator cannot verify (else a misarranged hostPath silently no-ops
 	// retention): the <path>/<pvc> arrangement-dependency and the multi-node
 	// nodeSelector hazard.
-	for _, want := range []string{"local-path-provisioner", "nodeSelector"} {
+	for _, want := range []string{"local-path", "nodeSelector"} {
 		if !strings.Contains(errBuf.String(), want) {
 			t.Errorf("reaper render must warn operators about %q on stderr, got %q", want, errBuf.String())
 		}
 	}
 }
+
+// TestManifestsReaperNodePin: --reaper-node pins the rendered CronJob's pod via
+// kubernetes.io/hostname and replaces the "no nodeSelector" hazard note with the
+// pin confirmation; using it without the worlds root is a fail-loud 2.
+func TestManifestsReaperNodePin(t *testing.T) {
+	var out, errBuf bytes.Buffer
+	code := run([]string{
+		"manifests",
+		"--felis-image", "registry.felis.svc:5000/felis:v1",
+		"--velocity-cidr", "10.0.0.5/32",
+		"--worlds-host-path", "/var/lib/felis/worlds",
+		"--archive-local-path", "/backups",
+		"--reaper-node", "node-a",
+	}, &out, &errBuf)
+	if code != 0 {
+		t.Fatalf("exit code = %d, want 0; stderr=%q", code, errBuf.String())
+	}
+	for _, want := range []string{
+		"kubernetes.io/hostname: node-a",
+	} {
+		if !strings.Contains(out.String(), want) {
+			t.Errorf("pinned render missing %q", want)
+		}
+	}
+	if !strings.Contains(errBuf.String(), "node-a") {
+		t.Errorf("stderr must confirm the pin, got %q", errBuf.String())
+	}
+
+	var out2, err2 bytes.Buffer
+	if code := run([]string{
+		"manifests",
+		"--felis-image", "registry.felis.svc:5000/felis:v1",
+		"--velocity-cidr", "10.0.0.5/32",
+		"--reaper-node", "node-a",
+	}, &out2, &err2); code != 2 {
+		t.Errorf("--reaper-node without --worlds-host-path: exit = %d, want 2", code)
+	}
+}

cmd/felis/operator.go

@@ -4,16 +4,19 @@ import (
 	"flag"
 	"fmt"
 	"io"
+	"log/slog"
 	"os"
 
 	"felis.lolicon.best/internal/apis/felis/v1alpha1"
 	felismetrics "felis.lolicon.best/internal/metrics"
 	"felis.lolicon.best/internal/operator"
+	"github.com/go-logr/logr"
 	"k8s.io/apimachinery/pkg/runtime"
 	utilruntime "k8s.io/apimachinery/pkg/util/runtime"
 	clientgoscheme "k8s.io/client-go/kubernetes/scheme"
 	ctrl "sigs.k8s.io/controller-runtime"
 	"sigs.k8s.io/controller-runtime/pkg/cache"
+	"sigs.k8s.io/controller-runtime/pkg/healthz"
 	ctrlmetrics "sigs.k8s.io/controller-runtime/pkg/metrics"
 	metricsserver "sigs.k8s.io/controller-runtime/pkg/metrics/server"
 )
@@ -25,6 +28,11 @@ func cmdOperator(args []string, _, stderr io.Writer) int {
 	fs := flag.NewFlagSet("operator", flag.ContinueOnError)
 	fs.SetOutput(stderr)
 	metricsAddr := fs.String("metrics-bind-address", ":8080", "address the metric endpoint binds to")
+	// healthAddr serves the manager's health endpoints (/healthz, /readyz) that the
+	// Deployment's probes dial. Without it the operator pod would carry no probe at
+	// all, and a wedged manager would keep its endpoint forever. It must differ from
+	// metricsAddr: the metrics server owns :8080.
+	healthAddr := fs.String("health-probe-bind-address", ":8081", "address the health probe endpoint binds to")
 	// namespace MUST equal the [k8s] namespace felis-api is configured with, and
 	// the deployment manifests (felis manifests) render both from one value. It
 	// scopes the manager's cache (informers) to a single namespace so the operator
@@ -42,9 +50,16 @@ func cmdOperator(args []string, _, stderr io.Writer) int {
 	utilruntime.Must(clientgoscheme.AddToScheme(scheme))
 	utilruntime.Must(v1alpha1.AddToScheme(scheme))
 
+	// controller-runtime logs through its own logr sink; without one, its first
+	// reconcile prints "log.SetLogger(...) was never called" ATTACHED TO A FULL
+	// GOROUTINE STACK — pure noise, not signal. Route it to slog's default handler
+	// so its messages appear as ordinary stderr lines.
+	ctrl.SetLogger(logr.FromSlogHandler(slog.Default().Handler()))
+
 	mgr, err := ctrl.NewManager(ctrl.GetConfigOrDie(), ctrl.Options{
-		Scheme:  scheme,
+		Scheme:                 scheme,
-		Metrics: metricsserver.Options{BindAddress: *metricsAddr},
+		Metrics:                metricsserver.Options{BindAddress: *metricsAddr},
+		HealthProbeBindAddress: *healthAddr,
 		// Scope every informer to the single watched namespace. Without this the
 		// cached client (mgr.GetClient) would LIST/WATCH cluster-wide, which a
 		// namespaced Role cannot grant — the operator would fail closed at runtime
@@ -60,6 +75,20 @@ func cmdOperator(args []string, _, stderr io.Writer) int {
 	}
 	fmt.Fprintf(stderr, "felis operator: watching namespace %q\n", *namespace)
 
+	// Register the two probe endpoints. controller-runtime only mounts /healthz and
+	// /readyz once at least one check is registered, so a bare listener would 404.
+	// The checks are the canonical always-pass ping: the probes' contract is "the
+	// manager process is up and serving", and a dependency hiccup (e.g. an API blip)
+	// must not restart the operator.
+	if err := mgr.AddHealthzCheck("ping", healthz.Ping); err != nil {
+		fmt.Fprintf(stderr, "felis operator: register healthz check: %v\n", err)
+		return 1
+	}
+	if err := mgr.AddReadyzCheck("ping", healthz.Ping); err != nil {
+		fmt.Fprintf(stderr, "felis operator: register readyz check: %v\n", err)
+		return 1
+	}
+
 	// Publish the named felis_* metrics (spec §23) on the endpoint the manager
 	// already serves (metricsAddr). controller-runtime's metrics server exposes
 	// its global Registry, so registering into it is all that is needed for

cmd/felis/reaper.go

@@ -1,9 +1,13 @@
 package main
 
 import (
+	"context"
+	"database/sql"
+	"errors"
 	"flag"
 	"fmt"
 	"io"
+	"os"
 	"path/filepath"
 	"strconv"
 	"strings"
@@ -12,8 +16,11 @@ import (
 	"felis.lolicon.best/internal/apis/felis/v1alpha1"
 	"felis.lolicon.best/internal/backup"
 	"felis.lolicon.best/internal/config"
+	"felis.lolicon.best/internal/mail"
+	"felis.lolicon.best/internal/platform"
 	"felis.lolicon.best/internal/reaper"
 	"felis.lolicon.best/internal/store"
+	corev1 "k8s.io/api/core/v1"
 	"k8s.io/apimachinery/pkg/runtime"
 	utilruntime "k8s.io/apimachinery/pkg/util/runtime"
 	clientgoscheme "k8s.io/client-go/kubernetes/scheme"
@@ -30,7 +37,7 @@ func cmdReaper(args []string, stdout, stderr io.Writer) int {
 	fs := flag.NewFlagSet("reaper", flag.ContinueOnError)
 	fs.SetOutput(stderr)
 	cfgPath := fs.String("config", "/etc/felis/felis.toml", "path to felis.toml")
-	worldsRoot := fs.String("worlds-root", "/worlds", "mount root under which world PVCs are visible (tarLocal: <root>/<pvc>)")
+	worldsRoot := fs.String("worlds-root", "/worlds", "mount root under which world PVCs are visible (tarLocal: <root>/<pvc>, else the stock local-path <root>/<pv-name>_<ns>_<pvc-name>)")
 	if err := fs.Parse(args); err != nil {
 		return 2
 	}
@@ -47,21 +54,8 @@ func cmdReaper(args []string, stdout, stderr io.Writer) int {
 		return 1
 	}
 
-	archiver, err := buildArchiver(cfg, *worldsRoot)
-	if err != nil {
-		fmt.Fprintf(stderr, "felis reaper: %v\n", err)
-		return 1
-	}
-
 	ctx := ctrl.SetupSignalHandler()
 
-	drv, err := store.Open(ctx, cfg.Database.URL)
-	if err != nil {
-		fmt.Fprintf(stderr, "felis reaper: open database: %v\n", err)
-		return 1
-	}
-	defer drv.Close()
-
 	scheme := runtime.NewScheme()
 	utilruntime.Must(clientgoscheme.AddToScheme(scheme))
 	utilruntime.Must(v1alpha1.AddToScheme(scheme))
@@ -71,6 +65,19 @@ func cmdReaper(args []string, stdout, stderr io.Writer) int {
 		return 1
 	}
 
+	archiver, err := buildArchiver(ctx, cfg, *worldsRoot, cl)
+	if err != nil {
+		fmt.Fprintf(stderr, "felis reaper: %v\n", err)
+		return 1
+	}
+
+	drv, err := store.Open(ctx, cfg.Database.URL)
+	if err != nil {
+		fmt.Fprintf(stderr, "felis reaper: open database: %v\n", err)
+		return 1
+	}
+	defer drv.Close()
+
 	r := &reaper.Reaper{
 		Cfg:      rcfg,
 		Store:    reaper.NewPGStore(drv.DB()),
@@ -78,6 +85,47 @@ func cmdReaper(args []string, stdout, stderr io.Writer) int {
 		Archiver: archiver,
 	}
 
+	// Pre-reap warnings go out by email when [smtp] is configured (the same
+	// relay and password_ref convention felis-api uses); without it the channel
+	// stays nil and the reaper logs each suppressed warning instead of stamping
+	// it, so a later SMTP setup still gets to warn. The owner must have a
+	// VERIFIED address — that flag is what proves the mailbox.
+	if cfg.SMTP.Host != "" {
+		passRef := cfg.SMTP.PasswordRef
+		if passRef == "" {
+			passRef = platform.SMTPPasswordEnv
+		}
+		password := os.Getenv(passRef)
+		if cfg.SMTP.Username != "" && password == "" {
+			fmt.Fprintf(stderr, "felis reaper: warning: [smtp] username is set but credentials env %s is empty — warning emails will fail AUTH\n", passRef)
+		}
+		db := drv.DB()
+		r.Warner = &mailWarner{
+			lookupEmail: func(ctx context.Context, ownerID string) (string, error) {
+				var email string
+				switch err := db.QueryRowContext(ctx,
+					`SELECT email FROM users
+					 WHERE id = $1 AND email_verified = true AND COALESCE(email, '') <> ''`,
+					ownerID).Scan(&email); {
+				case errors.Is(err, sql.ErrNoRows):
+					return "", fmt.Errorf("owner %s has no verified email", ownerID)
+				case err != nil:
+					return "", err
+				}
+				return email, nil
+			},
+			notifier: &mail.SMTP{
+				Host:     cfg.SMTP.Host,
+				Port:     cfg.SMTP.Port,
+				From:     cfg.SMTP.From,
+				Username: cfg.SMTP.Username,
+				Password: password,
+			},
+		}
+	} else {
+		fmt.Fprintln(stderr, "felis reaper: [smtp] not configured — pre-reap warnings are logged and NOT marked sent")
+	}
+
 	sum, err := r.RunOnce(ctx)
 	if err != nil {
 		fmt.Fprintf(stderr, "felis reaper: %v\n", err)
@@ -88,6 +136,38 @@ func cmdReaper(args []string, stdout, stderr io.Writer) int {
 	return 0
 }
 
+// mailWarner delivers a pre-reap notice to the owner's verified email — the
+// only channel this build can reach. Unowned owners and owners who never proved
+// a mailbox yield an error; the reaper retries such notices on its next run and
+// never lets them block the reap (red line ⑤).
+type mailWarner struct {
+	lookupEmail func(ctx context.Context, ownerID string) (string, error)
+	notifier    noticeNotifier
+}
+
+// noticeNotifier is the slice of mail.SMTP the warner needs (injected in tests).
+type noticeNotifier interface {
+	SendNotice(ctx context.Context, email, subject, body string) error
+}
+
+func (w *mailWarner) Warn(ctx context.Context, ownerID, server, remaining string) error {
+	email, err := w.lookupEmail(ctx, ownerID)
+	if err != nil {
+		return fmt.Errorf("resolve owner email: %w", err)
+	}
+	subject := fmt.Sprintf("Felis: 服务器 %s 将在 %s 后回收 · server reaped in %s", server, remaining, remaining)
+	body := fmt.Sprintf(
+		"Felis 世界回收提醒 / world-reaper notice\r\n"+
+			"\r\n"+
+			"服务器 / Server: %s\r\n"+
+			"距回收 / Time left: %s\r\n"+
+			"\r\n"+
+			"闲置的服务器会先自动备份，再释放世界；有人加入游戏即可重置倒计时。\r\n"+
+			"Idle servers are backed up and then released; any join resets the countdown.\r\n",
+		server, remaining)
+	return w.notifier.SendNotice(ctx, email, subject, body)
+}
+
 // reaperConfig derives the reaper's retention windows from felis.toml. The 15d
 // idle deadline is fixed by §18; only the warning offsets, retention, and the
 // store soft-cap are configurable (§24).
@@ -122,22 +202,57 @@ func reaperConfig(cfg *config.Config) (reaper.Config, error) {
 }
 
 // buildArchiver constructs the WorldArchiver. Only tarLocal is implemented in
-// this build; the resolver maps each world PVC to <worldsRoot>/<pvc>, the mount
+// this build; the resolver maps each world PVC to its directory under worldsRoot
-// convention the reaper Job is deployed with.
+// (resolveWorldDir).
-func buildArchiver(cfg *config.Config, worldsRoot string) (backup.WorldArchiver, error) {
+func buildArchiver(ctx context.Context, cfg *config.Config, worldsRoot string, cl client.Client) (backup.WorldArchiver, error) {
 	switch cfg.Archive.Store {
 	case "tarLocal":
 		return &backup.TarLocal{
 			BackupRoot: cfg.Archive.LocalPath,
-			Resolve: func(pvc string) (string, error) {
+			Resolve:    resolveWorldDir(ctx, cl, cfg.K8s.Namespace, worldsRoot),
-				return filepath.Join(worldsRoot, pvc), nil
-			},
 		}, nil
 	default:
 		return nil, fmt.Errorf("[archive] store %q is not implemented in this build (only tarLocal)", cfg.Archive.Store)
 	}
 }
 
+// resolveWorldDir maps a world PVC to its directory under worldsRoot, supporting
+// the two layouts a Felis host actually has:
+//
+//  1. <root>/<pvc> — the reaper's documented arrangement (worlds exposed by PVC
+//     name, e.g. via mounting each volume or a crafted storage class).
+//  2. <root>/<pv-name>_<namespace>_<pvc-name> — what a stock k3s install gets:
+//     local-path-provisioner stores every volume under its storage root as that
+//     exact directory name. Without this arm, retention on a default install could
+//     only ever fail to find a world (a no-op reaper, or worse an operator
+//     arranging paths by hand).
+//
+// The second path is derived EXACTLY from the live PVC's spec.volumeName, never
+// from a glob: a leftover directory of an old, deleted PV must never be mistaken
+// for the world the PVC currently binds, because the reaper archives the resolved
+// directory and then deletes that PVC — archiving stale bytes and deleting the
+// real world would be data loss. When neither path exists the first is returned,
+// so the archive walk fails loudly against the documented path.
+func resolveWorldDir(ctx context.Context, cl client.Client, namespace, worldsRoot string) backup.PVCResolver {
+	return func(pvc string) (string, error) {
+		direct := filepath.Join(worldsRoot, pvc)
+		if _, err := os.Stat(direct); err == nil {
+			return direct, nil
+		}
+		var claim corev1.PersistentVolumeClaim
+		if err := cl.Get(ctx, client.ObjectKey{Namespace: namespace, Name: pvc}, &claim); err != nil {
+			return "", fmt.Errorf("resolve world PVC %s: %w", pvc, err)
+		}
+		if pv := claim.Spec.VolumeName; pv != "" {
+			volDir := filepath.Join(worldsRoot, fmt.Sprintf("%s_%s_%s", pv, claim.Namespace, claim.Name))
+			if _, err := os.Stat(volDir); err == nil {
+				return volDir, nil
+			}
+		}
+		return direct, nil
+	}
+}
+
 // parseSpanDuration parses the human spans used in felis.toml's [archive] table:
 // "3mo" (months≈30d), "15d" (days), or any time.ParseDuration unit ("12h").
 func parseSpanDuration(s string) (time.Duration, error) {

cmd/felis/reaper_test.go

@@ -0,0 +1,123 @@
+package main
+
+import (
+	"context"
+	"errors"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+
+	corev1 "k8s.io/api/core/v1"
+	metav1 "k8s.io/apimachinery/pkg/apis/meta/v1"
+	"sigs.k8s.io/controller-runtime/pkg/client/fake"
+)
+
+// TestResolveWorldDir pins the two world layouts the reaper must find, and the
+// fail-closed miss. The stock local-path arm is derived from the live PVC's
+// volumeName — a name-based guess (glob) could tar a stale deleted PV's bytes and
+// then delete the current world, which is why it is read from the API instead.
+func TestResolveWorldDir(t *testing.T) {
+	ctx := context.Background()
+	root := t.TempDir()
+
+	// Arrange a world under the documented <root>/<pvc> layout.
+	named := filepath.Join(root, "world-named-0")
+	if err := os.MkdirAll(named, 0o750); err != nil {
+		t.Fatal(err)
+	}
+	// Arrange a second world the way k3s local-path stores it.
+	pvDir := filepath.Join(root, "pvc-11111111-2222-3333-4444-555555555555_minecraft_world-live-0")
+	if err := os.MkdirAll(pvDir, 0o750); err != nil {
+		t.Fatal(err)
+	}
+
+	claim := &corev1.PersistentVolumeClaim{
+		ObjectMeta: metav1.ObjectMeta{Name: "world-live-0", Namespace: "minecraft"},
+		Spec: corev1.PersistentVolumeClaimSpec{
+			VolumeName: "pvc-11111111-2222-3333-4444-555555555555",
+		},
+	}
+	cl := fake.NewClientBuilder().WithScheme(haltScheme(t)).WithObjects(claim).Build()
+	resolve := resolveWorldDir(ctx, cl, "minecraft", root)
+
+	t.Run("documented name layout wins", func(t *testing.T) {
+		got, err := resolve("world-named-0")
+		if err != nil || got != named {
+			t.Fatalf("resolve = (%q, %v), want (%q, nil)", got, err, named)
+		}
+	})
+
+	t.Run("stock local-path layout resolves exactly", func(t *testing.T) {
+		got, err := resolve("world-live-0")
+		if err != nil || got != pvDir {
+			t.Fatalf("resolve = (%q, %v), want (%q, nil)", got, err, pvDir)
+		}
+	})
+
+	t.Run("neither layout present falls back to the documented path", func(t *testing.T) {
+		// The claim exists but its directory does not: return the documented path so
+		// the archive walk fails there, and the reaper preserves the world.
+		missing := &corev1.PersistentVolumeClaim{
+			ObjectMeta: metav1.ObjectMeta{Name: "world-gone-0", Namespace: "minecraft"},
+			Spec:       corev1.PersistentVolumeClaimSpec{VolumeName: "pvc-99999999-0000-0000-0000-000000000000"},
+		}
+		cl := fake.NewClientBuilder().WithScheme(haltScheme(t)).WithObjects(missing).Build()
+		got, err := resolveWorldDir(ctx, cl, "minecraft", root)("world-gone-0")
+		if err != nil || got != filepath.Join(root, "world-gone-0") {
+			t.Fatalf("resolve = (%q, %v), want (%q, nil)", got, err, filepath.Join(root, "world-gone-0"))
+		}
+	})
+
+	t.Run("unknown pvc is an error, not a guess", func(t *testing.T) {
+		_, err := resolve("world-unknown-0")
+		if err == nil || !strings.Contains(err.Error(), "resolve world PVC world-unknown-0") {
+			t.Fatalf("err = %v, want a resolve-world-PVC error", err)
+		}
+	})
+}
+
+// The pre-reap warner resolves the owner's VERIFIED email and hands the notice
+// to the mailer. Every failure (no verified address, relay refusal) returns an
+// error so the reaper retries on its next run instead of stamping a notice
+// nobody received.
+func TestMailWarner(t *testing.T) {
+	lookup := func(email string, err error) func(context.Context, string) (string, error) {
+		return func(context.Context, string) (string, error) { return email, err }
+	}
+
+	n := &captureNotifier{}
+	w := &mailWarner{lookupEmail: lookup("[email protected]", nil), notifier: n}
+	if err := w.Warn(context.Background(), "u1", "survival", "3d"); err != nil {
+		t.Fatalf("Warn: %v", err)
+	}
+	if n.email != "[email protected]" || !strings.Contains(n.subject, "survival") || !strings.Contains(n.subject, "3d") {
+		t.Fatalf("notice envelope = (%q, %q)", n.email, n.subject)
+	}
+	if !strings.Contains(n.body, "survival") || !strings.Contains(n.body, "3d") {
+		t.Fatalf("body missing server/remaining:\n%s", n.body)
+	}
+
+	w = &mailWarner{lookupEmail: lookup("", errors.New("owner u2 has no verified email")), notifier: n}
+	if err := w.Warn(context.Background(), "u2", "survival", "3d"); err == nil || !strings.Contains(err.Error(), "verified email") {
+		t.Fatalf("unverified owner = %v, want the lookup error surfaced", err)
+	}
+
+	w = &mailWarner{lookupEmail: lookup("[email protected]", nil), notifier: &captureNotifier{err: errors.New("relay down")}}
+	if err := w.Warn(context.Background(), "u1", "survival", "3d"); err == nil || !strings.Contains(err.Error(), "relay down") {
+		t.Fatalf("relay failure = %v, want it surfaced", err)
+	}
+}
+
+type captureNotifier struct {
+	email, subject, body string
+	err                  error
+}
+
+func (n *captureNotifier) SendNotice(_ context.Context, email, subject, body string) error {
+	if n.err != nil {
+		return n.err
+	}
+	n.email, n.subject, n.body = email, subject, body
+	return nil
+}

cmd/felis/run.go

@@ -19,9 +19,11 @@ Commands:
   restore           Extract a world archive into a world volume (internal Job entrypoint)
   backup            Archive a world into the backup store and record it (internal Job entrypoint)
   files             List/read/write one file in a stopped server's world (internal Job entrypoint)
+  fetch-context     Fetch and extract a submission's build context (internal Job entrypoint)
   manifests         Render the control-plane RBAC + NetworkPolicy install bundle as YAML
   apply             Create a MinecraftServer CRD (direct K8s write; use -f server.json)
   setup             Run host bootstrap + first-run setup console (TUI; requires root/sudo)
+  converge          Fill in fields a newer desired spec added to already-installed system servers
   version           Print the build stamp of this binary
   update            Report which platform components have updates available
   breakGlass        Open the local break-glass emergency console (TUI; requires root/sudo)
@@ -47,9 +49,11 @@ var commands = map[string]func(args []string, stdout, stderr io.Writer) int{
 	"restore":          cmdRestore,
 	"backup":           cmdBackup,
 	"files":            cmdFiles,
+	"fetch-context":    cmdFetchContext,
 	"manifests":        cmdManifests,
 	"apply":            cmdApply,
 	"setup":            cmdSetup,
+	"converge":         cmdConverge,
 	"breakGlass":       cmdBreakGlass,
 	"bootstrap-assets": cmdBootstrapAssets,
 	"init-forwarding":  cmdInitForwarding,

cmd/felis/setup.go

@@ -233,13 +233,36 @@ func provisionSystemServers(ctx context.Context, cfg *config.Config, out io.Writ
 	// (the on-demand BACKUP Job runs in the minecraft namespace and mounts it to
 	// self-record its world_backups row; without the replica the Job's volume
 	// mount fails and every backup request strands in the cluster).
+	// An empty build_namespace means the build system's compiled-in default; the
+	// replica must target the namespace the Jobs actually run in.
+	buildNS := cfg.Registry.BuildNamespace
+	if buildNS == "" {
+		buildNS = platform.DefaultBuildNamespace
+	}
 	secretOutcomes := []systemServerOutcome{
 		ensureSecretReplica(ctx, cl, controlNS, cfg.K8s.Namespace,
-			naming.ServiceTokenSecretName, naming.ServiceTokenSecretKey, "service-token"),
+			naming.ServiceTokenSecretName, naming.ServiceTokenSecretKey, "service-token", "minecraft ns", false),
 		ensureSecretReplica(ctx, cl, controlNS, cfg.K8s.Namespace,
-			naming.ForwardingSecretName, naming.ForwardingSecretKey, "forwarding-secret"),
+			naming.ForwardingSecretName, naming.ForwardingSecretKey, "forwarding-secret", "minecraft ns", false),
+		// refresh=true: felis-config is the rendered config, not a credential. The
+		// backup/restore/fileedit Jobs and the reaper mount this copy, so a re-run
+		// must update it when the control plane's render has moved on (a stale copy
+		// e.g. keeps an old database URL after a credential rotation).
 		ensureSecretReplica(ctx, cl, controlNS, cfg.K8s.Namespace,
-			"felis-config", "felis.toml", "config"),
+			"felis-config", "felis.toml", "config", "minecraft ns", true),
+		// The reaper's pre-reap warning emails authenticate with the same relay
+		// password felis-api uses; the reaper pod runs in the minecraft namespace,
+		// where a secretKeyRef resolves only against a local mirror. Skipped while
+		// the relay is not configured yet — the "configure email" screen refreshes
+		// both mirrors when it applies.
+		ensureSecretReplica(ctx, cl, controlNS, cfg.K8s.Namespace,
+			"felis-smtp", "password", "smtp", "minecraft ns", false),
+		// The build namespace needs the same token: the build Job's fetch
+		// initContainer reads the submission context from the internal face. Best
+		// effort — a deployment that only installs the control plane simply never
+		// builds a user submission.
+		ensureSecretReplica(ctx, cl, controlNS, buildNS,
+			naming.ServiceTokenSecretName, naming.ServiceTokenSecretKey, "service-token", "felis-build ns", false),
 	}
 	outcomes := ensureSystemServers(ctx, cl, cfg.K8s.Namespace, cfg.Velocity.LoginImage, cfg.Velocity.LobbyImage, apiBaseURL, cfg.Server.RootDomain, defaultPanelHostname(cfg.Server.RootDomain, cfg.Auth.PanelHostname))
 	outcomes = append(secretOutcomes, outcomes...)
@@ -250,6 +273,8 @@ func provisionSystemServers(ctx context.Context, cfg *config.Config, out io.Writ
 			fmt.Fprintf(out, "  - %s: ERROR %v\n", o.name, o.err)
 		case o.created:
 			fmt.Fprintf(out, "  - %s: created (DesiredState=Running)\n", o.name)
+		case o.updated:
+			fmt.Fprintf(out, "  - %s: refreshed from the control namespace\n", o.name)
 		default:
 			fmt.Fprintf(out, "  - %s: skipped (%s)\n", o.name, o.skipped)
 		}

cmd/felis/systemservers.go

@@ -1,6 +1,7 @@
 package main
 
 import (
+	"bytes"
 	"context"
 	"fmt"
 	"time"
@@ -95,7 +96,7 @@ const felisLimboHealthPort int32 = 8080
 // fail-safes to readiness-only, so a login pod that has the URL/domain but not yet
 // the token is safe (it simply does not authenticate) rather than broken.
 const (
-	envAPIBaseURL    = "FELIS_API_BASE_URL"
+	envAPIBaseURL    = naming.EnvAPIBaseURL
 	envRootDomain    = "FELIS_ROOT_DOMAIN"
 	envPanelHostname = "FELIS_PANEL_HOSTNAME"
 	envLobbyServer   = "FELIS_LOBBY_SERVER"
@@ -255,10 +256,31 @@ func buildSystemServerClient() (client.Client, error) {
 // setup can report it without the provisioner deciding on the output format.
 type systemServerOutcome struct {
 	name      string
-	created   bool   // true = we created it this run
+	created   bool     // true = we created it this run
-	available bool   // true = the required object now exists
+	updated   bool     // true = we refreshed an existing replica from the source
-	skipped   string // non-empty = why it was skipped (image unset / already exists)
+	available bool     // true = the required object now exists
-	err       error  // non-nil = create failed
+	skipped   string   // non-empty = why it was skipped (image unset / already exists)
+	err       error    // non-nil = create failed
+	changes   []string // converge only: the fields this pass filled
+}
+
+// systemServerPlan is one system service in the provisioner's table: its name,
+// the image config gives it, and the pure builder for its desired CR.
+type systemServerPlan struct {
+	name  string
+	image string
+	build func(image, namespace string) (*v1alpha1.MinecraftServer, error)
+}
+
+// systemServerPlans is the single description of the login+lobby pair, shared by
+// ensureSystemServers (create-if-absent) and convergeSystemServers (field fill).
+func systemServerPlans(loginImage, lobbyImage, apiBaseURL, rootDomain, panelHostname string) []systemServerPlan {
+	return []systemServerPlan{
+		{name: naming.SystemLoginServer, image: loginImage, build: func(image, ns string) (*v1alpha1.MinecraftServer, error) {
+			return loginSystemServer(image, ns, apiBaseURL, rootDomain, panelHostname)
+		}},
+		{name: naming.SystemLobbyServer, image: lobbyImage, build: lobbySystemServer},
+	}
 }
 
 // ensureSystemServers idempotently creates the login and lobby system services.
@@ -269,17 +291,7 @@ type systemServerOutcome struct {
 // K8s client and namespace; this function performs no signal-handler or client
 // setup of its own.
 func ensureSystemServers(ctx context.Context, cl client.Client, namespace, loginImage, lobbyImage, apiBaseURL, rootDomain, panelHostname string) []systemServerOutcome {
-	type plan struct {
+	plans := systemServerPlans(loginImage, lobbyImage, apiBaseURL, rootDomain, panelHostname)
-		name  string
-		image string
-		build func(image, namespace string) (*v1alpha1.MinecraftServer, error)
-	}
-	plans := []plan{
-		{name: naming.SystemLoginServer, image: loginImage, build: func(image, ns string) (*v1alpha1.MinecraftServer, error) {
-			return loginSystemServer(image, ns, apiBaseURL, rootDomain, panelHostname)
-		}},
-		{name: naming.SystemLobbyServer, image: lobbyImage, build: lobbySystemServer},
-	}
 
 	outcomes := make([]systemServerOutcome, 0, len(plans))
 	for _, p := range plans {
@@ -379,12 +391,7 @@ var derivedSystemEnv = map[string]bool{
 // deliberate removal is indistinguishable from drift and re-adding it would fight the
 // operator every run.
 func refreshDerivedEnv(ctx context.Context, cl client.Client, existing, desired *v1alpha1.MinecraftServer) (bool, error) {
-	want := make(map[string]string, len(derivedSystemEnv))
+	want := derivedEnvWanted(desired)
-	for _, e := range desired.Spec.Env {
-		if derivedSystemEnv[e.Name] {
-			want[e.Name] = e.Value
-		}
-	}
 
 	changed := false
 	for i, e := range existing.Spec.Env {
@@ -402,6 +409,116 @@ func refreshDerivedEnv(ctx context.Context, cl client.Client, existing, desired
 	return true, nil
 }
 
+// derivedEnvWanted maps the derived env keys of desired onto their values.
+func derivedEnvWanted(desired *v1alpha1.MinecraftServer) map[string]string {
+	want := make(map[string]string, len(derivedSystemEnv))
+	for _, e := range desired.Spec.Env {
+		if derivedSystemEnv[e.Name] {
+			want[e.Name] = e.Value
+		}
+	}
+	return want
+}
+
+// convergeSystemServers is the explicit convergence pass over already-installed
+// system servers (#1). ensureSystemServers is create-if-absent by design — an
+// existing CR is left alone so a re-run cannot clobber an operator's edits — and
+// that leaves no path for a field the DESIRED spec gained after the install:
+// spec.rcon (the lobby's write channel), spec.startup.healthHTTPPort (the login
+// gate's readiness probe), or a config-derived env key that did not exist yet.
+// Such fields sit at their zero value forever while re-running setup reports
+// success, which is exactly the reported "configuration updates never reach an
+// installed deployment" symptom.
+//
+// This pass fills exactly those zero-value fields and the config-derived env keys,
+// and nothing else: a field already holding a non-zero value is the operator's and
+// is never overwritten. It is an explicit command rather than an implicit step of
+// setup because some fills need an ordering only the operator knows — enabling
+// RCON or the HTTP readiness gate on a server whose image predates the listener
+// would hold that server in Starting until it was marked Failed. Rebuild (or
+// upgrade) the images first, then run this.
+func convergeSystemServers(ctx context.Context, cl client.Client, namespace, loginImage, lobbyImage, apiBaseURL, rootDomain, panelHostname string) []systemServerOutcome {
+	outcomes := make([]systemServerOutcome, 0, 2)
+	for _, p := range systemServerPlans(loginImage, lobbyImage, apiBaseURL, rootDomain, panelHostname) {
+		if p.image == "" {
+			outcomes = append(outcomes, systemServerOutcome{name: p.name, skipped: "image not configured"})
+			continue
+		}
+		desired, err := p.build(p.image, namespace)
+		if err != nil {
+			outcomes = append(outcomes, systemServerOutcome{name: p.name, err: err})
+			continue
+		}
+
+		var existing v1alpha1.MinecraftServer
+		switch err := cl.Get(ctx, client.ObjectKeyFromObject(desired), &existing); {
+		case apierrors.IsNotFound(err):
+			outcomes = append(outcomes, systemServerOutcome{name: p.name,
+				skipped: "not present — run `sudo felis setup` first"})
+			continue
+		case err != nil:
+			outcomes = append(outcomes, systemServerOutcome{name: p.name, err: err})
+			continue
+		}
+		if existing.Labels[v1alpha1.LabelSystemRole] != p.name {
+			outcomes = append(outcomes, systemServerOutcome{name: p.name, err: fmt.Errorf(
+				"existing MinecraftServer %s/%s is not marked as the Felis %q system role; refusing to converge it",
+				namespace, p.name, p.name,
+			)})
+			continue
+		}
+
+		var changes []string
+		if existing.Spec.Rcon == (v1alpha1.RconSpec{}) && desired.Spec.Rcon != (v1alpha1.RconSpec{}) {
+			existing.Spec.Rcon = desired.Spec.Rcon
+			changes = append(changes, "spec.rcon")
+		}
+		if existing.Spec.Startup.HealthHTTPPort == 0 && desired.Spec.Startup.HealthHTTPPort != 0 {
+			existing.Spec.Startup.HealthHTTPPort = desired.Spec.Startup.HealthHTTPPort
+			changes = append(changes, "spec.startup.healthHTTPPort")
+		}
+		changes = append(changes, convergeDerivedEnv(&existing, desired)...)
+
+		if len(changes) == 0 {
+			outcomes = append(outcomes, systemServerOutcome{name: p.name, available: true, skipped: "already converged"})
+			continue
+		}
+		if err := cl.Update(ctx, &existing); err != nil {
+			outcomes = append(outcomes, systemServerOutcome{name: p.name, err: fmt.Errorf("converge %s: %w", p.name, err)})
+			continue
+		}
+		outcomes = append(outcomes, systemServerOutcome{name: p.name, available: true, updated: true, changes: changes})
+	}
+	return outcomes
+}
+
+// convergeDerivedEnv makes the config-derived env match the desired values: a key
+// whose value drifted is overwritten, and a key missing entirely is added. This is
+// the wider half of the same explicit pass — refreshDerivedEnv's present-only loop
+// can never introduce a NEW key, which is how a derived key added after an install
+// never reached it at all.
+func convergeDerivedEnv(existing, desired *v1alpha1.MinecraftServer) []string {
+	want := derivedEnvWanted(desired)
+	var changes []string
+	present := make(map[string]bool, len(existing.Spec.Env))
+	for i := range existing.Spec.Env {
+		e := &existing.Spec.Env[i]
+		present[e.Name] = true
+		if v, ok := want[e.Name]; ok && v != e.Value {
+			e.Value = v
+			changes = append(changes, "env "+e.Name)
+		}
+	}
+	for _, e := range desired.Spec.Env {
+		if !derivedSystemEnv[e.Name] || present[e.Name] {
+			continue
+		}
+		existing.Spec.Env = append(existing.Spec.Env, e)
+		changes = append(changes, "env "+e.Name)
+	}
+	return changes
+}
+
 // The login gate is a hard prerequisite of the Owner bind, so setup waits for it
 // rather than racing it. The ceiling covers a cold image pull on a fresh node;
 // the poll is fast enough that a warm start feels immediate.
@@ -471,26 +588,35 @@ func phaseOrPending(p v1alpha1.Phase) string {
 	return string(p)
 }
 
-// ensureSecretReplica copies one Secret from the control namespace into the minecraft
+// ensureSecretReplica copies one Secret from the control namespace into a workload
-// namespace so a backend pod can mount it via secretKeyRef. A secretKeyRef is
+// namespace (minecraft — or the build namespace, whose fetch initContainer reads the
-// namespace-local, but the backends run in the minecraft namespace while the sources
+// context from the felis-api internal face with the same token) so a pod can mount it
-// of truth live beside the control plane — so without this replica the operator's
+// via secretKeyRef. A secretKeyRef is namespace-local, but those workloads do not run
-// injected secretKeyRef would dangle and wedge the pod in CreateContainerConfigError.
+// beside the control plane — so without this replica the secretKeyRef would dangle and
+// wedge the pod in CreateContainerConfigError.
 //
-// Two Secrets need it, for different reasons: the service token (login only — it
+// Three Secrets need it, for different reasons: the service token (the login limbo and
-// authenticates the limbo plugin to the felis-api internal face) and the Velocity
+// the build Pod's context fetch — both authenticate to the felis-api internal face),
-// modern-forwarding secret (every backend — it is how a backend knows a login really
+// the Velocity modern-forwarding secret (every backend — it is how a backend knows
-// came from the proxy, and so that the player's UUID is Mojang-verified rather than
+// a login really came from the proxy, and so that the player's UUID is Mojang-verified
-// offline-derived).
+// rather than offline-derived), and the SMTP relay password (the reaper's pre-reap
+// warning emails; the felis-config mirror is what carries [smtp] into its pod).
 //
 // It is create-if-absent: an existing replica is left untouched so a hand-rotated
-// value in the minecraft namespace is never clobbered (to rotate, delete the replica
+// value in the workload namespace is never clobbered (to rotate, delete the replica
 // and re-run setup). Best-effort like the rest of the provisioner: a missing source or
 // a create failure degrades to a reported outcome, never a hard setup failure. It
 // copies only Type and Data — never labels/annotations/ownerRefs — so the replica
 // carries no accidental GC owner or managed-by lineage.
-func ensureSecretReplica(ctx context.Context, cl client.Client, controlNamespace, minecraftNamespace, secretName, secretKey, label string) systemServerOutcome {
+//
-	name := label + " (minecraft ns)"
+// refreshExisting switches the felis-config mirror to refresh-in-place: that Secret is
+// a rendered config, never a hand-rotated credential, and the workload Jobs that mount
+// it (backup/restore/fileedit) plus the reaper silently misbehave on a stale copy —
+// e.g. after a database credential rotation the control plane moves on while every
+// backup Job keeps failing auth. Credential Secrets keep the never-overwrite rule so a
+// rotated value survives; to rotate those, delete the replica and re-run setup.
+func ensureSecretReplica(ctx context.Context, cl client.Client, controlNamespace, minecraftNamespace, secretName, secretKey, label, where string, refreshExisting bool) systemServerOutcome {
+	name := label + " (" + where + ")"
 	validate := func(secret *corev1.Secret, location, skipped string) systemServerOutcome {
 		if len(secret.Data[secretKey]) == 0 {
 			return systemServerOutcome{name: name, skipped: fmt.Sprintf(
@@ -498,6 +624,33 @@ func ensureSecretReplica(ctx context.Context, cl client.Client, controlNamespace
 		}
 		return systemServerOutcome{name: name, available: true, skipped: skipped}
 	}
+	// refreshFromControl updates an existing replica from the control-namespace source
+	// when the rendered key differs. Only the felis-config mirror opts in.
+	refreshFromControl := func(existing *corev1.Secret) systemServerOutcome {
+		var src corev1.Secret
+		if err := cl.Get(ctx, client.ObjectKey{Namespace: controlNamespace, Name: secretName}, &src); err != nil {
+			if apierrors.IsNotFound(err) {
+				return systemServerOutcome{name: name, skipped: fmt.Sprintf(
+					"source Secret %s/%s not found — provision it (deploy/bootstrap.sh), then re-run setup",
+					controlNamespace, secretName)}
+			}
+			return systemServerOutcome{name: name, err: err}
+		}
+		if out := validate(&src, controlNamespace, ""); !out.available {
+			return out
+		}
+		if bytes.Equal(existing.Data[secretKey], src.Data[secretKey]) {
+			return validate(existing, minecraftNamespace, "already current")
+		}
+		if existing.Data == nil {
+			existing.Data = map[string][]byte{}
+		}
+		existing.Data[secretKey] = src.Data[secretKey]
+		if err := cl.Update(ctx, existing); err != nil {
+			return systemServerOutcome{name: name, err: err}
+		}
+		return systemServerOutcome{name: name, updated: true, available: true}
+	}
 	if controlNamespace == minecraftNamespace {
 		// Same namespace needs no replica, but the source still has to exist.
 		var existing corev1.Secret
@@ -516,6 +669,9 @@ func ensureSecretReplica(ctx context.Context, cl client.Client, controlNamespace
 	var existing corev1.Secret
 	getErr := cl.Get(ctx, client.ObjectKey{Namespace: minecraftNamespace, Name: secretName}, &existing)
 	if getErr == nil {
+		if refreshExisting {
+			return refreshFromControl(&existing)
+		}
 		return validate(&existing, minecraftNamespace, "already exists")
 	}
 	if !apierrors.IsNotFound(getErr) {
@@ -544,6 +700,9 @@ func ensureSecretReplica(ctx context.Context, cl client.Client, controlNamespace
 			if getErr := cl.Get(ctx, client.ObjectKey{Namespace: minecraftNamespace, Name: secretName}, &existing); getErr != nil {
 				return systemServerOutcome{name: name, err: getErr}
 			}
+			if refreshExisting {
+				return refreshFromControl(&existing)
+			}
 			return validate(&existing, minecraftNamespace, "already exists")
 		}
 		return systemServerOutcome{name: name, err: err}

cmd/felis/systemservers_test.go

@@ -156,7 +156,7 @@ func TestEnsureSecretReplica(t *testing.T) {
 	}
 	replicate := func(cl client.Client, controlNS, mcNS string) systemServerOutcome {
 		return ensureSecretReplica(ctx, cl, controlNS, mcNS,
-			naming.ServiceTokenSecretName, naming.ServiceTokenSecretKey, "service-token")
+			naming.ServiceTokenSecretName, naming.ServiceTokenSecretKey, "service-token", "minecraft ns", false)
 	}
 
 	t.Run("replicates when absent", func(t *testing.T) {
@@ -251,6 +251,87 @@ func TestEnsureSecretReplica(t *testing.T) {
 	})
 }
 
+// The felis-config mirror is the one replica that must refresh: it is a rendered
+// config, and a stale workload-side copy (backup/restore/fileedit Jobs, the reaper)
+// misbehaves silently — a rotated database credential keeps the control plane moving
+// while every backup Job keeps failing auth. Credential Secrets keep create-if-absent.
+func TestEnsureSecretReplicaRefresh(t *testing.T) {
+	scheme := newSystemServerScheme(t)
+	ctx := context.Background()
+	configSecret := func(ns, body string) *corev1.Secret {
+		return &corev1.Secret{
+			ObjectMeta: metav1.ObjectMeta{Name: "felis-config", Namespace: ns},
+			Type:       corev1.SecretTypeOpaque,
+			Data:       map[string][]byte{"felis.toml": []byte(body)},
+		}
+	}
+	refresh := func(cl client.Client) systemServerOutcome {
+		return ensureSecretReplica(ctx, cl, "felis", "minecraft",
+			"felis-config", "felis.toml", "config", "minecraft ns", true)
+	}
+	replicaBody := func(t *testing.T, cl client.Client) string {
+		t.Helper()
+		var got corev1.Secret
+		if err := cl.Get(ctx, client.ObjectKey{Namespace: "minecraft", Name: "felis-config"}, &got); err != nil {
+			t.Fatalf("get replica: %v", err)
+		}
+		return string(got.Data["felis.toml"])
+	}
+
+	t.Run("refreshes a stale config replica", func(t *testing.T) {
+		cl := fake.NewClientBuilder().WithScheme(scheme).WithObjects(
+			configSecret("felis", "current"),
+			configSecret("minecraft", "stale"),
+		).Build()
+		out := refresh(cl)
+		if out.err != nil || !out.updated || !out.available {
+			t.Fatalf("outcome = %+v, want refreshed", out)
+		}
+		if got := replicaBody(t, cl); got != "current" {
+			t.Errorf("replica = %q, want current", got)
+		}
+	})
+
+	t.Run("leaves a current config replica alone", func(t *testing.T) {
+		cl := fake.NewClientBuilder().WithScheme(scheme).WithObjects(
+			configSecret("felis", "same"),
+			configSecret("minecraft", "same"),
+		).Build()
+		out := refresh(cl)
+		if out.err != nil || out.updated || !out.available || out.skipped != "already current" {
+			t.Fatalf("outcome = %+v, want already current", out)
+		}
+	})
+
+	t.Run("fills an empty-key replica", func(t *testing.T) {
+		empty := &corev1.Secret{
+			ObjectMeta: metav1.ObjectMeta{Name: "felis-config", Namespace: "minecraft"},
+			Data:       map[string][]byte{},
+		}
+		cl := fake.NewClientBuilder().WithScheme(scheme).WithObjects(
+			configSecret("felis", "current"), empty).Build()
+		out := refresh(cl)
+		if out.err != nil || !out.updated {
+			t.Fatalf("outcome = %+v, want refreshed", out)
+		}
+		if got := replicaBody(t, cl); got != "current" {
+			t.Errorf("replica = %q, want current", got)
+		}
+	})
+
+	t.Run("missing source degrades to a skip", func(t *testing.T) {
+		cl := fake.NewClientBuilder().WithScheme(scheme).WithObjects(
+			configSecret("minecraft", "stale")).Build()
+		out := refresh(cl)
+		if out.err != nil || out.updated || out.available || out.skipped == "" {
+			t.Fatalf("outcome = %+v, want skipped (source missing)", out)
+		}
+		if got := replicaBody(t, cl); got != "stale" {
+			t.Errorf("replica = %q, want untouched stale", got)
+		}
+	})
+}
+
 func TestRequiredProvisioningError(t *testing.T) {
 	ready := []systemServerOutcome{
 		{name: "service-token (minecraft ns)", available: true},

cmd/felis/tui_backupnow.go

@@ -86,7 +86,7 @@ func (m *backupModel) loadCmd() tea.Cmd {
 		if err != nil {
 			return backupListMsg{err: fmt.Errorf("list servers: %w", err)}
 		}
-		return backupListMsg{cl: cl, servers: servers}
+		return backupListMsg{cl: cl, servers: backupPickable(servers)}
 	}
 }
 

cmd/felis/tui_edge_apply.go

@@ -133,6 +133,11 @@ func writeConfig(path string, cfg *config.Config) error {
 	return os.Rename(tmpPath, path)
 }
 
+// applyFelisConfigSecret applies the rendered config to the control namespace and
+// then converges the workload-namespace mirror best-effort. The mirror feeds the
+// backup/restore/fileedit Jobs and the reaper; without this refresh a reconfigure
+// here would leave those readers on the previous render until the next `felis
+// setup` run (startup pass) or installer re-run.
 func applyFelisConfigSecret(ctx context.Context) error {
 	out, err := kubectlOutput(ctx,
 		"-n", "felis", "create", "secret", "generic", "felis-config",
@@ -142,7 +147,41 @@ func applyFelisConfigSecret(ctx context.Context) error {
 	if err != nil {
 		return err
 	}
-	return kubectlWithInput(ctx, out, "apply", "-f", "-")
+	if err := kubectlWithInput(ctx, out, "apply", "-f", "-"); err != nil {
+		return err
+	}
+	if err := replicateFelisConfigToWorkloadNamespace(ctx); err != nil {
+		fmt.Fprintf(os.Stderr, "felis setup: warning: the control-plane config is applied, but the workload-namespace mirror could not be refreshed (%v); re-run felis setup once that is fixed\n", err)
+	}
+	return nil
+}
+
+// replicateFelisConfigToWorkloadNamespace overwrites the workload-namespace
+// felis-config mirror with the freshly rendered pod config. Deliberately a full
+// replace, not create-if-absent: a stale mirror is exactly what silently hands
+// the Jobs that mount it old settings after a reconfigure. No-op when the
+// workload namespace is unset or is the control namespace itself.
+func replicateFelisConfigToWorkloadNamespace(ctx context.Context) error {
+	cfg, err := config.Load(hostSetupConfigPath)
+	if err != nil {
+		return err
+	}
+	ns := cfg.K8s.Namespace
+	if ns == "" || ns == "felis" {
+		return nil
+	}
+	manifest, err := kubectlOutput(ctx,
+		"-n", ns, "create", "secret", "generic", "felis-config",
+		"--from-file=felis.toml="+podSetupConfigPath,
+		"--dry-run=client", "-o", "yaml",
+	)
+	if err != nil {
+		return fmt.Errorf("render felis-config for %s: %w", ns, err)
+	}
+	if err := kubectlWithInput(ctx, manifest, "-n", ns, "apply", "-f", "-"); err != nil {
+		return fmt.Errorf("replicate felis-config to %s: %w", ns, err)
+	}
+	return nil
 }
 
 func installCloudflaredService(ctx context.Context, cloudflaredBin, configPath string) error {

cmd/felis/tui_mc_bind.go

@@ -207,7 +207,11 @@ func (m *mcBindModel) doneView() string {
 		if box.Len() > 0 {
 			box.WriteString("\n")
 		}
-		box.WriteString(tuiLabel.Render("setup URL  ") + "\n" + tuiPassword.Render(m.setupTokenURL) + "\n\n")
+		box.WriteString(tuiLabel.Render("setup URL  ") + "\n")
+		for _, line := range wrapDisplayURL(m.setupTokenURL, 70) {
+			box.WriteString(tuiPassword.Render(line) + "\n")
+		}
+		box.WriteString("\n")
 		box.WriteString(tuiWarn.Render("Open this URL to complete passwordless login setup.\nIt is shown only once."))
 	}
 	if m.auditWarning != "" {

cmd/felis/tui_menu_test.go

@@ -151,19 +151,46 @@ func TestOwnerModelProvisionErrorRouting(t *testing.T) {
 		}
 	})
 
-	t.Run("a conflict on the Owner path is not a retry", func(t *testing.T) {
+	t.Run("the Owner seat refusal returns to the form naming the seat", func(t *testing.T) {
-		// Defensive: the Owner upserts and so never conflicts, but were one ever to
+		// Upserting a fresh username while a seat is occupied would mint a second
-		// surface it must end the session rather than loop the form — only the
+		// owner, so provisionOwner refuses with ownerSeatTakenError (Is
-		// insert-only operator path is retryable.
+		// api.ErrConflict) and the console must route back for a retype — the same
+		// recoverable contract as the operator clash, and the only Owner-path
+		// conflict there is.
 		m := newOwnerModel(ctx, &fakeOwnerStore{}, "root", true)
+		seatErr := &ownerSeatTakenError{seat: "seat-holder"}
 
-		next, cmd := m.Update(owProvisionMsg{err: conflict})
+		next, cmd := m.Update(owProvisionMsg{err: seatErr})
+		om := next.(*ownerModel)
+		if om.step != owProvision {
+			t.Fatalf("step = %v, want owProvision — the seat refusal is recoverable", om.step)
+		}
+		if om.provisionErr == nil || !errors.Is(om.provisionErr, api.ErrConflict) || !strings.Contains(om.provisionErr.Error(), "seat-holder") {
+			t.Errorf("provisionErr = %v, want the seat refusal naming the seat", om.provisionErr)
+		}
+		// Feed the rebuilt form's init message back through so its view renders;
+		// then the note must carry the seat name (the operator's retype cue).
+		if cmd != nil {
+			if msg := cmd(); msg != nil {
+				if n2, _ := om.Update(msg); n2 != nil {
+					om = n2.(*ownerModel)
+				}
+			}
+		}
+		if view := om.form.View(); !strings.Contains(view, "seat-holder") {
+			t.Errorf("the provision form must surface the seat refusal:\n%s", view)
+		}
+	})
+
+	t.Run("a generic Owner-path fault still tears the console down", func(t *testing.T) {
+		m := newOwnerModel(ctx, &fakeOwnerStore{}, "root", true)
+		next, cmd := m.Update(owProvisionMsg{err: errors.New("boom")})
 		om := next.(*ownerModel)
 		if om.provisionErr != nil {
-			t.Error("the Owner path recorded a retryable conflict; only the operator path retries")
+			t.Error("a generic fault must not be treated as a retryable refusal")
 		}
 		if res, ok := cmd().(ownerResultMsg); !ok || res.err == nil {
-			t.Error("an Owner-path conflict should tear down via an error result")
+			t.Error("a generic Owner-path fault should tear down via an error result")
 		}
 	})
 }

cmd/felis/tui_owner.go

@@ -174,12 +174,13 @@ func (m *ownerModel) Update(msg tea.Msg) (tea.Model, tea.Cmd) {
 
 	case owProvisionMsg:
 		if msg.err != nil {
-			// A taken Operator username is the expected, recoverable outcome of the
+			// api.ErrConflict marks the two recoverable refusals: a taken Operator
-			// insert-only operator path (refusing the clash is the whole reason it is
+			// username (insert-only clash) and an Owner reset naming anything but the
-			// insert-only, not an upsert). Route back to the form with a note so the
+			// occupied seat (ownerSeatTakenError Is ErrConflict). Route back to the
-			// operator can pick another name, rather than tearing down the console —
+			// form with a note so the operator can retype, rather than tearing down
-			// any other error is a genuine fault and still ends the session.
+			// the console — any other error is a genuine fault and still ends the
-			if m.operation == bgAddOperator && errors.Is(msg.err, api.ErrConflict) {
+			// session.
+			if errors.Is(msg.err, api.ErrConflict) {
 				m.provisionErr = msg.err
 				m.step = owProvision
 				m.form = m.sized(m.buildProvisionForm())
@@ -364,9 +365,15 @@ func (m *ownerModel) buildProvisionForm() *huh.Form {
 		}
 	}
 	if m.provisionErr != nil {
-		// The only error routed back to this form is a username clash on the insert-only
+		// Recoverable refusals routed back here: the seat refusal already names the
-		// operator path; show a concrete prompt to choose another name.
+		// username to enter, so show it verbatim; the operator-name clash gets the
-		desc = "That username is already taken — choose a different one.\n\n" + desc
+		// generic retry prompt.
+		note := "That username is already taken — choose a different one."
+		var seatErr *ownerSeatTakenError
+		if errors.As(m.provisionErr, &seatErr) {
+			note = seatErr.Error()
+		}
+		desc = note + "\n\n" + desc
 	}
 
 	fields := []huh.Field{
@@ -420,7 +427,11 @@ func (m *ownerModel) doneView() string {
 	var box strings.Builder
 	box.WriteString(tuiLabel.Render("username  ") + m.username + "\n")
 	if m.setupTokenURL != "" {
-		box.WriteString("\n" + tuiLabel.Render("setup URL  ") + "\n" + tuiPassword.Render(m.setupTokenURL) + "\n\n")
+		box.WriteString("\n" + tuiLabel.Render("setup URL  ") + "\n")
+		for _, line := range wrapDisplayURL(m.setupTokenURL, 70) {
+			box.WriteString(tuiPassword.Render(line) + "\n")
+		}
+		box.WriteString("\n")
 		box.WriteString(tuiWarn.Render("Open this URL to complete passwordless login setup. It is shown only once."))
 	}
 	if m.auditWarning != "" {

cmd/felis/tui_root.go

@@ -553,6 +553,7 @@ func (m *rootModel) showSummary() (tea.Model, tea.Cmd) {
 		storageLabel:  m.result.storageDetail,
 		routedHosts:   routed,
 		localHint:     m.result.connectMethod == connectLocal,
+		alreadySetUp:  m.result.alreadySetUp,
 	})
 }
 

cmd/felis/tui_root_test.go

@@ -249,6 +249,34 @@ func TestRootReconfigureSMTP(t *testing.T) {
 	}
 }
 
+// TestRootReconfigureStorageKeepsStatusFraming locks the same rule for the
+// "change storage" path: on a re-run, completing it must land back on the
+// alreadySetUp status framing (with the updated recap), not "Setup complete."
+func TestRootReconfigureStorageKeepsStatusFraming(t *testing.T) {
+	m := newTestRoot(true, consoleModeSetup, "")
+	m = drive(t, m, preflightDoneMsg{})
+	if _, ok := m.screen.(*summaryModel); !ok {
+		t.Fatalf("re-run after preflight, screen = %T, want *summaryModel", m.screen)
+	}
+
+	m = drive(t, m, reconfigureStorageMsg{})
+	if _, ok := m.screen.(*storageChooserModel); !ok {
+		t.Fatalf("reconfigure-storage screen = %T, want *storageChooserModel", m.screen)
+	}
+
+	m = drive(t, m, storageResultMsg{method: storageLocal, detail: "local disk · /var/lib/felis/uploads"})
+	sum, ok := m.screen.(*summaryModel)
+	if !ok {
+		t.Fatalf("after reconfigure-storage, screen = %T, want *summaryModel", m.screen)
+	}
+	if !sum.alreadySetUp {
+		t.Fatalf("after reconfigure-storage, summary should keep the alreadySetUp framing")
+	}
+	if sum.storageLabel != "local disk · /var/lib/felis/uploads" {
+		t.Fatalf("storageLabel = %q, want the updated recap", sum.storageLabel)
+	}
+}
+
 func TestRootRerunLandsOnStatus(t *testing.T) {
 	// adminExists at start of a setup run = re-run: preflight should skip straight
 	// to the "manage in panel" status screen, never touching owner/connect.

cmd/felis/tui_smtp.go

@@ -4,6 +4,7 @@ import (
 	"context"
 	"errors"
 	"fmt"
+	"os"
 	"strconv"
 	"strings"
 
@@ -338,20 +339,30 @@ func applySMTPConfig(ctx context.Context, in smtpInputs) error {
 	if err := applyFelisConfigSecret(ctx); err != nil {
 		return err
 	}
+	// Refresh the workload-namespace copies too (the reaper's warning path): the
+	// OTP path is already live in the control namespace, so a replica miss is
+	// reported but not fatal.
+	if err := replicateSMTPToWorkloadNamespace(ctx, in.password); err != nil {
+		fmt.Fprintf(os.Stderr, "felis setup: warning: email is configured, but refreshing the workload copies failed (pre-reap warning emails may stay suppressed): %v\n", err)
+	}
 	if err := kubectl(ctx, "-n", "felis", "rollout", "restart", "deployment/felis-api"); err != nil {
 		return err
 	}
 	return kubectl(ctx, "-n", "felis", "rollout", "status", "deployment/felis-api", "--timeout=180s")
 }
 
-// applySMTPSecret creates (or replaces) the felis-smtp Secret the felis-api
+// smtpSecretManifest renders the felis-smtp Secret for the given namespace, the
-// Deployment injects the relay password from. Rendered in-process and piped to
+// one the receiving Deployment/CronJob resolves its secretKeyRef against (felis
+// for felis-api, the workload namespace for the reaper's mirror). The namespace
+// must be IN the manifest: kubectl rejects a manifest whose namespace conflicts
+// with -n, so leaving the control namespace hardcoded made every workload-ns
+// replica fail before it started. Rendered in-process and piped to
 // `kubectl apply` — the password is never a command-line arg, so it never
 // appears in the host process table.
-func applySMTPSecret(ctx context.Context, password string) error {
+func smtpSecretManifest(password, namespace string) ([]byte, error) {
 	secret := &corev1.Secret{
 		TypeMeta:   metav1.TypeMeta{APIVersion: "v1", Kind: "Secret"},
-		ObjectMeta: metav1.ObjectMeta{Name: platform.SMTPSecretName, Namespace: "felis"},
+		ObjectMeta: metav1.ObjectMeta{Name: platform.SMTPSecretName, Namespace: namespace},
 		Type:       corev1.SecretTypeOpaque,
 		StringData: map[string]string{
 			platform.SMTPSecretPasswordKey: password,
@@ -359,7 +370,42 @@ func applySMTPSecret(ctx context.Context, password string) error {
 	}
 	manifest, err := yaml.Marshal(secret)
 	if err != nil {
-		return fmt.Errorf("render smtp secret: %w", err)
+		return nil, fmt.Errorf("render smtp secret: %w", err)
+	}
+	return manifest, nil
+}
+
+func applySMTPSecret(ctx context.Context, password string) error {
+	manifest, err := smtpSecretManifest(password, "felis")
+	if err != nil {
+		return err
 	}
 	return kubectlWithInput(ctx, manifest, "apply", "-f", "-")
 }
+
+// replicateSMTPToWorkloadNamespace refreshes the workload-namespace (minecraft)
+// copy of felis-smtp after email is reconfigured. The reaper's CronJob runs
+// there and resolves the password by local reference — a secretKeyRef is
+// namespace-local — so without this refresh a later SMTP change would never
+// reach the pre-reap warning emails. Deliberately OVERWRITES: this is a mirror
+// of the control-namespace source, and a stale mirror is exactly the failure
+// this closes. The felis-config mirror rides along in applyFelisConfigSecret,
+// which every apply path refreshes.
+func replicateSMTPToWorkloadNamespace(ctx context.Context, password string) error {
+	cfg, err := config.Load(hostSetupConfigPath)
+	if err != nil {
+		return err
+	}
+	ns := cfg.K8s.Namespace
+	if ns == "" || ns == "felis" {
+		return nil
+	}
+	smtpManifest, err := smtpSecretManifest(password, ns)
+	if err != nil {
+		return err
+	}
+	if err := kubectlWithInput(ctx, smtpManifest, "-n", ns, "apply", "-f", "-"); err != nil {
+		return fmt.Errorf("replicate %s to %s: %w", platform.SMTPSecretName, ns, err)
+	}
+	return nil
+}

cmd/felis/tui_smtp_test.go

@@ -0,0 +1,37 @@
+package main
+
+import (
+	"strings"
+	"testing"
+
+	"sigs.k8s.io/yaml"
+)
+
+// TestSMTPSecretManifestCarriesTargetNamespace pins the fix for the
+// workload-namespace replica: kubectl refuses a manifest whose namespace
+// conflicts with -n ("the namespace from the provided object ... does not
+// match"), so the mirror must render felis-smtp with the TARGET namespace —
+// otherwise the "configure email" refresh fails on the first apply and the
+// felis-config mirror never runs at all.
+func TestSMTPSecretManifestCarriesTargetNamespace(t *testing.T) {
+	for _, ns := range []string{"felis", "minecraft"} {
+		b, err := smtpSecretManifest("pw", ns)
+		if err != nil {
+			t.Fatalf("render for %s: %v", ns, err)
+		}
+		var got struct {
+			Metadata struct {
+				Namespace string `json:"namespace"`
+			} `json:"metadata"`
+		}
+		if err := yaml.Unmarshal(b, &got); err != nil {
+			t.Fatalf("unmarshal for %s: %v", ns, err)
+		}
+		if got.Metadata.Namespace != ns {
+			t.Fatalf("manifest namespace = %q, want %q", got.Metadata.Namespace, ns)
+		}
+		if !strings.Contains(string(b), "name: felis-smtp") {
+			t.Fatalf("manifest must still name felis-smtp: %s", b)
+		}
+	}
+}

cmd/felis/tui_widgets.go

@@ -29,6 +29,33 @@ func tuiSeparator() string {
 	return tuiHint.Render(strings.Repeat("─", 70))
 }
 
+// wrapDisplayURL breaks a long URL into lines no wider than width so the TUI
+// renderer never truncates it on a narrow terminal — the one-time setup URL
+// carries a 43-char token and overruns 80 columns. It prefers breaking right
+// after a '=' or '/' inside the window (the token then lands on its own line)
+// and hard-wraps only when no boundary is available. Lines concatenate back to
+// the original string.
+func wrapDisplayURL(u string, width int) []string {
+	if width <= 0 {
+		width = 70
+	}
+	var lines []string
+	for len(u) > width {
+		cut := width
+		if i := strings.LastIndexByte(u[:width], '='); i >= 0 && i >= width/2 {
+			cut = i + 1
+		} else if i := strings.LastIndexByte(u[:width], '/'); i >= 0 && i >= width/2 {
+			cut = i + 1
+		}
+		lines = append(lines, u[:cut])
+		u = u[cut:]
+	}
+	if u != "" {
+		lines = append(lines, u)
+	}
+	return lines
+}
+
 // tuiStepRail renders a breadcrumb of wizard stages. Steps before `current`
 // render as done, `current` is highlighted, and later steps are dimmed.
 func tuiStepRail(steps []string, current int) string {

cmd/felis/tui_widgets_test.go

@@ -0,0 +1,26 @@
+package main
+
+import (
+	"strings"
+	"testing"
+)
+
+func TestWrapDisplayURL(t *testing.T) {
+	u := "https://op.console.example.net/setup?token=" + strings.Repeat("A", 43)
+	lines := wrapDisplayURL(u, 70)
+	if got := strings.Join(lines, ""); got != u {
+		t.Fatalf("concatenated lines = %q, want the original URL back", got)
+	}
+	for i, l := range lines {
+		if len(l) > 70 {
+			t.Errorf("line %d is %d cols wide: %q", i, len(l), l)
+		}
+	}
+	if len(lines) < 2 || !strings.HasSuffix(lines[0], "token=") {
+		t.Fatalf("want the first line to end at the 'token=' boundary, got %q", lines)
+	}
+	short := "https://a/b"
+	if got := wrapDisplayURL(short, 70); len(got) != 1 || got[0] != short {
+		t.Errorf("short URL should pass through unsplit, got %q", got)
+	}
+}

cmd/felis/update.go

@@ -24,8 +24,9 @@ const updateTimeout = 60 * time.Second
 // The apply side is deliberately NOT implemented in this command. Every component
 // here is installed by deploy/bootstrap.sh, which is idempotent, already handles the
 // parts that are easy to get wrong (Velocity's pinned MINOR, the atomic jar install,
-// the k3s image re-import that a byte-identical StatefulSet template will not
+// the image re-import + registry push that a byte-identical StatefulSet template
-// trigger on its own), and is the path that gets exercised on every install. A
+// will not trigger on its own), and is the path that gets exercised on every
+// install. A
 // second installer living in this file would duplicate that policy, could drift from
 // it silently, and would be reachable only on a live node where a mistake takes the
 // proxy or the control plane down. So `felis update` reports, and hands the operator
@@ -42,6 +43,18 @@ type updateTarget struct {
 	command string
 }
 
+// installerRerun is the tested apply path for every planner-backed selector: re-run the
+// installer. It is idempotent, and it is the only path that fetches a newer version --
+// `felis setup` skips its host-bootstrap phase on a completed install (all four install
+// markers already exist), so there it opens the config console and moves no component,
+// and even on the bootstrap path it re-images felis-api from the binary setup is already
+// running (FELIS_BOOTSTRAP_BINARY), which looks like an update and changes nothing.
+//
+// The URL is the same one-liner both READMEs hand out. While the repo is private it
+// answers 404 (raw.githubusercontent.com hides private repos), which is why the trailer
+// below points at the README's token'd form for that case.
+const installerRerun = "curl -fsSL https://raw.githubusercontent.com/FelisMC/Felis/main/deploy/bootstrap.sh | sudo bash"
+
 // updateTargets is the selector table. panel and plugins both resolve to felis-api
 // because they are not separately versioned: the panel is compiled into the felis
 // binary with //go:embed, and the plugin jars are built from this same repo in the
@@ -51,19 +64,19 @@ var updateTargets = []updateTarget{
 		selector:  "panel",
 		component: "felis-api",
 		note:      "the panel is embedded in the felis binary (//go:embed), so updating it means rebuilding the felis image and rolling felis-api",
-		command:   "sudo felis setup",
+		command:   installerRerun,
 	},
 	{
 		selector:  "velocity",
 		component: "velocity",
 		note:      "re-runs install_velocity: newest BUILD of the pinned minor (FELIS_VELOCITY_VERSION), atomic jar install, then restarts felis-velocity",
-		command:   "sudo felis setup",
+		command:   installerRerun,
 	},
 	{
 		selector:  "plugins",
 		component: "felis-api",
-		note:      "felis-velocity.jar is a host-file swap, but felis-paper.jar and felis-limbo.jar are baked into the lobby/limbo images and need a rebuild + k3s image re-import",
+		note:      "felis-velocity.jar is a host-file swap, but felis-paper.jar and felis-limbo.jar are baked into the lobby/limbo images and need a rebuild + re-mirror into the in-cluster registry (the installer re-run does both)",
-		command:   "sudo felis setup",
+		command:   installerRerun,
 	},
 	{
 		selector:  "mc",
@@ -219,7 +232,6 @@ func renderApplyGuidance(res updater.Result, selected map[string]bool, force boo
 
 	var b strings.Builder
 	var offeredCommand bool
-	var offeredFelisAPI bool
 	for _, t := range updateTargets {
 		if !selected[t.selector] {
 			continue
@@ -248,28 +260,20 @@ func renderApplyGuidance(res updater.Result, selected map[string]bool, force boo
 		}
 		fmt.Fprintf(&b, "  run: %s\n", t.command)
 		offeredCommand = true
-		offeredFelisAPI = offeredFelisAPI || t.component == "felis-api"
 	}
 	// Only explain the command when one was actually offered; a --mc-only run has
 	// nothing to run and the trailer would be a non-sequitur.
-	if offeredCommand {
-		b.WriteString("\nfelis setup is idempotent and re-runs the installer that owns these components;\nit does not reinstall what is already current. Restart game servers afterwards.\n")
-	}
-	// Scoped to felis-api because it is the only component setup cannot move forward.
-	// velocity is fine: install_velocity re-resolves the newest build of the pinned minor
-	// on every run. But setup hands deploy/bootstrap.sh the binary it is itself running
-	// (FELIS_BOOTSTRAP_BINARY), and that arm skips the release lookup entirely, so it
-	// rebuilds the image and rolls the deployment from the SAME binary -- a run that looks
-	// like a successful update and leaves the version unchanged.
 	//
-	// The installer is the only thing that moves felis-api. It is safe to point at now
+	// One trailer serves every selector now: setup is not an apply path at all on a
-	// that detect_node_ip reuses the installed root domain, so what is left to warn about
+	// completed install (shouldRunHostBootstrapBeforeConfig only enters the host
-	// is the channel: FELIS_VERSION_BOOTSTRAP is not persisted anywhere and defaults to
+	// bootstrap while an install marker is missing), so the installer re-run is the one
-	// release, so a bare re-run on a host tracking main quietly moves it onto releases.
+	// worked path for all three components and there is no per-component exception left
-	// That is a channel change, not a broken install, which is why it is one clause and
+	// to scope. Two caveats stay because following the advice without them bites real
-	// not a paragraph.
+	// hosts: the channel is not persisted anywhere (a bare re-run on a main host quietly
-	if offeredFelisAPI {
+	// moves it onto releases), and the private repo's one-liner needs the read token
-		b.WriteString("\nfelis-api (panel, plugins) is the exception: setup re-images it from the felis binary\nalready on this host, so it cannot install a NEWER felis-api. Re-run the bootstrap\ninstaller for that -- it keeps this install's root domain. It does default to the\nrelease channel, so pass FELIS_VERSION_BOOTSTRAP=dev if this host tracks main.\n")
+	// back in the environment before it can resolve anything.
+	if offeredCommand {
+		b.WriteString("\nRe-running the installer applies everything above: it fetches the newest version on\nthe channel in effect and re-applies the bundle (release is the default). The channel\nis not persisted, so pass FELIS_VERSION_BOOTSTRAP=dev if this host tracks main. While\nthis repo is private, the one-liner above 404s without a token; the README's install\nsection has the token'd form that works. felis setup is not this path: on a completed\ninstall it opens the config console and installs nothing newer. Restart game servers\nafterwards.\n")
 	}
 	return b.String()
 }

cmd/felis/update_test.go

@@ -107,7 +107,7 @@ func TestApplyGuidanceMinecraftOffersNoCommand(t *testing.T) {
 	if !strings.Contains(out, "pinned by policy") {
 		t.Fatalf("want the pin explained:\n%s", out)
 	}
-	if strings.Contains(out, "run:") || strings.Contains(out, "felis setup is idempotent") {
+	if strings.Contains(out, "run:") || strings.Contains(out, "Re-running the installer") {
 		t.Fatalf("--mc must offer no command and no command trailer:\n%s", out)
 	}
 }
@@ -133,42 +133,37 @@ func TestUpdateTargetsMatchTopology(t *testing.T) {
 	}
 }
 
-// `sudo felis setup` is the right answer for velocity and the wrong one for felis-api,
+// Re-running the installer is the one apply path this table may hand out. setup is NOT an
-// so the caveat has to be scoped rather than appended to every run. setup hands
+// updater on a completed install -- its host-bootstrap phase only runs while an install
-// bootstrap the binary it is already running, and that arm skips the release lookup:
+// marker is missing, so it opens the config console and moves no component -- and even on
-// the run rebuilds the image and rolls the deployment off the SAME binary, which looks
+// the bootstrap path it re-images felis-api from the binary setup is already running. The
-// like a successful update and changes nothing. install_velocity, by contrast, really
+// table used to answer with "sudo felis setup" and scope a felis-api-only exception; both
-// does re-resolve the newest build on every run.
+// taught a model that does not survive contact with an installed host.
-func TestApplyGuidanceScopesTheFelisAPICaveat(t *testing.T) {
+func TestApplyGuidancePointsEveryComponentAtTheInstaller(t *testing.T) {
-	const caveat = "cannot install a NEWER felis-api"
-
 	api := renderApplyGuidance(
 		planResult([]updates.Action{{Component: "felis-api", Kind: updates.ActionNotify, LatestKnown: true}}),
 		map[string]bool{"panel": true}, false)
-	if !strings.Contains(api, caveat) {
+	for _, want := range []string{"deploy/bootstrap.sh", "FELIS_VERSION_BOOTSTRAP=dev", "felis setup is not this path"} {
-		t.Fatalf("--panel resolves to felis-api and must carry the caveat:\n%s", api)
+		if !strings.Contains(api, want) {
+			t.Fatalf("--panel guidance missing %q:\n%s", want, api)
+		}
 	}
-	// Naming the installer obliges us to name what a bare re-run still changes. The domain
+	if strings.Contains(api, "run: sudo felis setup") {
-	// is handled -- detect_node_ip reuses the installed one -- but the channel is not
+		t.Fatalf("setup must never be offered as the apply command:\n%s", api)
-	// persisted at all and defaults to release, so a host tracking main gets moved onto
-	// releases by following this advice.
-	if !strings.Contains(api, "FELIS_VERSION_BOOTSTRAP=dev") {
-		t.Fatalf("pointing at the installer without the channel caveat misleads a dev host:\n%s", api)
 	}
 
+	// The same path serves velocity; a scoped caveat would re-teach the old model that
+	// setup fixes velocity.
 	vel := renderApplyGuidance(
 		planResult([]updates.Action{{Component: "velocity", Kind: updates.ActionNotify, LatestKnown: true}}),
 		map[string]bool{"velocity": true}, false)
-	if strings.Contains(vel, caveat) {
+	if !strings.Contains(vel, "run: curl -fsSL") || !strings.Contains(vel, "felis setup is not this path") {
-		t.Fatalf("velocity IS fixed by setup; the caveat would misdirect the operator:\n%s", vel)
+		t.Fatalf("velocity gets the same installer path:\n%s", vel)
-	}
-	if !strings.Contains(vel, "felis setup is idempotent") {
-		t.Fatalf("velocity still wants the ordinary trailer:\n%s", vel)
 	}
 
 	// --mc offers no command at all, so neither trailer belongs.
 	mc := renderApplyGuidance(planResult(nil), map[string]bool{"mc": true}, true)
-	if strings.Contains(mc, caveat) {
+	if strings.Contains(mc, "deploy/bootstrap.sh") || strings.Contains(mc, "FELIS_VERSION_BOOTSTRAP") {
-		t.Fatalf("--mc offers no command; the caveat is a non-sequitur:\n%s", mc)
+		t.Fatalf("--mc offers no command; the trailer is a non-sequitur:\n%s", mc)
 	}
 }

deploy/alerts/felis-alerts.yaml

@@ -0,0 +1,69 @@
+# Felis alert rules — plain Prometheus format (also the promtool-tested source
+# for felis-prometheusrule.yaml). See docs/troubleshooting.md §14 for scraping
+# and loading instructions.
+#
+# felis_* series come from two processes:
+#   - felis-operator pod :8080/metrics      → felis_servers_total, felis_start_duration_seconds
+#   - felis-api internal :8081/metrics      → felis_image_build_failures_total
+# node_* / kube_* series come from node-exporter / kube-state-metrics.
+groups:
+  - name: felis.rules
+    rules:
+      - alert: FelisImageBuildFailures
+        expr: increase(felis_image_build_failures_total[6h]) > 0
+        for: 5m
+        labels:
+          severity: warning
+        annotations:
+          summary: "modpack/image build failed in the last 6h"
+          description: >-
+            felis_image_build_failures_total increased. Inspect the failed build Job
+            (kubectl logs -n felis-build job/<build-job>); the same error text is on
+            GET /api/v1/images/build/{id} and in the submitter's row in the panel.
+      - alert: FelisSlowServerStarts
+        expr: histogram_quantile(0.9, sum by (le) (rate(felis_start_duration_seconds_bucket[30m]))) > 300
+        for: 15m
+        labels:
+          severity: warning
+        annotations:
+          summary: "p90 server start time exceeds 5 minutes"
+          description: >-
+            Starts regularly take over five minutes (felis_start_duration_seconds,
+            observed when readiness is first reached). A start that never completes
+            records nothing — cross-check desiredState=Running servers with no ready
+            phase (troubleshooting §1).
+  - name: felis.node.rules
+    rules:
+      - alert: FelisNodeDiskSpaceLow
+        expr: >-
+          node_filesystem_avail_bytes{fstype=~"ext4|xfs|btrfs"}
+          / node_filesystem_size_bytes{fstype=~"ext4|xfs|btrfs"} < 0.15
+        for: 15m
+        labels:
+          severity: warning
+        annotations:
+          summary: "node filesystem {{ $labels.mountpoint }} below 15% available"
+          description: >-
+            Sustained disk pressure evicts game pods and garbage-collects images
+            (troubleshooting §13b). Free space before kubelet raises DiskPressure.
+      - alert: FelisNodeDiskPressure
+        expr: kube_node_status_condition{condition="DiskPressure",status="true"} == 1
+        for: 5m
+        labels:
+          severity: critical
+        annotations:
+          summary: "kubelet reports DiskPressure on {{ $labels.node }}"
+          description: >-
+            The eviction chain is in progress: control-plane pods hold
+            system-cluster-critical and survive, game pods do not. Free disk now
+            (troubleshooting §13b).
+      - alert: FelisNodeMemoryLow
+        expr: node_memory_MemAvailable_bytes / node_memory_MemTotal_bytes < 0.10
+        for: 15m
+        labels:
+          severity: warning
+        annotations:
+          summary: "node memory available below 10% for 15m"
+          description: >-
+            PostgreSQL, the control plane, the registry and game servers share one
+            node; sustained memory pressure risks OOM kills.

deploy/alerts/felis-alerts_test.yml

@@ -0,0 +1,107 @@
+# promtool unit tests: `promtool test rules felis-alerts_test.yml`
+# Proves every shipped rule actually fires on its target condition (and stays
+# silent before it).
+rule_files:
+  - felis-alerts.yaml
+evaluation_interval: 1m
+tests:
+  - name: build failure and slow starts
+    interval: 1m
+    input_series:
+      # counter: quiet for 5m, then one failure per step.
+      - series: 'felis_image_build_failures_total'
+        values: '0x5 1x15'
+      # histogram: all observations land in the (300,600] bucket.
+      - series: 'felis_start_duration_seconds_bucket{le="120"}'
+        values: '0x22'
+      - series: 'felis_start_duration_seconds_bucket{le="300"}'
+        values: '0x22'
+      - series: 'felis_start_duration_seconds_bucket{le="600"}'
+        values: '0+10x21'
+      - series: 'felis_start_duration_seconds_bucket{le="+Inf"}'
+        values: '0+10x21'
+    alert_rule_test:
+      - eval_time: 2m
+        alertname: FelisImageBuildFailures
+        exp_alerts: []
+      - eval_time: 20m
+        alertname: FelisImageBuildFailures
+        exp_alerts:
+          - exp_labels:
+              severity: warning
+            exp_annotations:
+              summary: "modpack/image build failed in the last 6h"
+              description: >-
+                felis_image_build_failures_total increased. Inspect the failed build Job
+                (kubectl logs -n felis-build job/<build-job>); the same error text is on
+                GET /api/v1/images/build/{id} and in the submitter's row in the panel.
+      - eval_time: 20m
+        alertname: FelisSlowServerStarts
+        exp_alerts:
+          - exp_labels:
+              severity: warning
+            exp_annotations:
+              summary: "p90 server start time exceeds 5 minutes"
+              description: >-
+                Starts regularly take over five minutes (felis_start_duration_seconds,
+                observed when readiness is first reached). A start that never completes
+                records nothing — cross-check desiredState=Running servers with no ready
+                phase (troubleshooting §1).
+  - name: node disk and memory thresholds
+    interval: 1m
+    input_series:
+      - series: 'node_filesystem_avail_bytes{device="/dev/vda1",fstype="xfs",instance="node1",job="node-exporter",mountpoint="/"}'
+        values: '10x26'
+      - series: 'node_filesystem_size_bytes{device="/dev/vda1",fstype="xfs",instance="node1",job="node-exporter",mountpoint="/"}'
+        values: '100x26'
+      - series: 'kube_node_status_condition{condition="DiskPressure",node="n1",status="true"}'
+        values: '0x4 1x22'
+      - series: 'node_memory_MemAvailable_bytes{instance="node1",job="node-exporter"}'
+        values: '5x26'
+      - series: 'node_memory_MemTotal_bytes{instance="node1",job="node-exporter"}'
+        values: '100x26'
+    alert_rule_test:
+      - eval_time: 2m
+        alertname: FelisNodeDiskPressure
+        exp_alerts: []
+      - eval_time: 20m
+        alertname: FelisNodeDiskSpaceLow
+        exp_alerts:
+          - exp_labels:
+              device: /dev/vda1
+              fstype: xfs
+              instance: node1
+              job: node-exporter
+              mountpoint: /
+              severity: warning
+            exp_annotations:
+              summary: "node filesystem / below 15% available"
+              description: >-
+                Sustained disk pressure evicts game pods and garbage-collects images
+                (troubleshooting §13b). Free space before kubelet raises DiskPressure.
+      - eval_time: 20m
+        alertname: FelisNodeDiskPressure
+        exp_alerts:
+          - exp_labels:
+              condition: DiskPressure
+              node: n1
+              status: "true"
+              severity: critical
+            exp_annotations:
+              summary: "kubelet reports DiskPressure on n1"
+              description: >-
+                The eviction chain is in progress: control-plane pods hold
+                system-cluster-critical and survive, game pods do not. Free disk now
+                (troubleshooting §13b).
+      - eval_time: 20m
+        alertname: FelisNodeMemoryLow
+        exp_alerts:
+          - exp_labels:
+              instance: node1
+              job: node-exporter
+              severity: warning
+            exp_annotations:
+              summary: "node memory available below 10% for 15m"
+              description: >-
+                PostgreSQL, the control plane, the registry and game servers share one
+                node; sustained memory pressure risks OOM kills.

deploy/alerts/felis-prometheusrule.yaml

@@ -0,0 +1,74 @@
+# prometheus-operator twin of felis-alerts.yaml (kube-prometheus-stack loads
+# rules through the PrometheusRule CRD, not rule_files). The plain file is the
+# promtool-tested source; keep the groups in sync.
+apiVersion: monitoring.coreos.com/v1
+kind: PrometheusRule
+metadata:
+  name: felis-alerts
+  namespace: monitoring
+  labels:
+    # Change to match your stack's ruleSelector (kube-prometheus-stack's
+    # default selects on the Helm release name).
+    release: kube-prometheus-stack
+spec:
+  groups:
+    - name: felis.rules
+      rules:
+        - alert: FelisImageBuildFailures
+          expr: increase(felis_image_build_failures_total[6h]) > 0
+          for: 5m
+          labels:
+            severity: warning
+          annotations:
+            summary: "modpack/image build failed in the last 6h"
+            description: >-
+              felis_image_build_failures_total increased. Inspect the failed build Job
+              (kubectl logs -n felis-build job/<build-job>); the same error text is on
+              GET /api/v1/images/build/{id} and in the submitter's row in the panel.
+        - alert: FelisSlowServerStarts
+          expr: histogram_quantile(0.9, sum by (le) (rate(felis_start_duration_seconds_bucket[30m]))) > 300
+          for: 15m
+          labels:
+            severity: warning
+          annotations:
+            summary: "p90 server start time exceeds 5 minutes"
+            description: >-
+              Starts regularly take over five minutes (felis_start_duration_seconds,
+              observed when readiness is first reached). A start that never completes
+              records nothing — cross-check desiredState=Running servers with no ready
+              phase (troubleshooting §1).
+    - name: felis.node.rules
+      rules:
+        - alert: FelisNodeDiskSpaceLow
+          expr: >-
+            node_filesystem_avail_bytes{fstype=~"ext4|xfs|btrfs"}
+            / node_filesystem_size_bytes{fstype=~"ext4|xfs|btrfs"} < 0.15
+          for: 15m
+          labels:
+            severity: warning
+          annotations:
+            summary: "node filesystem {{ $labels.mountpoint }} below 15% available"
+            description: >-
+              Sustained disk pressure evicts game pods and garbage-collects images
+              (troubleshooting §13b). Free space before kubelet raises DiskPressure.
+        - alert: FelisNodeDiskPressure
+          expr: kube_node_status_condition{condition="DiskPressure",status="true"} == 1
+          for: 5m
+          labels:
+            severity: critical
+          annotations:
+            summary: "kubelet reports DiskPressure on {{ $labels.node }}"
+            description: >-
+              The eviction chain is in progress: control-plane pods hold
+              system-cluster-critical and survive, game pods do not. Free disk now
+              (troubleshooting §13b).
+        - alert: FelisNodeMemoryLow
+          expr: node_memory_MemAvailable_bytes / node_memory_MemTotal_bytes < 0.10
+          for: 15m
+          labels:
+            severity: warning
+          annotations:
+            summary: "node memory available below 10% for 15m"
+            description: >-
+              PostgreSQL, the control plane, the registry and game servers share one
+              node; sustained memory pressure risks OOM kills.

deploy/bootstrap.sh

@@ -59,10 +59,24 @@
 #   FELIS_GITHUB_TOKEN GitHub token; REQUIRED while the repo is private
 #   FELIS_REF         branch/tag/sha — pins the build, overrides the channel, and forces a
 #                     source build (naming a ref asks for that tree, not a published asset)
-#   FELIS_IMAGE       local image tag         (default: felis:demo  — never :latest)
+#   FELIS_IMAGE       control-plane image ref (default:
+#                     registry.felis.svc:5000/felis/felis:demo — never :latest;
+#                     anything not under the registry is used as-is but is NOT
+#                     mirrored into it, so it has no pull source after an image GC)
 #   FELIS_ROOT_DOMAIN deployment root domain  (default: <node-ip>.nip.io)
 #   FELIS_PANEL_NODEPORT local HTTPS panel/API NodePort (default: 30443)
 #   FELIS_EGRESS_MODE loadbalancer|nodeport   (default: nodeport — no MetalLB on a demo box)
+#   FELIS_BACKUP_PVC  world-archive PVC the installer renders and felis-api hands to its
+#                     backup/restore Jobs (default: felis-backups; empty string disables
+#                     backups — the endpoints answer 503)
+#   FELIS_ARCHIVE_LOCAL_PATH path that PVC is mounted at inside those Jobs; written into
+#                     felis.toml [archive] local_path (default: /var/lib/felis/archives)
+#   FELIS_WORLDS_HOST_PATH node directory holding the world volumes (on the k3s this
+#                     installer provisions: /var/lib/rancher/k3s/storage). Setting it
+#                     enables the daily retention reaper, which archives and then deletes
+#                     worlds idle beyond the retention window, and grants the reaper's
+#                     uid (1000) traverse access to that root — k3s ships it 0700
+#                     root:root (default: unset = no reaper)
 #   PKG_LOCK_TIMEOUT seconds to wait for package-manager locks (default: 900)
 #   APT_LOCK_TIMEOUT legacy alias for PKG_LOCK_TIMEOUT
 set -Eeuo pipefail
@@ -70,7 +84,7 @@ set -Eeuo pipefail
 # ---------------------------------------------------------------------------
 # Configuration & constants
 # ---------------------------------------------------------------------------
-FELIS_REPO_URL="${FELIS_REPO_URL:-https://github.com/MliroLirrorsIngenuity/Felis.git}"
+FELIS_REPO_URL="${FELIS_REPO_URL:-https://github.com/FelisMC/Felis.git}"
 # Which version to install. "release" builds the newest published GitHub release;
 # "dev" builds the tip of main. Release is the default because an installer that
 # tracks a moving branch by default hands every new host a different, untested
@@ -103,9 +117,29 @@ export FELIS_GITHUB_TOKEN
 # Set by resolve_install_ref/stamp_version and linked into the binary as main.version.
 FELIS_VERSION=""
 FELIS_VERSION_BASE=""
-FELIS_IMAGE="${FELIS_IMAGE:-felis:demo}"
+# Where the installer parks every image it builds, so kubelet can re-pull one the
+# image GC has collected (the disk-pressure drill's dead end: ImagePullBackOff with
+# nothing to pull from). The node pulls through the loopback hostPort the registry
+# Deployment binds (configure_registry_mirror below); pushes go through
+# REGISTRY_PUSH_HOST — docker treats 127.0.0.1 as insecure by default, so the
+# daemon needs no insecure-registries entry for it.
+REGISTRY_URL="registry.felis.svc:5000"
+REGISTRY_PUSH_HOST="127.0.0.1:${REGISTRY_URL##*:}"
+FELIS_IMAGE="${FELIS_IMAGE:-${REGISTRY_URL}/felis/felis:demo}"
 FELIS_EGRESS_MODE="${FELIS_EGRESS_MODE:-nodeport}"
 FELIS_PANEL_NODEPORT="${FELIS_PANEL_NODEPORT:-30443}"
+# World-archive storage. The installer renders this PVC (minecraft namespace) and felis-api
+# advertises it to its backup/restore Jobs; emptying it disables backups (503). The archive
+# path is written into felis.toml so the Jobs' mount and [archive] local_path agree by
+# construction — a mismatch would leave tarLocal's absolute archive refs unresolvable.
+FELIS_BACKUP_PVC="${FELIS_BACKUP_PVC:-felis-backups}"
+FELIS_ARCHIVE_LOCAL_PATH="${FELIS_ARCHIVE_LOCAL_PATH:-/var/lib/felis/archives}"
+# Retention is opt-in because it DELETES worlds (after a verified archive): point this at the
+# node directory the world volumes live under. On the k3s this installer provisions that is
+# /var/lib/rancher/k3s/storage — the reaper resolves each PVC's local-path directory exactly
+# from its volumeName. Left unset, no reaper CronJob renders and archives accumulate until
+# the backup PVC fills (then backups fail loudly; nothing is deleted).
+FELIS_WORLDS_HOST_PATH="${FELIS_WORLDS_HOST_PATH:-}"
 INSTALL_MODE="${FELIS_INSTALL_MODE:-}"
 # Loopback by default: hasJoined is an unauthenticated endpoint by protocol (Velocity
 # sends no token), so a public bind is a free auth relay — anyone can point their own
@@ -133,12 +167,12 @@ PKG_LOCK_TIMEOUT="${PKG_LOCK_TIMEOUT:-${APT_LOCK_TIMEOUT:-900}}"
 APT_LOCK_TIMEOUT="${APT_LOCK_TIMEOUT:-$PKG_LOCK_TIMEOUT}"
 
 # --- the game stack: proxy on the host, the two always-on backends in k3s ---
-FELIS_LIMBO_IMAGE="${FELIS_LIMBO_IMAGE:-felis-limbo:demo}"
+FELIS_LIMBO_IMAGE="${FELIS_LIMBO_IMAGE:-${REGISTRY_URL}/felis/limbo:demo}"
-FELIS_LOBBY_IMAGE="${FELIS_LOBBY_IMAGE:-felis-lobby:demo}"
+FELIS_LOBBY_IMAGE="${FELIS_LOBBY_IMAGE:-${REGISTRY_URL}/felis/lobby:demo}"
 # Plain Paper base recommended for a user's own server (deploy/paper). Not a system
 # server — forwarding is applied by the operator's init-forwarding initContainer, so it
 # needs no secret. Seeded recommended in 0019_recommended_paper.sql.
-FELIS_PAPER_IMAGE="${FELIS_PAPER_IMAGE:-felis-paper:demo}"
+FELIS_PAPER_IMAGE="${FELIS_PAPER_IMAGE:-${REGISTRY_URL}/felis/paper:demo}"
 # The Velocity MINOR is pinned, not discovered. PaperMC's Fill v3 groups velocity
 # builds by version group, and "newest across all groups" today means 4.0.0-SNAPSHOT —
 # an UNRELEASED proxy (the 4.0.0 group has zero published builds) that needs a Java 25
@@ -191,7 +225,6 @@ POD_CIDR="10.42.0.0/16"          # k3s default cluster CIDR
 SERVICE_CIDR="10.43.0.0/16"      # k3s default service CIDR
 DB_NAME="felis"
 DB_USER="felis"
-REGISTRY_URL="registry.felis.svc:5000"
 
 STATE_DIR="/etc/felis"
 SECRETS_ENV="${STATE_DIR}/secrets.env"
@@ -210,6 +243,10 @@ VELOCITY_SERVICE="/etc/systemd/system/felis-velocity.service"
 JRE_DIR="/opt/felis/jre"
 K3S_BIN_DIR="${K3S_BIN_DIR:-/usr/local/bin}"
 K3S_BIN="${K3S_BIN_DIR}/k3s"
+# k3s's containerd mirror config, written by configure_registry_mirror. A variable
+# (not just the literal path) so bootstrap_test.sh can point the writer at a
+# scratch file.
+K3S_REGISTRIES_FILE="/etc/rancher/k3s/registries.yaml"
 APT_LOCK_FILES=(
   /var/lib/dpkg/lock-frontend
   /var/lib/dpkg/lock
@@ -307,6 +344,21 @@ apply_literal_secret() {
   rm -f "$tmp"
 }
 
+# The control plane mounts felis-config from its own namespace; the workload
+# namespace's backup/restore/fileedit Jobs and the reaper mount a local copy (a
+# secretKeyRef is namespace-local). The installer owns the rendered config, so both
+# copies are (re)applied on every run — unlike the create-if-absent credential
+# replicas `felis setup` makes, because a stale config copy keeps an old database URL
+# or archive policy after an upgrade or a credential rotation.
+apply_felis_config_secrets() {
+  kube -n "$CONTROL_NS" create secret generic felis-config \
+    --from-file=felis.toml="${STATE_DIR}/felis.pod.toml" \
+    --dry-run=client -o yaml | kube apply -f -
+  kube -n "$MINECRAFT_NS" create secret generic felis-config \
+    --from-file=felis.toml="${STATE_DIR}/felis.pod.toml" \
+    --dry-run=client -o yaml | kube apply -f -
+}
+
 as_postgres() {
   if command -v runuser >/dev/null 2>&1; then
     runuser -u postgres -- "$@"
@@ -790,6 +842,13 @@ install_k3s() {
   systemctl enable --now k3s
   export KUBECONFIG=/etc/rancher/k3s/k3s.yaml
   log "waiting for the node to become Ready"
+  wait_for_node_ready
+}
+
+# Waits for the (single) node to report Ready. Shared by the k3s install and the
+# registry-mirror restart below: both restart the agent, and a bootstrap that
+# proceeds early fails later with a misleading "not found"/timeout instead.
+wait_for_node_ready() {
   local i
   for i in $(seq 1 60); do
     if kube get nodes 2>/dev/null | grep -q ' Ready '; then
@@ -802,6 +861,78 @@ install_k3s() {
   die "k3s node did not become Ready in time"
 }
 
+# ---------------------------------------------------------------------------
+# 4b. The in-cluster registry: the node-side pull path, the registry's own
+#     image, and the hosting of every image this installer builds.
+#
+#     Kubelet's image GC collects an unused image under disk pressure (drilled:
+#     the game images were collected and ImagePullBackOff had nothing to pull
+#     from). The fix is a pull source that is always there — the registry the
+#     bundle already renders. Two node-level facts make that work:
+#       * kubelet cannot reach the registry Service VIP (the live stack answered
+#         "Empty reply"), so containerd is told to go through the loopback
+#         hostPort the registry Deployment binds (the Deployment renders it) —
+#         that is configure_registry_mirror below;
+#       * the registry's own image (registry:2) must already be in containerd
+#         before the registry Deployment can start at all —
+#         import_registry_image below caches it.
+#     After deploy_bundle, push_images_to_registry mirrors the built images into
+#     the registry, so containerd's imported copies are a first-boot cache
+#     rather than the only copy.
+# ---------------------------------------------------------------------------
+
+# The node's containerd cannot dial the registry Service VIP, so pulls arrive
+# over the loopback hostPort the registry Deployment binds. k3s reads this file
+# when the agent starts and regenerates containerd's certs.d from it — no
+# restart, no effect — so a CONTENT change restarts k3s; an identical file
+# (every re-run) restarts nothing. K3S_REGISTRIES_FILE is a variable so
+# bootstrap_test.sh can point the function at a scratch file.
+configure_registry_mirror() {
+  local file="$K3S_REGISTRIES_FILE" tmp
+  tmp="$(mktemp)"
+  remember_temp "$tmp"
+  cat > "$tmp" <<EOF
+mirrors:
+  "${REGISTRY_URL}":
+    endpoint:
+      - "http://${REGISTRY_PUSH_HOST}"
+EOF
+  if [ -f "$file" ] && cmp -s "$tmp" "$file"; then
+    rm -f "$tmp"
+    ok "registry mirror already configured (${REGISTRY_URL} -> http://${REGISTRY_PUSH_HOST})"
+    return 0
+  fi
+  mkdir -p "$(dirname "$file")"
+  mv "$tmp" "$file"
+  log "restarting k3s to load the registry mirror (${REGISTRY_URL} -> http://${REGISTRY_PUSH_HOST})"
+  systemctl restart k3s
+  export KUBECONFIG=/etc/rancher/k3s/k3s.yaml
+  wait_for_node_ready
+}
+
+# The registry Deployment runs registry:2 (platform.defaultRegistryImage; the
+# renderer's default — this script never passes --registry-image). On a box
+# that cannot reach Docker Hub the Deployment can never start without a local
+# copy, so the installer caches one whenever it can. Best-effort by design: if
+# the pull fails the registry rollout still fails loudly at deploy_bundle, with
+# the regular diagnostics — but for every box that CAN pull, the image is
+# fetched exactly once, here, instead of at first pod start.
+import_registry_image() {
+  # The name containerd normalizes "registry:2" to after any docker-save import.
+  if k3s_cmd ctr images ls -q 2>/dev/null | grep -qx 'docker.io/library/registry:2'; then
+    ok "registry image registry:2 already in k3s containerd"
+    return 0
+  fi
+  log "importing the registry's own image (registry:2) into k3s containerd"
+  systemctl start docker
+  if docker pull registry:2 && docker save registry:2 | k3s_cmd ctr images import -; then
+    ok "registry image registry:2 imported"
+  else
+    warn "could not import registry:2: the in-cluster registry will start only if the node can pull it from Docker Hub; on an air-gapped box import it by hand (docs/troubleshooting.md §8e)"
+  fi
+  systemctl stop docker docker.socket 2>/dev/null || true
+}
+
 # ---------------------------------------------------------------------------
 # 5. Source/binary + image build + containerd import
 # ---------------------------------------------------------------------------
@@ -1299,7 +1430,7 @@ resolve_game_jars() {
 luckperms_latest_jar() {
   local json url
   json="$(curl -fsSL --retry 5 --retry-delay 2 \
-    -A "felis-bootstrap (+https://github.com/MliroLirrorsIngenuity/Felis)" \
+    -A "felis-bootstrap (+https://github.com/FelisMC/Felis)" \
     "https://metadata.luckperms.net/data/all")" || return 1
   url="$(printf '%s' "$json" \
     | grep -o 'https://download\.luckperms\.net/[0-9]\{1,\}/bukkit/loader/[^"]*\.jar' || true)"
@@ -1321,7 +1452,7 @@ luckperms_latest_jar() {
 papermc_latest_jar() {
   local project="$1" version="$2" json urls url sha
   json="$(curl -fsSL --retry 5 --retry-delay 2 \
-    -A "felis-bootstrap (+https://github.com/MliroLirrorsIngenuity/Felis)" \
+    -A "felis-bootstrap (+https://github.com/FelisMC/Felis)" \
     "https://fill.papermc.io/v3/projects/${project}/versions/${version}/builds/latest")" || return 1
   urls="$(printf '%s' "$json" | grep -o 'https://fill-data\.papermc\.io/[^"]*\.jar' || true)"
   url="${urls%%$'\n'*}"
@@ -1971,24 +2102,30 @@ EOF
 # family as the root_domain loss fixed in ecbeb20 -- generated file, hand-set
 # value, no carry-forward.
 #
-# Cached on first call because write_felis_toml clobbers felis.host.toml before
+# The carry is the section's header and key lines only. Printing every line up to
-# it is called again for felis.pod.toml: by then the file this would read from
+# the next section header hoarded the generated [[auth_source]] comment block
-# no longer has the block. The pod toml is the fallback for exactly that window.
+# that sits below [smtp] into this carry: each re-run then re-emitted the hoard
+# plus a fresh template copy, growing both config files by one comment block per
+# run (audit #50). The extraction is idempotent, which is also why re-reading the
+# freshly rewritten host file on the pod pass is safe. The pod toml remains the
+# fallback for a host file with no [smtp] section at all.
 persisted_smtp_block() {
-  if [ -z "${SMTP_BLOCK_CACHED:-}" ]; then
+  local f out
-    SMTP_BLOCK_CACHED=1
+  for f in "${STATE_DIR}/felis.host.toml" "${STATE_DIR}/felis.pod.toml"; do
-    SMTP_BLOCK=""
+    [ -r "$f" ] || continue
-    local f
+    out="$(awk '
-    for f in "${STATE_DIR}/felis.host.toml" "${STATE_DIR}/felis.pod.toml"; do
+      /^[[:space:]]*\[/ {
-      [ -r "$f" ] || continue
+        if (insmtp) exit
-      # Print from [smtp] up to (not including) the next section header.
+        insmtp = ($0 ~ /^[[:space:]]*\[smtp\][[:space:]]*$/)
-      SMTP_BLOCK="$(awk '/^[[:space:]]*\[smtp\]/ { f=1 }
+        if (insmtp) print
-                         f && /^[[:space:]]*\[/ && !/^[[:space:]]*\[smtp\]/ { exit }
+        next
-                         f { print }' "$f")"
+      }
-      [ -n "$SMTP_BLOCK" ] && break
+      insmtp && /^[[:space:]]*("[A-Za-z_][A-Za-z0-9_]*"|[A-Za-z_][A-Za-z0-9_]*)[[:space:]]*=/ { print }
-    done
+    ' "$f")"
-  fi
+    [ -n "$out" ] || continue
-  printf '%s' "$SMTP_BLOCK"
+    printf '%s' "$out"
+    return 0
+  done
 }
 
 # persisted_auth_source_blocks echoes the [[auth_source]] tables an earlier run left
@@ -2013,17 +2150,78 @@ persisted_auth_source_blocks() {
     'url = "https://littleskin.cn/api/yggdrasil/sessionserver/session/minecraft/hasJoined"'
 }
 
+# persisted_archive_block echoes the operator-owned [archive] keys an earlier run
+# left behind — the retention window, the pre-reap warn offsets, the local cap —
+# so a re-run does not silently revert them to the built-ins the reaper carries
+# (felis reaper reads these from the config Secret at run time; defaults: 90d
+# retention, 3d/1d warnings, no cap). store and local_path are NOT carried: this
+# script owns them (FELIS_ARCHIVE_LOCAL_PATH must equal the mount). Same
+# first-readable-file rule as persisted_smtp_block; warn_before must be a
+# single-line TOML array (the shape every writer here emits).
+persisted_archive_block() {
+  local f out
+  for f in "${STATE_DIR}/felis.host.toml" "${STATE_DIR}/felis.pod.toml"; do
+    [ -r "$f" ] || continue
+    out="$(awk '
+      /^[[:space:]]*\[/ { sect = $0; next }
+      sect ~ /^[[:space:]]*\[archive\][[:space:]]*$/ &&
+        /^[[:space:]]*(retention|warn_before|max_local_bytes)[[:space:]]*=/ { print }
+    ' "$f")"
+    [ -n "$out" ] || continue
+    printf '%s\n' "$out"
+    return 0
+  done
+}
+
+# persisted_registry_block echoes the operator-owned [registry] keys an earlier run
+# left behind — the build-lane executor mirrors, the resource caps, the uploads
+# backend and its [registry.s3] subtable — so §15's upgrade path (re-run the
+# installer) does not silently revert them. Nothing in this script's inputs
+# derives these: they are hand-written per docs/troubleshooting.md §8e or stamped
+# by the storage wizard. url and build_namespace are NOT carried: this script
+# owns them (they must match REGISTRY_URL / BUILD_NS). Same first-readable-file
+# rule as persisted_smtp_block.
+persisted_registry_block() {
+  local f out
+  for f in "${STATE_DIR}/felis.host.toml" "${STATE_DIR}/felis.pod.toml"; do
+    [ -r "$f" ] || continue
+    out="$(awk '
+      /^[[:space:]]*\[/ { sect = $0; next }
+      sect ~ /^[[:space:]]*\[registry\][[:space:]]*$/ &&
+        /^[[:space:]]*(kaniko_image|trivy_image|trivy_db_repository|trivy_java_db_repository|build_cpu_limit|build_mem_limit|user_uploads_context)[[:space:]]*=/ { print }
+      sect ~ /^[[:space:]]*\[registry\.s3\][[:space:]]*$/ && /^[[:space:]]*[A-Za-z_]+[[:space:]]*=/ {
+        if (!s3hdr) { printf "[registry.s3]\n"; s3hdr = 1 }
+        print
+      }
+    ' "$f")"
+    [ -n "$out" ] || continue
+    printf '%s\n' "$out"
+    return 0
+  done
+}
+
 write_felis_toml() {
-  local target="$1" db_host="$2" smtp_block auth_source_blocks
+  local target="$1" db_host="$2" smtp_block auth_source_blocks registry_block archive_block
   smtp_block="$(persisted_smtp_block)"
   if [ -n "$smtp_block" ]; then
     log "carrying forward the configured [smtp] relay"
     smtp_block="${smtp_block}"$'\n' # keep a blank line before the next section
   fi
   auth_source_blocks="$(persisted_auth_source_blocks)"
+  registry_block="$(persisted_registry_block)"
+  if [ -n "$registry_block" ]; then
+    log "carrying forward the configured [registry] overrides"
+    registry_block="${registry_block}"$'\n' # keep a blank line before the next section
+  fi
+  archive_block="$(persisted_archive_block)"
+  if [ -n "$archive_block" ]; then
+    log "carrying forward the configured [archive] overrides"
+    archive_block="${archive_block}"$'\n' # keep a blank line before the next section
+  fi
   cat > "$target" <<EOF
 # Generated by deploy/bootstrap.sh; rerun the installer to regenerate. Hand edits are
-# overwritten, except [smtp] and [[auth_source]], which carry forward.
+# overwritten, except [smtp], [[auth_source]], and the operator-owned [registry] /
+# [archive] overrides, which carry forward.
 [server]
 listen = "0.0.0.0:8080"
 root_domain = "${FELIS_ROOT_DOMAIN}"
@@ -2044,11 +2242,11 @@ lobby_image = "${FELIS_LOBBY_IMAGE}"
 [registry]
 url = "${REGISTRY_URL}"
 build_namespace = "${BUILD_NS}"
-
+${registry_block}
 [archive]
 store = "tarLocal"
-local_path = "/var/lib/felis/archives"
+local_path = "${FELIS_ARCHIVE_LOCAL_PATH}"
-
+${archive_block}
 [auth]
 admin_hostname = "op.console.${FELIS_ROOT_DOMAIN}"
 panel_hostname = "console.${FELIS_ROOT_DOMAIN}"
@@ -2121,10 +2319,12 @@ deploy_bundle() {
   done
 
   log "provisioning felis-config + felis-service-token + felis-forwarding-secret + panel TLS secrets (out-of-band, never in the bundle)"
-  kube -n "$CONTROL_NS" create secret generic felis-config \
+  apply_felis_config_secrets
-    --from-file=felis.toml="${STATE_DIR}/felis.pod.toml" \
-    --dry-run=client -o yaml | kube apply -f -
   apply_literal_secret "$CONTROL_NS" felis-service-token token "$SERVICE_TOKEN"
+  # The build namespace needs the same token: the build Job's fetch initContainer
+  # streams a submission's build context from the felis-api internal face, and a
+  # secretKeyRef is namespace-local (a PVC cannot carry it across either).
+  apply_literal_secret "$BUILD_NS" felis-service-token token "$SERVICE_TOKEN"
   # The forwarding key every backend verifies the proxy's handshake with. `felis setup`
   # replicates it into the minecraft namespace (ensureSecretReplica) before it creates
   # the pods that mount it; the operator injects it into EVERY backend, because Velocity's
@@ -2137,11 +2337,40 @@ deploy_bundle() {
     --dry-run=client -o yaml | kube apply -f -
 
   log "rendering + applying the control-plane bundle"
-  "$HOST_BIN" manifests \
+  local -a manifest_args=(
-    --felis-image "$FELIS_IMAGE" \
+    --felis-image "$FELIS_IMAGE"
-    --panel-node-port "$FELIS_PANEL_NODEPORT" \
+    --panel-node-port "$FELIS_PANEL_NODEPORT"
-    --velocity-cidr "${NODE_IP}/32" \
+    --velocity-cidr "${NODE_IP}/32"
-    | kube apply -f -
+  )
+  # Backups are on by default (the renderer's own default names felis-backups); an emptied
+  # FELIS_BACKUP_PVC asks for the no-backup shape explicitly, and a custom name must be
+  # passed through or the api would advertise a PVC the bundle never created.
+  if [ -n "$FELIS_BACKUP_PVC" ]; then
+    manifest_args+=(--backup-pvc "$FELIS_BACKUP_PVC")
+  else
+    manifest_args+=(--backup-pvc=)
+  fi
+  # Retention renders only when the operator names where the worlds live; the archive path
+  # always travels with it because it must equal the [archive] local_path written above.
+  if [ -n "$FELIS_WORLDS_HOST_PATH" ]; then
+    log "retention enabled: the daily reaper will read worlds from ${FELIS_WORLDS_HOST_PATH}"
+    # The reaper pod runs as the tree's non-root uid (1000, platform.workloads.nonRootUID)
+    # and must traverse into the per-volume directories under this root. k3s's own storage
+    # root ships 0700 root:root, so grant traverse — an ACL entry when the host has setfacl,
+    # otherwise the equivalent o+x. Traverse only: no listing either way, and the per-volume
+    # directories themselves are world-accessible (local-path creates them 0777).
+    if [ -d "$FELIS_WORLDS_HOST_PATH" ]; then
+      if command -v setfacl >/dev/null 2>&1; then
+        setfacl -m u:1000:x "$FELIS_WORLDS_HOST_PATH" || chmod o+x "$FELIS_WORLDS_HOST_PATH"
+      else
+        chmod o+x "$FELIS_WORLDS_HOST_PATH"
+      fi
+    else
+      warn "worlds root ${FELIS_WORLDS_HOST_PATH} does not exist yet; the reaper CronJob cannot start until it does (hostPath type Directory)"
+    fi
+    manifest_args+=(--worlds-host-path "$FELIS_WORLDS_HOST_PATH" --archive-local-path "$FELIS_ARCHIVE_LOCAL_PATH")
+  fi
+  "$HOST_BIN" manifests "${manifest_args[@]}" | kube apply -f -
   restart_existing_control_plane "$had_api" "$had_operator"
 
   log "waiting for control-plane rollouts"
@@ -2167,10 +2396,54 @@ restart_existing_control_plane() {
   if [ "$had_operator" = "1" ]; then kube -n "$CONTROL_NS" rollout restart deployment/felis-operator; fi
 }
 
-# The login/lobby images use local mutable tags. Importing a replacement updates
+# push_image_to_registry <ref> re-tags a locally built image for the node's
-# containerd, but an existing StatefulSet template is byte-for-byte unchanged and
+# loopback push endpoint and uploads it. The registry keys a repository by the
-# Kubernetes will not roll it. Recreate only the two always-on system pods so a
+# path AFTER the host, so pushing 127.0.0.1:5000/felis/felis:demo lands exactly
-# convergent bootstrap actually starts the images it just imported.
+# where a later kubelet pull of registry.felis.svc:5000/felis/felis:demo (the
+# mirror rewrites the host) will look. A ref not under REGISTRY_URL is not
+# mirrored — warn, don't fail: the install is still self-consistent, that image
+# just has no pull source once GC collects its containerd copy.
+push_image_to_registry() {
+  local ref="$1" push_ref
+  case "$ref" in
+    "${REGISTRY_URL}/"*)
+      push_ref="${REGISTRY_PUSH_HOST}/${ref#"${REGISTRY_URL}/"}"
+      ;;
+    *)
+      warn "not mirroring ${ref} into the internal registry: it is not under ${REGISTRY_URL}; once the image GC collects that tag, nothing can re-pull it"
+      return 0
+      ;;
+  esac
+  log "mirroring ${ref} into the internal registry"
+  docker tag "$ref" "$push_ref" || die "could not tag ${ref} as ${push_ref} — is docker healthy?"
+  docker push "$push_ref" || die "could not mirror ${ref} into the internal registry — check the registry Deployment/pod and its PVC"
+  docker rmi "$push_ref" >/dev/null 2>&1 || true
+}
+
+# Every image this installer builds is hosted in the registry, so the copies it
+# imported into containerd are a first-boot cache, not the only copy: kubelet
+# re-pulls from the registry after any image GC. Runs AFTER deploy_bundle — the
+# registry it pushes into does not exist before that.
+#
+# Docker is started once for the whole batch and stopped once at the end. A
+# start/stop pair per image trips systemd's start rate limit — observed live on
+# a re-run: three fast pushes, then "Start request repeated too quickly /
+# start-limit-hit" and the fourth image never got mirrored. docker.service is
+# socket-triggered, so each cycle counts twice against the burst limit.
+push_images_to_registry() {
+  local img
+  systemctl start docker
+  for img in "$FELIS_IMAGE" "$FELIS_LIMBO_IMAGE" "$FELIS_LOBBY_IMAGE" "$FELIS_PAPER_IMAGE"; do
+    [ -n "$img" ] || continue
+    push_image_to_registry "$img"
+  done
+  systemctl stop docker docker.socket 2>/dev/null || true
+}
+
+# The login/lobby images use mutable :demo tags. Importing/pushing a replacement
+# updates containerd, but an existing StatefulSet template is byte-for-byte
+# unchanged and Kubernetes will not roll it. Recreate only the two always-on
+# system pods so a convergent bootstrap actually starts the images it just built.
 restart_existing_system_servers() {
   local name pods
   for name in "$LOGIN_SERVER" "$LOBBY_SERVER"; do
@@ -2575,6 +2848,11 @@ main() {
   ensure_panel_tls_cert
   install_docker
   install_k3s
+  # The registry mirror must exist before the bundle's pods start pulling (and
+  # before any re-run's rollouts); the registry's own image must be in containerd
+  # before its Deployment can start at all.
+  configure_registry_mirror
+  import_registry_image
   # Three ways to end up with a felis binary, in preference order. The release download is
   # the only one that skips compiling: it is the CI artifact for this exact tag, panel
   # included. Both other arms leave HAVE_PREBUILT_BINARY unset where a source build is what
@@ -2592,6 +2870,9 @@ main() {
   configure_postgres
   run_migrations
   deploy_bundle
+  # AFTER deploy_bundle: the registry the built images are mirrored into is part
+  # of that bundle.
+  push_images_to_registry
   restart_existing_system_servers
   # After deploy_bundle: the proxy dials felis-api's internal ClusterIP, which does not
   # exist until the bundle is applied.

deploy/bootstrap_test.sh

@@ -192,6 +192,83 @@ for hdr in '[[ auth_source ]]' '[["auth_source"]]' "[['auth_source']]"; do
   esac
 done
 
+# --- [smtp] carry-forward does not hoard the auth_source comment block -------------------
+# persisted_smtp_block used to print every line between [smtp] and the next section
+# header -- which includes the generated Yggdrasil comment block that sits above
+# [[auth_source]]. Each re-run re-emitted that hoard plus a fresh template copy, so both
+# config files grew by one comment block per run (audit #50). The carry must be the
+# section's header and keys only, and must be byte-stable when written back.
+
+sblock="$(awk '/^persisted_smtp_block\(\) \{/,/^}/' "$BS")"
+[ -n "$sblock" ] || { echo "FAIL: no persisted_smtp_block found in $BS"; exit 1; }
+[ "$(printf '%s\n' "$sblock" | wc -l)" -lt 40 ] \
+  || { echo "FAIL: the extracted block is not the function -- did its closing brace move?"; exit 1; }
+
+sfn="$(mktemp)"
+printf '%s\n' "$sblock" > "$sfn"
+smtp_dir="$(mktemp -d)"
+trap 'rm -f "$jar" "$sfn"; rm -rf "$vdir" "$sdir" "$smtp_dir"' EXIT
+
+run_smtp() { # state-dir
+  STATE_DIR="$1" SBLOCK_FILE="$sfn" bash -c '. "$SBLOCK_FILE"; persisted_smtp_block'
+}
+
+cat > "$smtp_dir/felis.host.toml" <<'TOML'
+[server]
+listen = "0.0.0.0:8080"
+
+[smtp]
+  host = "mail.example"
+  port = 587
+  from = "[email protected]"
+  username = "relay-user"
+  password_ref = "smtp-password"
+
+# Third-party Yggdrasil sources federated by the hasJoined multiplexer. Mojang is
+# always the code-owned identity anchor (premium-first), prepended in Go; sources here
+# append as namespace-rewritten guests. A fresh install federates LittleSkin. Edit the
+# list in /etc/felis/felis.host.toml and rerun the installer; re-runs keep it as it
+# is, and with no [[auth_source]] at all the server is Mojang-only.
+
+[[auth_source]]
+  tag = "littleskin"
+  prefix = "LS"
+  url = "https://littleskin.cn/api/yggdrasil/sessionserver/session/minecraft/hasJoined"
+TOML
+
+out="$(run_smtp "$smtp_dir")"
+expect "a configured [smtp] relay is carried" 'host = "mail.example"' "$out"
+expect "its port survives the carry" 'port = 587' "$out"
+expect "its credentials reference survives" 'password_ref = "smtp-password"' "$out"
+case "$out" in
+  *"#"*)
+    echo "FAIL: the carry hoards comment lines:"; printf '%s\n' "$out"; fails=$((fails + 1)) ;;
+  *) echo "PASS the carry is header and keys only -- no comment hoard" ;;
+esac
+case "$out" in
+  *"[["*)
+    echo "FAIL: the carry ran into the next section:"; printf '%s\n' "$out"; fails=$((fails + 1)) ;;
+  *) echo "PASS the carry stops at the next section header" ;;
+esac
+
+# Write the carry back the way write_felis_toml does (carry + one fresh template block +
+# the tables) and extract again: a second re-run must add nothing.
+{
+  printf '%s\n' "$out"
+  printf '\n%s\n' '# Third-party Yggdrasil sources federated by the hasJoined multiplexer. Mojang is'
+  printf '%s\n' '[[auth_source]]' '  tag = "littleskin"' '  prefix = "LS"' \
+    '  url = "https://littleskin.cn/api/yggdrasil/sessionserver/session/minecraft/hasJoined"'
+} > "$smtp_dir/felis.host.toml"
+out2="$(run_smtp "$smtp_dir")"
+printf '%s\n' "$out" > "$smtp_dir/first"
+printf '%s\n' "$out2" > "$smtp_dir/second"
+if cmp -s "$smtp_dir/first" "$smtp_dir/second"; then
+  echo "PASS a carried-forward [smtp] converges (a second re-run adds nothing)"
+else
+  echo "FAIL: carrying [smtp] is not idempotent:"; diff "$smtp_dir/first" "$smtp_dir/second" | head
+  fails=$((fails + 1))
+fi
+
 # --- write_nano_config leaves the unit able to read its config ---------------------------
 # felis-nano runs as a DynamicUser, so the directory must be searchable by others under a
 # hardened umask too, including one an older installer left at 0750 -- but the full
@@ -574,6 +651,295 @@ mkdir -p "$sdir/src/.git"
 expect "a failed fetch into an existing checkout names the token" "set FELIS_GITHUB_TOKEN" \
   "$(run_fetch "$sdir/src")"
 
+# --- default install keeps backups, and retention envs reach the renderer ----------------
+# A default install must render the world-archive PVC (without one, backup/restore answer an
+# honest 503), and FELIS_WORLDS_HOST_PATH must turn into the reaper's two flags or an
+# operator's retention enablement silently renders no CronJob. Extracted, not retyped.
+
+mblock="$(awk '/^  log "rendering \+ applying the control-plane bundle"/,/kube apply -f -/' "$BS")"
+[ -n "$mblock" ] || { echo "FAIL: no manifest_args block found in $BS"; exit 1; }
+[ "$(printf '%s\n' "$mblock" | wc -l)" -lt 40 ] \
+  || { echo "FAIL: the extracted block is not the manifest_args block -- did it move?"; exit 1; }
+
+run_bundle_flags() { # backup-pvc worlds-host-path
+  FELIS_IMAGE=reg/felis:test FELIS_PANEL_NODEPORT=30443 NODE_IP=10.0.0.5 \
+  FELIS_BACKUP_PVC="$1" FELIS_WORLDS_HOST_PATH="$2" FELIS_ARCHIVE_LOCAL_PATH=/var/lib/felis/archives \
+  HOST_BIN=myManifests bash -c '
+    log() { :; }
+    warn() { printf "WARN: %s\n" "$*"; }
+    kube() { cat; }
+    myManifests() { printf "%s\n" "$@"; }
+    setfacl() { printf "SETFACL %s\n" "$*"; }
+    run_bundle() {
+    '"$mblock"'
+    }
+    run_bundle'
+}
+
+out="$(run_bundle_flags felis-backups '')"
+expect "a default install asks the renderer for the archive PVC" "--backup-pvc
+felis-backups" "$out"
+case "$out" in
+  *--worlds-host-path*) echo "FAIL: no reaper flags may render without FELIS_WORLDS_HOST_PATH"; fails=$((fails + 1)) ;;
+esac
+
+out="$(run_bundle_flags '' '')"
+expect "an emptied FELIS_BACKUP_PVC is the explicit no-backup shape" "--backup-pvc=" "$out"
+
+out="$(run_bundle_flags felis-backups /var/lib/rancher/k3s/storage)"
+expect "enabling retention passes the worlds root" "--worlds-host-path
+/var/lib/rancher/k3s/storage" "$out"
+expect "enabling retention passes the archive mount that must match felis.toml" "--archive-local-path
+/var/lib/felis/archives" "$out"
+
+# The warn fires only when the root is ABSENT (hostPath type Directory would fail);
+# the case above passes a path that exists on any host already running k3s, so it
+# must not also demand the warning — probing the real /var/lib/rancher path made
+# this suite red on exactly the hosts the installer is for. Point the warn case at
+# a path guaranteed missing.
+missing="/tmp/felis-worlds-root-must-not-exist-$$"
+out="$(run_bundle_flags felis-backups "$missing")"
+expect "a missing worlds root is warned about, not silently skipped" "WARN: worlds root $missing does not exist yet" "$out"
+
+# The reaper pod is non-root (uid 1000) and k3s ships the storage root 0700 root:root, so
+# the installer must grant traverse or every archive dies with permission denied.
+wdir="$(mktemp -d)"
+out="$(run_bundle_flags felis-backups "$wdir")"
+expect "enabling retention grants the reaper uid traverse on the worlds root" "SETFACL -m u:1000:x $wdir" "$out"
+
+# --- the registry mirror writer -----------------------------------------------------------
+# k3s only consults registries.yaml at agent start, so a CONTENT change must restart k3s and
+# an identical file (every re-run) must restart nothing. The k3s restart is the expensive,
+# disruptive half of the pair -- getting the idempotence wrong bounces the whole cluster on
+# every installer re-run, so both halves are pinned here against the extracted function.
+
+cmblock="$(awk '/^configure_registry_mirror\(\) \{/,/^}/' "$BS")"
+[ -n "$cmblock" ] || { echo "FAIL: no configure_registry_mirror found in $BS"; exit 1; }
+
+run_mirror() { # scratch-file
+  K3S_REGISTRIES_FILE="$1" REGISTRY_URL=registry.felis.svc:5000 REGISTRY_PUSH_HOST=127.0.0.1:5000 \
+  bash -c '
+    log() { printf "LOG: %s\n" "$*"; }
+    ok() { printf "OK: %s\n" "$*"; }
+    die() { printf "DIE: %s\n" "$*"; exit 1; }
+    warn() { printf "WARN: %s\n" "$*"; }
+    remember_temp() { :; }
+    systemctl() { printf "SYSTEMCTL %s\n" "$*"; }
+    kube() { printf "n Ready \n"; }
+    wait_for_node_ready() { kube get nodes | grep -q " Ready " && ok "k3s node Ready"; }
+    '"$cmblock"'
+    configure_registry_mirror'
+}
+
+mfile="$(mktemp -u)"
+out="$(run_mirror "$mfile")"
+expect "a missing registries.yaml is written" "\"registry.felis.svc:5000\":" "$(cat "$mfile" 2>/dev/null)"
+expect "the mirror endpoint is the node loopback push/pull host" "\"http://127.0.0.1:5000\"" "$(cat "$mfile" 2>/dev/null)"
+expect "a content change restarts k3s" "SYSTEMCTL restart k3s" "$out"
+
+out="$(run_mirror "$mfile")"
+expect "an identical registries.yaml is recognised" "already configured" "$out"
+case "$out" in
+  *"SYSTEMCTL restart"*) echo "FAIL: a re-run with identical content must not restart k3s"; fails=$((fails + 1)) ;;
+esac
+
+printf 'mirrors: {}\n' >"$mfile"
+out="$(run_mirror "$mfile")"
+expect "changed content restarts k3s again" "SYSTEMCTL restart k3s" "$out"
+rm -f "$mfile"
+
+# --- image mirroring ----------------------------------------------------------------------
+# The registry keys a repository by the path AFTER the host, so the push must swap the
+# registry host for the node's loopback endpoint and nothing else. A ref outside the
+# registry must be warned about, not silently pushed somewhere unintended.
+
+pblock="$(awk '/^push_image_to_registry\(\) \{/,/^}/' "$BS")"
+[ -n "$pblock" ] || { echo "FAIL: no push_image_to_registry found in $BS"; exit 1; }
+
+run_push() { # ref [docker-push-exit]
+  REF="$1" PUSH_EXIT="${2:-0}" \
+  REGISTRY_URL=registry.felis.svc:5000 REGISTRY_PUSH_HOST=127.0.0.1:5000 \
+  bash -c '
+    log() { printf "LOG: %s\n" "$*"; }
+    warn() { printf "WARN: %s\n" "$*"; }
+    die() { printf "DIE: %s\n" "$*"; exit 1; }
+    ok() { :; }
+    systemctl() { :; }
+    docker() {
+      case "$1" in
+        push) printf "DOCKER %s\n" "$*"; return "$PUSH_EXIT" ;;
+        *) printf "DOCKER %s\n" "$*" ;;
+      esac
+    }
+    '"$pblock"'
+    push_image_to_registry "$REF"'
+}
+
+out="$(run_push registry.felis.svc:5000/felis/felis:demo)"
+expect "a registry ref is re-tagged onto the node loopback endpoint" \
+  "DOCKER tag registry.felis.svc:5000/felis/felis:demo 127.0.0.1:5000/felis/felis:demo" "$out"
+expect "and pushed to exactly that endpoint" "DOCKER push 127.0.0.1:5000/felis/felis:demo" "$out"
+
+out="$(run_push registry.felis.svc:50000/felis/felis:demo)"
+expect "a ref outside the registry is refused with a warning" "WARN: not mirroring" "$out"
+case "$out" in
+  *"DOCKER push"*) echo "FAIL: a non-registry ref must not be pushed"; fails=$((fails + 1)) ;;
+esac
+
+out="$(run_push registry.felis.svc:5000/felis/felis:demo 1)"
+expect "a failed push fails the install loudly" "DIE: could not mirror" "$out"
+
+# docker must be started ONCE for the whole batch: a start/stop pair per image trips
+# systemd's start rate limit ("start-limit-hit" — observed live; the 4th image was never
+# mirrored because docker.service is socket-triggered and each cycle counts twice).
+wiblock="$(awk '/^push_images_to_registry\(\) \{/,/^}/' "$BS")"
+[ -n "$wiblock" ] || { echo "FAIL: no push_images_to_registry found in $BS"; exit 1; }
+out="$(
+  FELIS_IMAGE=a FELIS_LIMBO_IMAGE=b FELIS_LOBBY_IMAGE=c FELIS_PAPER_IMAGE=d bash -c '
+    systemctl() { printf "SYSTEMCTL %s\n" "$*"; }
+    push_image_to_registry() { printf "PUSH %s\n" "$1"; }
+    '"$wiblock"'
+    push_images_to_registry'
+)"
+starts="$(printf '%s\n' "$out" | grep -c 'SYSTEMCTL start docker')"
+stops="$(printf '%s\n' "$out" | grep -c 'SYSTEMCTL stop docker')"
+[ "$starts" = 1 ] && [ "$stops" = 1 ] && [ "$(printf '%s\n' "$out" | grep -c '^PUSH')" = 4 ] \
+  && echo "PASS the batch wraps all four pushes in ONE docker start/stop" \
+  || { echo "FAIL: expected 1 start / 1 stop / 4 pushes, got:"; printf '%s\n' "$out"; fails=$((fails + 1)); }
+
+# --- the registry's own image must not be re-pulled on every run --------------------------
+iblock="$(awk '/^import_registry_image\(\) \{/,/^}/' "$BS")"
+[ -n "$iblock" ] || { echo "FAIL: no import_registry_image found in $BS"; exit 1; }
+
+out="$(
+  bash -c '
+    log() { printf "LOG: %s\n" "$*"; }
+    ok() { printf "OK: %s\n" "$*"; }
+    warn() { printf "WARN: %s\n" "$*"; }
+    die() { printf "DIE: %s\n" "$*"; exit 1; }
+    systemctl() { :; }
+    k3s_cmd() { case "$*" in "ctr images ls -q") printf "docker.io/library/registry:2\n" ;; esac; }
+    docker() { printf "DOCKER %s\n" "$*"; return 1; }
+    '"$iblock"'
+    import_registry_image'
+)"
+expect "an already-imported registry:2 is left alone" "already in k3s containerd" "$out"
+case "$out" in
+  *DOCKER*) echo "FAIL: a present registry:2 must not trigger a docker pull"; fails=$((fails + 1)) ;;
+esac
+
+# --- installer re-runs refresh the workload namespace's felis-config copy ---------------
+# The backup/restore/fileedit Jobs and the reaper mount the workload namespace's own
+# felis-config (a secretKeyRef is namespace-local). `felis setup` makes that replica
+# create-if-absent -- right for credentials, wrong for a rendered config -- so the
+# installer must refresh it every run; a stale copy keeps old DB/archive settings.
+
+fcblock="$(awk '/^apply_felis_config_secrets\(\) \{/,/^}/' "$BS")"
+[ -n "$fcblock" ] || { echo "FAIL: no apply_felis_config_secrets found in $BS"; exit 1; }
+[ "$(printf '%s\n' "$fcblock" | wc -l)" -lt 20 ] \
+  || { echo "FAIL: the extracted block is not the function -- did its closing brace move?"; exit 1; }
+
+kubcalls="$(mktemp)"
+run_fc() {
+  : > "$kubcalls"
+  CONTROL_NS=felis MINECRAFT_NS=minecraft STATE_DIR=/tmp/fc KUBCALLS="$kubcalls" bash -c '
+    kube() { printf "%s\n" "$*" >> "$KUBCALLS"; }
+    '"$fcblock"'
+    apply_felis_config_secrets'
+  cat "$kubcalls"
+}
+out="$(run_fc)"
+expect "the control plane's felis-config is applied" \
+  "-n felis create secret generic felis-config" "$out"
+expect "the workload namespace's copy is applied too" \
+  "-n minecraft create secret generic felis-config" "$out"
+expect "both copies render from the pod config" \
+  "felis.toml=/tmp/fc/felis.pod.toml" "$out"
+applies="$(printf '%s\n' "$out" | grep -c '^apply -f -$')"
+if [ "$applies" -eq 2 ]; then
+  echo "PASS both rendered copies are piped to kubectl apply"
+else
+  echo "FAIL: expected 2 applies, got $applies:"; printf '%s\n' "$out"; fails=$((fails + 1))
+fi
+rm -f "$kubcalls"
+
+# --- installer re-runs keep the operator's [registry] overrides --------------------------
+# §15's upgrade path is re-running the installer, but the build-lane mirrors and the
+# uploads backend live in [registry] as hand-written keys (docs/troubleshooting.md §8e or
+# the storage wizard) that nothing in this script's inputs derives. A re-run must carry
+# them forward — without letting a stale url/build_namespace survive (installer-owned).
+
+wrblock="$(awk '/^write_felis_toml\(\) \{/,/^}/' "$BS")"
+prblock="$(awk '/^persisted_registry_block\(\) \{/,/^}/' "$BS")"
+pablock="$(awk '/^persisted_archive_block\(\) \{/,/^}/' "$BS")"
+{ [ -n "$wrblock" ] && [ -n "$prblock" ] && [ -n "$pablock" ]; } \
+  || { echo "FAIL: write_felis_toml / persisted_{registry,archive}_block not found in $BS"; exit 1; }
+# The blocks quote themselves (the awk program uses single quotes), so they are
+# sourced from a file instead of being spliced into a single-quoted bash -c.
+fnfile="$(mktemp)"
+printf '%s\n%s\n%s\n' "$prblock" "$pablock" "$wrblock" > "$fnfile"
+
+rdir="$(mktemp -d)"
+cat > "$rdir/felis.host.toml" <<'TOML'
+[registry]
+url = "stale.invalid:5000"
+build_namespace = "stale-ns"
+kaniko_image = "registry.felis.svc:5000/mirror/kaniko-executor:v1.24.0"
+trivy_db_repository = "registry.felis.svc:5000/mirror/trivy-db:2"
+trivy_java_db_repository = "registry.felis.svc:5000/mirror/trivy-java-db:1"
+
+[registry.s3]
+endpoint = "https://s3.example"
+region = "us-east-1"
+
+[archive]
+store = "tarLocal"
+local_path = "/stale/path"
+retention = "30d"
+TOML
+
+run_write() { # out-file
+  STATE_DIR="$rdir" OUT_TOML="$1" FNFILE="$fnfile" bash -c '
+    log() { :; }
+    persisted_smtp_block() { :; }
+    persisted_auth_source_blocks() { :; }
+    . "$FNFILE"
+    FELIS_ROOT_DOMAIN=r.example.com DB_USER=u DB_PASSWORD=p DB_NAME=d MINECRAFT_NS=minecraft \
+    FELIS_EGRESS_MODE=nodeport FELIS_LIMBO_IMAGE=li FELIS_LOBBY_IMAGE=lo \
+    REGISTRY_URL=registry.felis.svc:5000 BUILD_NS=felis-build FELIS_ARCHIVE_LOCAL_PATH=/a \
+    write_felis_toml "$OUT_TOML" 127.0.0.1'
+}
+
+run_write "$rdir/out.toml"
+out="$(cat "$rdir/out.toml")"
+expect "a re-run carries the build-lane executor mirrors" \
+  'kaniko_image = "registry.felis.svc:5000/mirror/kaniko-executor:v1.24.0"' "$out"
+expect "a re-run carries the trivy vulnerability-DB mirror" \
+  'trivy_db_repository = "registry.felis.svc:5000/mirror/trivy-db:2"' "$out"
+expect "a re-run carries the trivy java-DB mirror" \
+  'trivy_java_db_repository = "registry.felis.svc:5000/mirror/trivy-java-db:1"' "$out"
+expect "a re-run carries the [registry.s3] uploads subtable" "[registry.s3]" "$out"
+expect "the carried subtable keeps its keys" 'endpoint = "https://s3.example"' "$out"
+expect "url stays installer-owned" 'url = "registry.felis.svc:5000"' "$out"
+expect "a re-run carries the archive retention window" 'retention = "30d"' "$out"
+expect "the archive mount stays installer-owned" 'local_path = "/a"' "$out"
+case "$out" in
+  *stale.invalid* | *stale-ns* | *stale/path*)
+    echo "FAIL: stale installer-owned values survived the re-run"; fails=$((fails + 1)) ;;
+esac
+
+cp "$rdir/out.toml" "$rdir/felis.host.toml"
+run_write "$rdir/out2.toml"
+if cmp -s "$rdir/out.toml" "$rdir/out2.toml"; then
+  echo "PASS a carried-forward config converges (the second re-run is a no-op)"
+else
+  echo "FAIL: carrying [registry] overrides is not idempotent"
+  diff "$rdir/out.toml" "$rdir/out2.toml" | head
+  fails=$((fails + 1))
+fi
+
+rm -f "$fnfile"
+
 # ---------------------------------------------------------------------------------------
 if [ "$fails" -eq 0 ]; then
   echo "ALL PASS"

deploy/crd/felis.lolicon.best_minecraftservers.yaml

@@ -347,6 +347,14 @@ spec:
                   - type
                   type: object
                 type: array
+              emptySince:
+                description: |-
+                  EmptySince is when the operator first observed 0 online players during
+                  a Running phase (spec §8 idle auto-stop). It is reset when a player joins
+                  or the server stops, so the empty-duration counter starts fresh each time
+                  the server becomes unoccupied.
+                format: date-time
+                type: string
               endpoint:
                 description: Endpoint is where the proxy should route traffic.
                 properties:

deploy/demo-up.sh

@@ -1,30 +1,28 @@
 #!/bin/bash
 # demo-up.sh — one-shot Felis demo bring-up.
 #
-# Collapses the four manual steps (bootstrap -> build/import limbo+lobby images ->
-# edit felis.toml -> felis setup) into a single command:
-#
 #     sudo bash deploy/demo-up.sh
 #
-# It ends by exec'ing the interactive `felis setup` TUI (create the Owner account) —
+# Every piece a demo box needs — base platform (k3s + felis + docker + cloudflared +
-# that human step is the only thing this script cannot do for you.
+# control plane), the limbo/lobby/paper images, the felis-velocity proxy plugin, and
+# the [velocity] wiring in felis.host.toml — is built by deploy/bootstrap.sh. This
+# wrapper adds only the one step the installer cannot do: the interactive
+# `felis setup` TUI that creates the Owner account.
 #
-# Image source, in order of preference:
+# It used to rebuild the game images here with its own copy of that logic, written
-#   1. Prebuilt tars at deploy/images/felis-limbo.tar + felis-lobby.tar (imported as-is).
+# before bootstrap grew the job. The copy drifted: it pinned Paper 1.21.8 while the
-#   2. Otherwise built on this host with docker, resolving the LOOHP/Limbo CI jar and
+# installer derives one version from the Limbo login gate (both hops of a login must
-#      the latest stable Paper jar automatically. Override any of:
+# speak one protocol), never built felis-velocity.jar (so the proxy it wired had
-#        LIMBO_JAR_URL LIMBO_SCHEM_URL LIMBO_VERSION PAPER_JAR_URL PAPER_JAR_SHA256
+# nowhere to route), and left docker running. The installer is the single origin.
-#        PAPER_MC_VERSION
 #
-# Toggles: SKIP_BOOTSTRAP=1 (base already up), SKIP_SETUP=1 (stop before the TUI).
+# Toggles: SKIP_BOOTSTRAP=1 (base + game stack already installed by a full bootstrap),
+#          SKIP_SETUP=1 (stop before the TUI).
 set -Eeuo pipefail
 
 SRC_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
 STATE_DIR=/etc/felis
 HOST_TOML="$STATE_DIR/felis.host.toml"
-IMG_DIR="$SRC_DIR/deploy/images"
+PLUGIN_JAR=/opt/felis/velocity/plugins/felis-velocity.jar
-LIMBO_IMAGE="felis-limbo:demo"
-LOBBY_IMAGE="felis-lobby:demo"
 K3S=/usr/local/bin/k3s
 FELIS=/usr/local/bin/felis
 
@@ -33,11 +31,11 @@ die() { printf '\033[1;31mERROR: %s\033[0m\n' "$*" >&2; exit 1; }
 
 [ "$(id -u)" -eq 0 ] || die "run as root (sudo bash deploy/demo-up.sh)"
 
-# 1. base platform (k3s + felis + docker + cloudflared + control plane) ----------
+# 1. base platform + full game stack (deploy/bootstrap.sh) -----------------------
 if [ "${SKIP_BOOTSTRAP:-0}" = 1 ]; then
   log "SKIP_BOOTSTRAP=1 — assuming the base platform is already up"
 else
-  log "bringing up the base platform (deploy/bootstrap.sh)"
+  log "bringing up the base platform and the game stack (deploy/bootstrap.sh)"
   bash "$SRC_DIR/deploy/bootstrap.sh"
 fi
 
@@ -46,83 +44,19 @@ command -v "$K3S"   >/dev/null 2>&1 || K3S=k3s
 command -v "$K3S"   >/dev/null 2>&1 || die "k3s not found — did bootstrap complete?"
 command -v "$FELIS" >/dev/null 2>&1 || die "felis not found — did bootstrap complete?"
 
-# 2. get the two game images into k3s containerd --------------------------------
+# SKIP_BOOTSTRAP=1 trusts an earlier run to be complete. Check that it actually left
-if [ -f "$IMG_DIR/felis-limbo.tar" ] && [ -f "$IMG_DIR/felis-lobby.tar" ]; then
+# the full stack behind: a base from before the game-stack installer, or one whose
-  log "importing prebuilt image tars from $IMG_DIR"
+# pieces were pruned by hand, must fail here with a pointer — not present as a proxy
-  "$K3S" ctr images import "$IMG_DIR/felis-limbo.tar"
+# that accepts logins and routes nowhere, with nothing in any log to say why.
-  "$K3S" ctr images import "$IMG_DIR/felis-lobby.tar"
-  # Optional: the plain-Paper recommended base, if a tar was staged for it.
-  [ -f "$IMG_DIR/felis-paper.tar" ] && "$K3S" ctr images import "$IMG_DIR/felis-paper.tar"
-else
-  log "no prebuilt tars in $IMG_DIR — building on this host with docker"
-  command -v docker >/dev/null 2>&1 || die "docker not found; cannot build images"
-
-  rel=$(curl -fsSL --max-time 30 "https://ci.loohpjames.com/job/Limbo/lastSuccessfulBuild/api/json" \
-          | grep -oE 'target/Limbo-[0-9][^"]+\.jar' | head -1) || true
-  : "${LIMBO_JAR_URL:=https://ci.loohpjames.com/job/Limbo/lastSuccessfulBuild/artifact/$rel}"
-  : "${LIMBO_SCHEM_URL:=https://ci.loohpjames.com/job/Limbo/lastSuccessfulBuild/artifact/spawn.schem}"
-  : "${LIMBO_VERSION:=$(basename "$rel" | sed -E 's/^Limbo-//; s/\.jar$//; s/-[0-9]+\.[0-9]+$//')}"
-  [ -n "$rel" ] || [ -n "${LIMBO_JAR_URL##*artifact/}" ] || die "could not resolve the Limbo jar; set LIMBO_JAR_URL"
-  log "building $LIMBO_IMAGE (Limbo $LIMBO_VERSION)"
-  docker build -f "$SRC_DIR/deploy/limbo/Dockerfile" \
-    --build-arg LIMBO_JAR_URL="$LIMBO_JAR_URL" \
-    --build-arg LIMBO_SCHEM_URL="$LIMBO_SCHEM_URL" \
-    --build-arg LIMBO_VERSION="$LIMBO_VERSION" \
-    -t "$LIMBO_IMAGE" "$SRC_DIR"
-  docker save "$LIMBO_IMAGE" | "$K3S" ctr images import -
-
-  : "${PAPER_MC_VERSION:=1.21.8}"
-  : "${PAPER_JAR_URL:=$(curl -fsSL --max-time 30 "https://fill.papermc.io/v3/projects/paper/versions/${PAPER_MC_VERSION}/builds/latest" | grep -oE 'https://fill-data\.papermc\.io/[^"]+\.jar' | head -1)}"
-  [ -n "$PAPER_JAR_URL" ] || die "could not resolve the Paper jar; set PAPER_JAR_URL"
-  # Both Dockerfiles require the jar's digest. The fill-data URL is content-addressed
-  # (the objects/ path segment IS the sha256), so it is derived rather than asked for;
-  # a mirror override carries no such segment and must bring its own digest.
-  if [ -z "${PAPER_JAR_SHA256:-}" ]; then
-    sha="${PAPER_JAR_URL#*/objects/}"
-    sha="${sha%%/*}"
-    case "$sha" in
-      *[!0-9a-f]*|"") sha="" ;;
-    esac
-    if [ "${#sha}" -ne 64 ]; then
-      die "cannot derive the Paper jar sha256 from PAPER_JAR_URL (not a content-addressed fill-data URL); set PAPER_JAR_SHA256"
-    fi
-    PAPER_JAR_SHA256="$sha"
-  fi
-  log "building $LOBBY_IMAGE (Paper $PAPER_MC_VERSION)"
-  docker build -f "$SRC_DIR/deploy/lobby/Dockerfile" \
-    --build-arg PAPER_JAR_URL="$PAPER_JAR_URL" \
-    --build-arg PAPER_JAR_SHA256="$PAPER_JAR_SHA256" \
-    -t "$LOBBY_IMAGE" "$SRC_DIR"
-  docker save "$LOBBY_IMAGE" | "$K3S" ctr images import -
-
-  # Plain Paper recommended base — same PAPER_JAR_URL, no plugins, no secret gate.
-  : "${PAPER_IMAGE:=felis-paper:demo}"
-  log "building $PAPER_IMAGE (plain Paper $PAPER_MC_VERSION, forwarding via the operator initContainer)"
-  docker build -f "$SRC_DIR/deploy/paper/Dockerfile" \
-    --build-arg PAPER_JAR_URL="$PAPER_JAR_URL" \
-    --build-arg PAPER_JAR_SHA256="$PAPER_JAR_SHA256" \
-    -t "$PAPER_IMAGE" "$SRC_DIR"
-  docker save "$PAPER_IMAGE" | "$K3S" ctr images import -
-fi
-
-# 3. wire the images into the config `felis setup` reads ------------------------
-log "wiring [velocity] images into $HOST_TOML"
 [ -f "$HOST_TOML" ] || die "missing $HOST_TOML — did bootstrap run?"
-if grep -q '^\[velocity\]' "$HOST_TOML"; then
+grep -q '^\[velocity\]' "$HOST_TOML" \
-  echo "  [velocity] table already present — leaving it untouched"
+  || die "$HOST_TOML has no [velocity] section — re-run the installer without SKIP_BOOTSTRAP so the system servers get wired"
-else
+[ -f "$PLUGIN_JAR" ] \
-  cat >> "$HOST_TOML" <<EOF
+  || die "$PLUGIN_JAR missing — this base did not finish the full installer, and a proxy without it silently routes nothing; re-run the installer without SKIP_BOOTSTRAP"
 
-[velocity]
+# 2. interactive Owner creation + system-server provisioning --------------------
-login_image = "$LIMBO_IMAGE"
-lobby_image = "$LOBBY_IMAGE"
-EOF
-  echo "  appended login_image=$LIMBO_IMAGE / lobby_image=$LOBBY_IMAGE"
-fi
-
-# 4. interactive Owner creation + system-server provisioning -------------------
 if [ "${SKIP_SETUP:-0}" = 1 ]; then
-  log "SKIP_SETUP=1 — base + images + config ready. Finish with:  sudo felis setup"
+  log "SKIP_SETUP=1 — base + game stack ready. Finish with:  sudo felis setup"
 else
   log "launching 'felis setup' — create the Owner account (this is the only interactive step)"
   exec "$FELIS" setup

deploy/limbo/README.md

@@ -90,14 +90,21 @@ docker build -f deploy/limbo/Dockerfile \
   version `2026.0.2-ALPHA` (the `-26.2` CI qualifier is not published to the
   maven repo).
 
-Import into k3s and point config at it:
+Publish it into the cluster's registry and point config at it. On the node
+itself (docker treats `127.0.0.1` as insecure by default):
 
 ```
-docker save felis-limbo:demo | sudo k3s ctr images import -
+docker tag  felis-limbo:demo 127.0.0.1:5000/felis/limbo:demo
-# felis.toml → [velocity] login_image = "felis-limbo:demo"
+docker push 127.0.0.1:5000/felis/limbo:demo
+# felis.toml → [velocity] login_image = "registry.felis.svc:5000/felis/limbo:demo"
 sudo felis setup
 ```
 
+The registry keys a repository by the path after the host, so pushing through a
+`kubectl -n felis port-forward svc/registry 5000:5000` from another machine is
+equivalent. Hosting the image in the registry (rather than only importing it
+into containerd) is what lets kubelet re-pull it after an image GC.
+
 ## Ports (handled for you)
 
 The entrypoint (`deploy/limbo/entrypoint.sh`) pins Limbo's `server-port` to

deploy/lobby/README.md

@@ -31,8 +31,12 @@ docker build -f deploy/lobby/Dockerfile \
   --build-arg PAPER_JAR_URL=https://<mirror>/paper-1.21.x-<build>.jar \
   --build-arg PAPER_JAR_SHA256=<sha256 of that jar> \
   -t felis-lobby:demo .
-docker save felis-lobby:demo | sudo k3s ctr images import -
+# Publish into the cluster's registry (on the node; docker treats 127.0.0.1 as
-# felis.toml → [velocity] lobby_image = "felis-lobby:demo"
+# insecure by default — or through a `kubectl -n felis port-forward svc/registry
+# 5000:5000`, which is equivalent: only the path after the host matters).
+docker tag  felis-lobby:demo 127.0.0.1:5000/felis/lobby:demo
+docker push 127.0.0.1:5000/felis/lobby:demo
+# felis.toml → [velocity] lobby_image = "registry.felis.svc:5000/felis/lobby:demo"
 sudo felis setup
 ```
 

docs/deferred-seams.md

@@ -42,9 +42,27 @@ A grep across `*.md` and `*.go` returns both sets; only the Go ones are seams.
   does not exist. `felis update` runs with a zero window, under which every
   `Scheduled` component degrades to a notify, so no path can currently claim an
   apply is under way.
-- `internal/submit/blobstore.go:40` — the uploads PVC is mounted into felis-api but
+- `internal/submit/blobstore.go` — CLOSED 2026-09-22. The uploads PVC still cannot
-  not into the Kaniko build Pod, so a submitted context is durable at the derived
+  cross namespaces, so the transport went through the API instead of a mount: the
-  location without yet being readable by the build that consumes it.
+  derived context ref is now the internal-face URL
+  (`/api/v1/internal/submissions/{id}/context`, service-token gated), the build
+  Job's `context-fetch` initContainer streams it with `felis fetch-context` and
+  extracts under a zip-slip guard into a size-limited emptyDir, and Kaniko builds
+  `--context=/context`. The token reaches the build namespace through the same
+  Secret-replica mechanism the login gate uses (bootstrap + `felis setup`), and the
+  build egress lock allows exactly the control namespace on the internal port.
+  Uniform for local and s3:// stores — neither hands the sandboxed build Pod a
+  filesystem view or object-store credentials. Kaniko/Trivy images are
+  external-only by default; `[registry] kaniko_image / trivy_image /
+  build_cpu_limit / build_mem_limit` override them for mirrored or air-gapped
+  installs. Trivy's vulnerability DB is the same story, and now has its own knob:
+  `[registry] trivy_db_repository` points `--db-repository` at an internal mirror
+  (recipe in docs/troubleshooting.md §8e); `trivy_java_db_repository` does the
+  same for the Java DB, which Trivy fetches so soon as the scanned image contains
+  a jar — i.e. for every real modpack build. Left unset on an egress-locked box
+  the scan step fails closed — Kaniko pushes, Trivy exits on the DB download —
+  which is the correct fail direction but leaves the build unfinished, so the
+  mirrors are part of a production build install.
 
 ## Built; only its I/O is unverifiable from this repo
 
@@ -58,17 +76,25 @@ or a real upstream account to run it against — not an implementation.
   drives the whole flow through a fake.
 - `internal/api/console.go:39`, `internal/api/logstream.go:236`,
   `internal/api/logstream.go:306`, `internal/fileedit/k8sjobs.go:45` — each needs a
-  live cluster (RCON, `pods/log` follow, a Job).
+  live cluster (RCON, `pods/log` follow, a Job). **Verified live 2026-09-22/23
+  (auditfix7–25):** the RCON command spine (wake → probe → `command`/access
+  mutations/stop), the log SSE stream, and the fileedit Job have each run
+  end-to-end on the drill cluster.
 - `internal/api/handlers_access.go:170,490` — parsing real vanilla and LuckPerms
-  command output.
+  command output. **Verified live 2026-09-23:** players / whitelist / banlist
+  parses matched a live Paper server's replies (LuckPerms not installed → the raw
+  reply falls through as documented; the input guards held on four negative cases).
 
 ## Deliberately accepted, not scheduled to close
 
 These are decisions, not backlog. Each names the condition under which it would be
 worth revisiting.
 
-- `internal/api/pgrepo.go:281` — the quota check and `ClaimServer` are two statements
+- ~~`internal/api/pgrepo.go:281` — the quota check and `ClaimServer` are two statements
-  (audit #4 TOCTOU). Closeable only against a real Postgres.
+  (audit #4 TOCTOU). Closeable only against a real Postgres.~~ **Closed** — the gate
+  moved inside `ClaimServer` (advisory lock + re-check + UPDATE in one transaction),
+  red-then-green in the pgint suite, which is exactly the real-Postgres harness this
+  line was waiting for.
 - `internal/api/api.go:671` — `cooldownLimiter` is process-local, so across N api
   replicas a caller could draw up to N OTP codes per window. The intra-replica burst
   is closed; cross-replica bounding needs a shared store, out of scope for a

docs/openapi.yaml

@@ -337,6 +337,19 @@ components:
         build_id:
           type: string
           description: image_builds.id, set only after the build hand-off succeeds.
+        build_status:
+          type: string
+          enum: [pending, building, succeeded, failed, cancelled]
+          description: >-
+            The linked build's outcome, attached by the LIST routes
+            (/me/submissions, /submissions) — for a submitter this is the only
+            visible outlet for a failed build. Omitted until a build is linked
+            and its row is readable.
+        build_error:
+          type: string
+          description: >-
+            The build's recorded failure text (e.g. a CRITICAL CVE scan
+            failure), attached alongside build_status.
         reviewed_by: { type: string }
         reject_reason: { type: string }
         created_at: { type: string, format: date-time }
@@ -454,6 +467,26 @@ paths:
         '503':
           $ref: '#/components/responses/ServiceUnavailable'
 
+  /metrics:
+    get:
+      tags: [metrics]
+      operationId: metrics
+      summary: Prometheus metrics (felis_* collectors) on the internal face.
+      description: >-
+        Scrape-only infrastructure route, not a product API: the internal listener is
+        ClusterIP-only and a Prometheus scrape carries no token, the same stance as the
+        probes. Serves the felis_* exposition documented in troubleshooting §14; the
+        external face never serves it.
+      x-felis-face: [internal]
+      x-felis-tier: public
+      security: []
+      responses:
+        '200':
+          description: Prometheus text exposition format.
+          content:
+            text/plain:
+              schema: { type: string }
+
   /session/minecraft/hasJoined:
     get:
       tags: [nano]
@@ -607,6 +640,34 @@ paths:
         '404':
           $ref: '#/components/responses/NotFound'
 
+  /api/v1/internal/submissions/{id}/context:
+    get:
+      tags: [submissions-internal]
+      operationId: internalSubmissionContext
+      summary: Stream a submission's stored build-context tarball to the build Pod.
+      description: >-
+        The build Job's fetch initContainer cannot mount the control-plane uploads
+        PVC (a PVC does not cross namespaces) and holds no object-store
+        credentials, so the API that stored the blob streams it here. Served on
+        the internal face (service token, no Zero Trust).
+      x-felis-face: [internal]
+      x-felis-tier: service
+      security: [{ serviceToken: [] }]
+      parameters:
+        - { name: id, in: path, required: true, schema: { type: string } }
+      responses:
+        '200':
+          description: The stored gzip tarball, verbatim.
+          content:
+            application/gzip:
+              schema: { type: string, format: binary }
+        '401':
+          $ref: '#/components/responses/Unauthorized'
+        '404':
+          $ref: '#/components/responses/NotFound'
+        '503':
+          $ref: '#/components/responses/ServiceUnavailable'
+
   /api/v1/internal/servers/{name}/join-event:
     post:
       tags: [servers-internal]
@@ -2690,6 +2751,13 @@ paths:
                                 The owner's display identity (email, or username when
                                 the address is absent). Absent for an unclaimed server
                                 or when the best-effort owner lookup failed.
+                            system:
+                              type: boolean
+                              description: >-
+                                True for a platform-provisioned system service (the login
+                                gate, the lobby). Their reserved names are rejected by
+                                every per-server route, so the cockpit renders them
+                                read-only instead of offering actions that would 400.
         '401':
           $ref: '#/components/responses/Unauthorized'
         '403':
@@ -2771,11 +2839,12 @@ paths:
     post:
       tags: [backups]
       operationId: backupNow
-      summary: Back up a server's world on demand (owner-or-admin; server must be stopped).
+      summary: Back up a server's data volume on demand (owner-or-admin; server must be stopped).
       description: >-
-        Snapshots the server's world into the archive store as a first-class
+        Snapshots the server's whole data volume (worlds, config, plugins/mods,
-        world_backups row (reason "manual"), restorable later like an inactivity
+        jars, libraries — not just world folders) into the archive store as a
-        backup. The world PVC is RWO and held by a running server, so the server must
+        first-class world_backups row (reason "manual"), restorable later like an
+        inactivity backup. A restore replaces the volume with the archive. The world PVC is RWO and held by a running server, so the server must
         be fully stopped first (409 not_stopped otherwise). The backup runs
         asynchronously as a Job, so success is 202 (backing_up).
       x-felis-face: [external]
@@ -2811,6 +2880,56 @@ paths:
         '503':
           $ref: '#/components/responses/ServiceUnavailable'
 
+  # -------------------------------------------------- async job status (app) ---
+  /api/v1/servers/{name}/jobs:
+    get:
+      tags: [backups]
+      operationId: listServerJobs
+      summary: Latest async world operations (backup/restore) for a server (owner-or-admin).
+      description: >-
+        Backup and restore run as cluster Jobs, so a 202 that later failed left
+        its only trace in the Job object. This route projects the newest such
+        Jobs, newest first, so failures are observable without kubectl. State is
+        "running" | "succeeded" | "failed".
+      x-felis-face: [external]
+      x-felis-tier: app
+      security: [{ accessJWT: [] }]
+      parameters:
+        - { name: name, in: path, required: true, schema: { type: string } }
+      responses:
+        '200':
+          description: The server's newest backup/restore jobs.
+          content:
+            application/json:
+              schema:
+                type: object
+                required: [server, jobs]
+                properties:
+                  server: { type: string }
+                  jobs:
+                    type: array
+                    items:
+                      type: object
+                      required: [name, kind, state]
+                      properties:
+                        name: { type: string }
+                        kind: { type: string, enum: [backup, restore] }
+                        state: { type: string, enum: [running, succeeded, failed] }
+                        message: { type: string }
+                        started_at: { type: string, format: date-time }
+                        finished_at: { type: string, format: date-time }
+        '401':
+          $ref: '#/components/responses/Unauthorized'
+        '403':
+          $ref: '#/components/responses/Forbidden'
+        '404':
+          description: Unknown server.
+          content:
+            application/json:
+              schema: { $ref: '#/components/schemas/Error' }
+        '503':
+          $ref: '#/components/responses/ServiceUnavailable'
+
   # ------------------------------------------------- server file editor (app) ---
   /api/v1/servers/{name}/files:
     get:
@@ -4147,18 +4266,30 @@ paths:
           $ref: '#/components/responses/BadRequest'
         '401':
           $ref: '#/components/responses/Unauthorized'
+        '403':
+          description: >-
+            The per-user submission allowance is spent — too many of the
+            caller's submissions are awaiting review, or their stored-upload
+            budget is full (submission_quota_exceeded).
+        '429':
+          description: >-
+            A submission was created within the per-user cooldown window
+            (submission_cooldown).
         '503':
           $ref: '#/components/responses/ServiceUnavailable'
     get:
       tags: [submissions]
       operationId: mySubmissions
-      summary: List the caller's own modpack submissions (user-directed lane over §16).
+      summary: List the caller's own modpack submissions with each linked build's outcome (user-directed lane over §16).
       x-felis-face: [external]
       x-felis-tier: app
       security: [{ accessJWT: [] }]
       responses:
         '200':
-          description: The caller's submissions, newest first.
+          description: >-
+            The caller's submissions, newest first; rows with a linked build
+            additionally carry build_status/build_error so the submitter can see
+            whether their build succeeded or failed (and why).
           content:
             application/json:
               schema:
@@ -4185,8 +4316,10 @@ paths:
         principal; a submission the caller does not own is reported as 404, so
         this endpoint cannot upload to or probe another user's submission. Only a
         pending_review submission accepts a context (409 otherwise); a wrong-format
-        or oversize body is rejected with 400. Returns 503 when the deployment's
+        or oversize body is rejected with 400, and an upload that would push the
-        context store has no implemented upload transport.
+        caller past their per-user stored-context budget is refused with 403
+        before the excess is persisted. Returns 503 when the deployment's context
+        store has no implemented upload transport.
       x-felis-face: [external]
       x-felis-tier: app
       security: [{ accessJWT: [] }]
@@ -4207,10 +4340,53 @@ paths:
           $ref: '#/components/responses/BadRequest'
         '401':
           $ref: '#/components/responses/Unauthorized'
+        '403':
+          description: >-
+            The upload would exceed the caller's per-user stored-context budget
+            (submission_quota_exceeded).
         '404':
           $ref: '#/components/responses/NotFound'
         '409':
           $ref: '#/components/responses/Conflict'
+        '429':
+          description: >-
+            An upload was accepted within the per-user cooldown window
+            (submission_cooldown).
+        '503':
+          $ref: '#/components/responses/ServiceUnavailable'
+
+  /api/v1/me/submissions/{id}:
+    delete:
+      tags: [submissions]
+      operationId: withdrawSubmission
+      summary: Withdraw your own pending submission (user side; user-directed lane over §16).
+      description: >-
+        Retracts the caller's own submission while it is still pending review:
+        the row and its uploaded build context are deleted, freeing the pending
+        slot and the per-user storage budget for a fresh submission. A reviewed
+        submission is frozen (409 — its build may already be consuming the
+        context), and a submission the caller does not own reads back as 404, so
+        this endpoint cannot probe or clear another user's uploads.
+      x-felis-face: [external]
+      x-felis-tier: app
+      security: [{ accessJWT: [] }]
+      parameters:
+        - { name: id, in: path, required: true, schema: { type: string } }
+      responses:
+        '200':
+          description: The withdrawn submission, as it was before the deletion.
+          content:
+            application/json:
+              schema: { $ref: '#/components/schemas/Submission' }
+        '401':
+          $ref: '#/components/responses/Unauthorized'
+        '404':
+          $ref: '#/components/responses/NotFound'
+        '409':
+          description: Submission has already been reviewed and cannot be withdrawn.
+          content:
+            application/json:
+              schema: { $ref: '#/components/schemas/Error' }
         '503':
           $ref: '#/components/responses/ServiceUnavailable'
 
@@ -4285,10 +4461,24 @@ paths:
               type: object
               required: [image_ref, dockerfile, context_ref]
               properties:
-                image_ref: { type: string }
+                image_ref:
-                dockerfile: { type: string }
+                  type: string
-                context_ref: { type: string }
+                  description: Push target under the internal registry (e.g. registry.felis.svc:5000/foo:1.0).
-                base_image: { type: string }
+                dockerfile:
+                  type: string
+                  description: >-
+                    Audit archive of the recipe, recorded on the build row and shown in the
+                    panel — the executed Dockerfile is the file named `Dockerfile` at the
+                    root of the context tarball (Kaniko runs --dockerfile=Dockerfile), so
+                    this field is never executed.
+                context_ref:
+                  type: string
+                  description: >-
+                    Location of the uploaded gzip build context; its root must contain the
+                    Dockerfile that gets executed.
+                base_image:
+                  type: string
+                  description: Resolved FROM, recorded for audit only — not a build gate.
       responses:
         '202':
           description: Build accepted.
@@ -4522,6 +4712,39 @@ paths:
         '503':
           $ref: '#/components/responses/ServiceUnavailable'
 
+  /api/v1/submissions/{id}/context:
+    get:
+      tags: [submissions]
+      operationId: downloadSubmissionContext
+      summary: Download a submission's uploaded build context (admin; user-directed lane over §16).
+      description: >-
+        The reviewer's read path to the artifact they are about to approve: the
+        executed Dockerfile lives inside this tarball (Kaniko runs the context's
+        root `Dockerfile`), so without it the human gate would be blind. Streams
+        the stored context.tar.gz verbatim with an attachment disposition — the
+        same bytes the build Pod fetches over the internal face. 404 when the
+        submission is unknown or has no uploaded context; 503 when the
+        deployment's context store has no implemented transport.
+      x-felis-face: [external]
+      x-felis-tier: admin
+      security: [{ accessJWT: [] }]
+      parameters:
+        - { name: id, in: path, required: true, schema: { type: string } }
+      responses:
+        '200':
+          description: The stored build context (gzip tarball), served as an attachment.
+          content:
+            application/gzip:
+              schema: { type: string, format: binary }
+        '401':
+          $ref: '#/components/responses/Unauthorized'
+        '403':
+          $ref: '#/components/responses/Forbidden'
+        '404':
+          $ref: '#/components/responses/NotFound'
+        '503':
+          $ref: '#/components/responses/ServiceUnavailable'
+
   /api/v1/submissions/{id}/reject:
     post:
       tags: [submissions]
@@ -4562,3 +4785,35 @@ paths:
               schema: { $ref: '#/components/schemas/Error' }
         '503':
           $ref: '#/components/responses/ServiceUnavailable'
+
+  /api/v1/submissions/{id}:
+    delete:
+      tags: [submissions]
+      operationId: deleteSubmission
+      summary: Retire a submission outright — row and uploaded context (admin; user-directed lane over §16).
+      description: >-
+        Removes the submission and its uploaded build context, any status — the
+        lane's only lifecycle valve, and the path that reclaims a rejected or
+        consumed upload from the uploads PVC. The reviewer identity is recorded
+        in the audit event, not on the (now deleted) row. Deleting an approved
+        submission whose build is still running fails that build's context
+        fetch; the admin has explicitly chosen to retire the artifact.
+      x-felis-face: [external]
+      x-felis-tier: admin
+      security: [{ accessJWT: [] }]
+      parameters:
+        - { name: id, in: path, required: true, schema: { type: string } }
+      responses:
+        '200':
+          description: The deleted submission, as it was before the deletion.
+          content:
+            application/json:
+              schema: { $ref: '#/components/schemas/Submission' }
+        '401':
+          $ref: '#/components/responses/Unauthorized'
+        '403':
+          $ref: '#/components/responses/Forbidden'
+        '404':
+          $ref: '#/components/responses/NotFound'
+        '503':
+          $ref: '#/components/responses/ServiceUnavailable'

docs/sequence-diagrams.md

@@ -87,7 +87,7 @@ sequenceDiagram
         else quota available
             Repo-->>API: true
             API->>Repo: ClaimServer(name, user_id)
-            Note over Repo: SELECT EXISTS(server); then atomic UPDATE servers SET owner_id=$2, claimed_at=now() WHERE name=$1 AND owner_id IS NULL AND deleted_at IS NULL
+            Note over Repo: ONE transaction: pg_advisory_xact_lock(user_id) serializes this user's claim lane; SELECT FROM servers WHERE name=$1 AND deleted_at IS NULL FOR UPDATE; re-run the four-dimension quota gate (authoritative — the pre-check above is a fast path); then UPDATE servers SET owner_id=$2, claimed_at=now() WHERE name=$1 AND owner_id IS NULL AND deleted_at IS NULL
             alt server missing
                 Repo-->>API: ErrNotFound
                 API-->>Panel: 404 not_found
@@ -137,14 +137,14 @@ sequenceDiagram
     Panel->>APIExternal: POST /api/v1/account/link/verify {code}
     APIExternal->>APIExternal: trim and uppercase code
     APIExternal->>Repo: VerifyLinkCode(user_id, code, now)
-    Repo->>Repo: SELECT non-expired code
+    Repo->>Repo: SELECT mc_uuid, auth_source FROM account_link_codes WHERE code=$1 AND expires_at>$2 FOR UPDATE
     alt missing or expired code
         Repo-->>APIExternal: ErrLinkCodeInvalid
         APIExternal-->>Panel: 400 invalid_code
-    else UUID linked to another user
+    else UUID linked to a different, live user
         Repo-->>APIExternal: ErrConflict
         APIExternal-->>Panel: 409 already_linked
-    else valid code
+    else valid code (re-verify by the same user is idempotent; a retired/soft-deleted owner's link is taken over)
         Repo->>Repo: INSERT account_links(user_id, mc_uuid, auth_source) ON CONFLICT (user_id, mc_uuid) DO UPDATE auth_source
         Repo->>Repo: DELETE account_link_codes WHERE code=$1
         Repo-->>APIExternal: mc_uuid, auth_source

docs/troubleshooting.md

@@ -20,8 +20,11 @@ graded for how far the in-repo Go test suite proves the behaviour:
   containerd, Postgres, or the network, not by Felis Go code; you will see it in
   `kubectl describe` / pod logs, never in `MinecraftServer.status`.
 - **[INERT]** — the configuration field exists in the CRD but no controller
-  reads it. Tuning it does nothing. §12 lists the one field this still applies
+  reads it. Tuning it does nothing. **No CRD field carries this status today**;
-  to, alongside the fields that *are* read and the condition each depends on.
+  the last one, `spec.storage.retainOnDelete`, was removed rather than
+  implemented (§13 records why). A field whose change seems ignored is almost
+  always a condition instead — §12 lists the fields that *are* read and the
+  condition each depends on.
 
 The operator never invents the parent domain; routing identity is
 `spec.subdomain` under the deployment zone. Examples below use
@@ -391,6 +394,95 @@ Inspect:
 kubectl logs -n felis-build job/<build-job>
 ```
 
+### 8e. Build Pods never start: executor images and air-gapped installs
+
+The build Job runs Kaniko and Trivy from external registries by default
+(`gcr.io/kaniko-project/executor:latest`, `aquasec/trivy:latest`). On a box whose
+build namespace cannot reach those registries (the egress policy allows only
+DNS, the internal registry and `--package-cidr` mirrors — and an air-gapped box
+has no route at all), the Pods sit in `ImagePullBackOff`/`ErrImagePull` and the
+build stays `building` until its deadline. Point the overrides at images **in
+the internal registry** — the one pull source that survives an image GC (a bare
+node-containerd import does not: kubelet's image GC collects unused images under
+disk pressure, and an air-gapped box then has nothing to restore them from) —
+in `felis.toml`:
+
+```toml
+[registry]
+url = "registry.felis.svc:5000"
+build_namespace = "felis-build"
+kaniko_image = "registry.felis.svc:5000/mirror/kaniko-executor:v1.24.0"
+trivy_image  = "registry.felis.svc:5000/mirror/trivy:0.74.0"
+trivy_db_repository = "registry.felis.svc:5000/mirror/trivy-db:2"
+trivy_java_db_repository = "registry.felis.svc:5000/mirror/trivy-java-db:1"
+build_cpu_limit = "2"
+build_mem_limit = "4Gi"
+```
+
+Mirror the executor images into the registry once. On the node itself, push
+through the loopback hostPort the registry Deployment binds (docker treats
+`127.0.0.1` as insecure by default; the installer leaves the daemon stopped, so
+`sudo systemctl start docker` first):
+
+```sh
+docker pull gcr.io/kaniko-project/executor:v1.24.0   # any versions you pin
+docker pull aquasec/trivy:0.74.0
+docker pull mirror.gcr.io/aquasec/trivy-java-db:1
+docker tag gcr.io/kaniko-project/executor:v1.24.0 127.0.0.1:5000/mirror/kaniko-executor:v1.24.0
+docker tag aquasec/trivy:0.74.0                   127.0.0.1:5000/mirror/trivy:0.74.0
+docker tag mirror.gcr.io/aquasec/trivy-java-db:1  127.0.0.1:5000/mirror/trivy-java-db:1
+docker push 127.0.0.1:5000/mirror/kaniko-executor:v1.24.0
+docker push 127.0.0.1:5000/mirror/trivy:0.74.0
+docker push 127.0.0.1:5000/mirror/trivy-java-db:1
+```
+
+From another machine, port-forward the registry instead (`kubectl -n felis
+port-forward svc/registry 5000:5000`) and push to `localhost:5000/...` — the
+registry keys a repository by the path after the host, so pushes through either
+door land in the same place the build Pods will pull from.
+
+Put them in **both** `/etc/felis/felis.host.toml` (host-side CLI) and
+`/etc/felis/felis.pod.toml` (the file rendered into the API's `felis-config`
+Secret — the two differ only in the database URL; the setup screens re-render
+the Secret from the pod file, so edits made only through `kubectl` on the live
+Secret are lost at the next reconfigure). A Deployment restart alone is NOT
+enough — the API Pod mounts the Secret, never the host file. Re-render the
+Secret from the pod file, then roll `felis-api`:
+
+```sh
+kubectl -n felis create secret generic felis-config \
+  --from-file=felis.toml=/etc/felis/felis.pod.toml --dry-run=client -o yaml | kubectl apply -f -
+kubectl -n felis rollout restart deployment/felis-api
+```
+
+Unset fields keep the defaults.
+
+`trivy_db_repository` is not optional on an egress-locked box. Trivy fetches its
+vulnerability DB from `mirror.gcr.io`/`ghcr.io` unless told otherwise, and the
+build egress policy denies those hosts — so the scan step fails closed
+(`failed to download vulnerability DB`) and NO build ever completes, even though
+Kaniko pushed the image. Mirror the DB into the internal registry once:
+
+```
+# On the node (docker treats 127.0.0.1 as insecure by default), or through the
+# port-forward above:
+#   docker pull mirror.gcr.io/aquasec/trivy-db:2
+#   docker tag  mirror.gcr.io/aquasec/trivy-db:2 127.0.0.1:5000/mirror/trivy-db:2
+#   docker push 127.0.0.1:5000/mirror/trivy-db:2
+```
+
+The Job's Trivy container already runs with `--insecure`, so the internal
+registry's plain HTTP works for the DB pull exactly as it does for the scanned
+image. Re-mirror the tag periodically (Trivy refreshes the DB several times a
+day upstream; a stale mirror only means stale CVE data, never a failed gate).
+
+`trivy_java_db_repository` is the same story one step lazier: Trivy downloads
+the Java DB on demand the first time it scans an image containing Java
+artifacts — every real modpack — and that download fails closed too. Mirror
+`mirror.gcr.io/aquasec/trivy-java-db:1` alongside the vulnerability DB (commands
+above); the Java DB refreshes far less often than the vulnerability DB, so a
+one-off mirror is usually fine.
+
 ---
 
 ## 9. Registry push/pull failures (spec §15)
@@ -408,6 +500,16 @@ control namespace (or `--registry-namespace`):
 - **Storage:** PVC is RWO, `10Gi`, mounted at `/var/lib/registry`, **no
   `storageClassName`** → binds the cluster default class. If the cluster has no
   default StorageClass the PVC stays `Pending` and the registry never starts.
+- **Node-side pulls:** containerd cannot dial the Service VIP (the live stack
+  answered "Empty reply"), so the registry Deployment binds a loopback hostPort
+  (`127.0.0.1:<port>`) and the installer writes a `/etc/rancher/k3s/registries.yaml`
+  mirror relaying `registry.<ns>.svc:<port>` onto it. That pair is what lets
+  kubelet re-pull a garbage-collected image; both halves must survive together
+  (remove either and every pull after an image GC fails).
+- **Memory:** the registry's limit is 2Gi, deliberately larger than the other
+  control-plane pods' 256Mi — a live 475MB-layer push OOM-killed the 256Mi
+  template mid-upload (audit #46). Very large layers need headroom here, not
+  more CPU.
 - **Selector quirk worth knowing:** the registry Service selector is only
   `name + component=registry` — it deliberately lacks the
   `part-of=felis-control-plane` label, so the registry is *invisible* to the
@@ -426,6 +528,16 @@ reaps a world only when `now - last_active_at > 15d` (`inactive_15d`); the 15-da
 deadline is **hard-fixed in code** (only `warn_before` / `retention` /
 `max_local_bytes` are configurable from `felis.toml [archive]`).
 
+### What a "backup" contains
+
+A backup tars the server's ENTIRE data volume — the same volume the server mounts
+at `/data`: world folders, `server.properties`, plugins/mods, configs, jars,
+libraries, logs and cache, not just the `world/` directory. A restore replaces the
+volume's contents with the archive (files added since the backup are pruned), so a
+restore also rolls config/plugin changes back. Sizes are dominated by
+libraries/cache on stock Paper servers (~170MB for a fresh instance before any
+world growth) — do not size the archive PVC as if only world data were stored.
+
 ### The backup-before-delete invariant
 
 The reap sequence (all [GO-TESTED] hermetically) preserves the world unless a
@@ -449,6 +561,29 @@ So a missing backup never results in a deleted world. [GO-TESTED:
 - CRD missing → logs `reaper: CRD missing, skipping`, skipped.
 - Idle `≤ 15d` → not yet eligible.
 
+### Pre-reap warnings (the `warn_before` offsets)
+
+An OWNED server inside a warning window gets an email notice (`3d`/`1d` before
+the deadline, `warn_before` from `[archive]`) to the owner's **verified** email —
+the same `[smtp]` relay felis-api uses. The `warned_3d_at` / `warned_1d_at`
+stamps record a **delivered** notice:
+
+- No `[smtp]` configured (or owner has no verified address): the run logs
+  `reaper: warning suppressed — no warner wired` / a delivery error and does
+  NOT stamp. Nothing is falsely recorded as sent, and the day SMTP is
+  configured the pending warning can still go out.
+- Delivery failure (relay down): logged and retried on the next daily run —
+  bounded by the warning window, since the reap removes the candidate anyway.
+- `warned=` in the run output counts DELIVERED notices, not attempts.
+
+The reaper runs in the minecraft namespace and reads the **mirrors** of
+`felis-smtp` and `felis-config` there (a `secretKeyRef` is namespace-local). The
+installed `felis setup`'s "configure email" screen refreshes both mirrors when it
+applies, so configuring SMTP after install is enough; a manual edit of the
+control-namespace Secret alone is not. [GO-TESTED: the delivered/retried/
+suppressed matrix in `internal/reaper`; live-drilled end to end against a local
+SMTP sink.]
+
 ### Genuine false-delete risk vectors
 
 - **Stale `last_active_at`.** The keep-alive is `RecordJoin`, called from the
@@ -468,6 +603,32 @@ Only `TarLocal` (tar+gzip) archiving is implemented; VolumeSnapshot/Longhorn
 backends return `not implemented in this build`. The live PVC delete / Postgres
 store paths are [INTEGRATION-ONLY].
 
+### Where worlds are read from (hostPath resolution)
+
+The CronJob mounts `--worlds-host-path` read-only at `/worlds`; the resolver
+runs `cmd/felis/reaper.resolveWorldDir`: it looks for `<root>/<pvc>`, then for
+the stock local-path directory `<root>/<pv-name>_<ns>_<pvc-name>` derived from
+the live PVC's `spec.volumeName` (never a glob — a leftover directory of a
+deleted PV must not stand in for the world the PVC currently binds). Pointing
+the flag at k3s's storage root (`/var/lib/rancher/k3s/storage`) is therefore the
+supported way to enable retention on a stock install. Two deployment facts the
+resolver cannot fix:
+
+- **Permissions.** The reaper Pod runs as **root** and carries `DAC_OVERRIDE`:
+  worlds are written by the game image's own UID (root for every Paper image we
+  ship), and Paper saves `level.dat` mode-0600, so any fixed non-root identity
+  (the previous uid-1000 convention, and the ACL setup that went with it) could
+  neither walk the tree nor read the files — every archive failed
+  `open …/level.dat: permission denied` and the same defect failed on-demand
+  backups/restores. Root is the same identity the game container itself runs as
+  (see the operator's forwarding-init note); `DAC_OVERRIDE` extends the archive
+  to game images with a different UID. If a world is still **preserved** while a
+  reap was expected, it is now a different cause: check the run's ERROR logs for
+  the resolver's `lstat` messages before suspecting permissions.
+- **Node placement.** Multi-node clusters: the world's directory exists only on
+  the node holding its volume, and the CronJob sets no `nodeSelector`, so add
+  one (single-node starters are pinned implicitly).
+
 ---
 
 ## 11. Idle auto-stop never fires; player count always shows 0
@@ -502,6 +663,32 @@ Both fields set and still nothing happens? Then the probe is failing rather than
 disabled: the server would be stuck in `Starting` with `RconNotReachable`
 (`reconciler.go:156`), which is §1's symptom, not this one.
 
+This path used to fail even with everything configured correctly, through three
+stacked defects proven and fixed on a live cluster (auditfix21/22): the
+`emptySince` stamp was pruned by a missing CRD status field, a quiescent empty
+server produced no watch events to re-check the timer, and the Role lacked the
+`minecraftservers:patch` grant the stop write needs. If auto-stop ever looks
+dead again, check these three in order (each is now pinned by a test):
+
+```sh
+# ① The stamp must persist — should print a timestamp, not an empty string,
+#    a few seconds after a server goes Ready with zero players.
+kubectl get minecraftserver <name> -o jsonpath='{.status.emptySince}'
+
+# ② The operator must be able to write spec.desiredState (403 in the operator
+#    log = missing patch grant on Role felis-operator).
+kubectl auth can-i patch minecraftservers -n <ns> --as=system:serviceaccount:<ctl-ns>:felis-operator
+
+# ③ A wake-up must be scheduled: while empty, expect whatever you set
+#    as emptySecondsBeforeStop to elapse and the box to flip to Stopped without
+#    any external action.
+```
+
+While players are online the operator re-probes on a 30s cadence so it notices
+the moment the last one leaves; while empty it schedules a wake-up exactly at
+the deadline. Quiet operator logs on an idle server are normal — the action is
+the scheduled wake-up, not a stream of reconciles.
+
 Note the reaper's `last_active_at` (§10) is a *different* subsystem (Postgres
 business layer, bumped by join events) — it keeps worlds alive against the
 reaper, but it does **not** auto-stop empty running servers.
@@ -526,6 +713,30 @@ deadline for the whole start, applied on the RCON-probe branch.
 
 ---
 
+## 12b. A newer field never reaches an already-installed system server (`felis converge`)
+
+Provisioning is create-if-absent: `felis setup` never rewrites an existing
+`login`/`lobby` `MinecraftServer` beyond the config-derived env it owns, so a
+field the desired spec gained after your install sits absent forever — this is
+how a deployment ends up with a lobby that has no `spec.rcon` (a dead console
+and an online-player count that is always 0) and a login gate without
+`spec.startup.healthHTTPPort`. `felis converge` is the explicit pass that fills
+exactly those zero-valued fields (and re-adds a derived env key that is
+missing). It never overwrites a value that already holds one — an operator's
+RCON secretRef or tuning survives.
+
+```
+sudo felis converge
+```
+
+Run it **after the images are in place**. Enabling RCON, or the HTTP readiness
+gate, on a server whose image predates the listener would hold that server in
+`Starting` until the operator marks it `Failed` — that ordering is the reason
+this is a command you run rather than something setup does on every re-run.
+System servers that are already current report `already converged`.
+
+---
+
 ## 13. World PVC survives after I deleted the MinecraftServer
 
 This is expected. The world PVC is a StatefulSet `VolumeClaimTemplate`. There is
@@ -557,6 +768,69 @@ changes, because retention was never conditional in the first place.
 
 ---
 
+## 13b. Node runs out of disk: what survives, and how to recover
+
+A full disk is the most destructive failure this stack sees: kubelet evicts game
+pods (the control plane is protected below), and its image GC then collects
+images nothing is running. The images have a pull source now — the in-cluster
+registry — so they come back without an operator re-import; freeing space is
+what completes the recovery.
+
+**Eviction.** Every control-plane pod (api, operator, reaper, registry) runs
+under the BUILT-IN `system-cluster-critical` PriorityClass (value 2e9). Kubelet's
+node-pressure eviction refuses to touch those pods — the log shows
+*"Eviction manager: cannot evict a critical pod"* for each of them — while
+game-server pods at the default priority 0 are evicted first. A drill that filled
+the disk to 1.7G free saw exactly this: login/lobby evicted, the whole control
+plane still Running (before the fix the same drill evicted the api, operator and
+registry too, and the image-GC stage below followed). User-defined
+PriorityClasses cannot substitute: the API caps them at 1e9, below kubelet's
+critical threshold. The built-in class allows preemption (its policy is fixed),
+so a control-plane pod that cannot fit may preempt a game pod — deliberate: the
+management plane must be placeable.
+
+**The pressure condition clears slowly.** After you free space, the node can stay
+`DiskPressure:True` for up to ~5 minutes
+(`--eviction-pressure-transition-period` defaults to 5m, to stop the condition
+flapping); pods that need scheduling wait for it. This is the bulk of the
+"recovery takes minutes" observation, not a stuck node.
+
+**The images may be gone — they come back on their own.** If pods were evicted,
+the kubelet can garbage-collect their images (unused > 2 minutes under imagefs
+pressure). Every image this platform runs is ALSO hosted in the in-cluster
+registry: the installer builds each one as `registry.<ns>.svc:5000/felis/…`
+and mirrors it there, and the node's containerd is configured (a registries.yaml
+mirror onto the registry's loopback hostPort) to relay those refs back through
+it. So a GC'd image is re-pulled on the next attempt with no operator action —
+delete the stuck pod to force an immediate retry (or wait out the backoff), and
+the workload converges.
+
+If a pull does NOT come back:
+
+1. Free disk on the node (`df -h /var/lib/rancher`; the biggest consumers are
+   `k3s ctr images ls -q` and the world/backup PVCs under
+   `/var/lib/rancher/k3s/storage`).
+2. Check the registry: `kubectl -n felis get pods -l
+   app.kubernetes.io/component=registry` and, on the node,
+   `curl -s http://127.0.0.1:5000/v2/` (expect `{}`).
+3. Check the mirror file: `/etc/rancher/k3s/registries.yaml` must map
+   `registry.felis.svc:5000` to `http://127.0.0.1:5000`. Missing or changed:
+   re-run the installer (it rewrites the file and restarts k3s only when the
+   content changed).
+4. Re-mirror a tag the registry does not have (hand-built images were never
+   pushed): `sudo systemctl start docker` (the installer leaves the daemon
+   stopped), then `docker tag <ref> 127.0.0.1:5000/<repo>:<tag> && docker push
+   127.0.0.1:5000/<repo>:<tag>`.
+
+For an image that is in neither place, the old fallback still stands: re-run the
+installer (it rebuilds/re-imports from the local Docker store AND mirrors into
+the registry), or for a single image
+`docker save felis:<tag> | k3s ctr images import -`. The Docker store remains a
+deliberate second copy on the node; treat it as the recovery path, not as free
+space.
+
+---
+
 ## 14. Metrics for diagnosis (spec §23)
 
 All four mandated metrics have real producers; scrape them when triaging:
@@ -571,8 +845,65 @@ All four mandated metrics have real producers; scrape them when triaging:
   actually deleted post-backup (§10); a spike here means worlds crossed the 15d
   idle line — cross-check that join events are flowing (§10 risk vectors).
 
+### Scraping
+
+The series come from two processes:
+
+- `felis-operator` pod `:8080/metrics` — `felis_servers_total`,
+  `felis_start_duration_seconds` (no Service; scrape pod-scoped, e.g. a
+  PodMonitor targeting port `metrics`).
+- `felis-api` internal face `:8081/metrics` (Service `felis-api-internal`) —
+  `felis_image_build_failures_total`. Unauthenticated like the probes;
+  ClusterIP-only, and the external face never serves it.
+- `felis_reaper_worlds_deleted_total` is produced inside the one-shot reaper
+  CronJob, which exits long before any scrape interval — without a pushgateway
+  it has no scrape path. Read the reaper Pod log or the `world_backups` table
+  for deletions instead.
+
+### Alert rules
+
+`deploy/alerts/` ships ready-made rules: build failures, slow starts, node
+disk/memory thresholds, and the kubelet `DiskPressure` condition.
+
+- Plain Prometheus: add `felis-alerts.yaml` to `rule_files`. Check and unit-test
+  it standalone with `promtool check rules felis-alerts.yaml` and
+  `promtool test rules felis-alerts_test.yml` (the tests pin exactly when each
+  alert fires).
+- kube-prometheus-stack / prometheus-operator: `kubectl apply -f
+  felis-prometheusrule.yaml` (adjust its `release:` label to your stack's
+  ruleSelector).
+
 ---
 
+## 15. Control-plane upgrades, and rolling back a bad one
+
+There is no in-place updater: an upgrade is re-running the installer
+(`curl -fsSL <installer URL> | sudo bash`), which rebuilds/re-imports the image
+and re-applies the bundle. (`sudo felis setup` is not this path; on a completed
+install it only opens the config console.) The channel is not persisted across
+the re-run, so pass `FELIS_VERSION_BOOTSTRAP=dev` on a host that tracks main.
+Two properties of the control plane matter when you do:
+
+- Both Deployments use strategy **Recreate** (single replica, no leader election:
+  two overlapping instances would fight over the same cluster). An upgrade takes
+  the panel/API down for the rollout window — seconds normally, longer if the new
+  image still has to be imported.
+- If the new pod cannot start (bad tag, missing image), the installer's rollout
+  wait fails after 180s and prints `kubectl describe` diagnostics: you see
+  `ErrImagePull`/`ImagePullBackOff` there instead of a silent hang.
+
+Roll back with:
+
+```
+kubectl -n felis rollout undo deploy/felis-api
+kubectl -n felis rollout status deploy/felis-api
+```
+
+(the same for `felis-operator` and `registry`). `rollout undo` returns to the
+previous ReplicaSet, whose image is normally still on the node; if the image GC
+collected it, the registry re-serves it automatically (§13b) for every tag the
+installer built — only hand-built tags need a manual re-mirror.
+
 ## Quick reference: symptom → section
 
 | Symptom | Section |
@@ -588,10 +919,12 @@ All four mandated metrics have real producers; scrape them when triaging:
 | Local password login rejected | §5c |
 | Internal callers 401 (service token) | §6 |
 | Link/claim 400/409/412/403/404 | §7 |
-| Build push 400 / SA denied / egress hang / Failed | §8 |
+| Build push 400 / SA denied / egress hang / Failed / executor ImagePullBackOff | §8, §8e |
 | Registry push/pull unreachable | §9 |
 | World deleted unexpectedly / backup skipped | §10 |
 | Idle auto-stop not firing; player count 0 | §11 |
 | A config field seems ignored | §12 |
 | PVC left behind after delete | §13 |
+| Node out of disk; pods evicted / ImagePullBackOff | §13b |
 | Which metric to scrape | §14 |
+| Upgrade / roll back a bad control-plane image | §15 |

go.mod

@@ -9,6 +9,7 @@ require (
 	github.com/charmbracelet/huh v1.0.0
 	github.com/charmbracelet/lipgloss v1.1.0
 	github.com/descope/virtualwebauthn v1.0.5
+	github.com/go-logr/logr v1.4.2
 	github.com/go-webauthn/webauthn v0.17.4
 	github.com/golang-jwt/jwt/v5 v5.3.1
 	github.com/google/uuid v1.6.0
@@ -42,7 +43,6 @@ require (
 	github.com/erikgeiser/coninput v0.0.0-20211004153227-1c3628e74d0f // indirect
 	github.com/evanphx/json-patch/v5 v5.9.0 // indirect
 	github.com/fxamacker/cbor/v2 v2.9.2 // indirect
-	github.com/go-logr/logr v1.4.2 // indirect
 	github.com/go-openapi/jsonpointer v0.19.6 // indirect
 	github.com/go-openapi/jsonreference v0.20.2 // indirect
 	github.com/go-openapi/swag v0.22.4 // indirect

internal/api/api.go

@@ -63,6 +63,11 @@ type API struct {
 	// authorization boundary is exercised before the backup-Job executor is wired.
 	Backuper Backuper
 
+	// JobStatus reads the latest backup/restore Job outcomes for GET
+	// /servers/{name}/jobs — the status outlet for async failures (the enqueue
+	// endpoints only answer 202). Optional: nil → that route reports 503.
+	JobStatus JobStatusReader
+
 	// Files is the server file editor (list / read / write a file in a stopped
 	// server's world volume — the "one wrong line in server.properties" repair).
 	// Like Restorer and Backuper it is optional: when nil the file routes report
@@ -118,6 +123,15 @@ type API struct {
 	// on the wake lever). Zero disables throttling.
 	WakeCooldown time.Duration
 
+	// SubmitCreateCooldown / SubmitUploadCooldown throttle the user-modpack
+	// submission lane per user: create bounds how quickly review-queue rows can
+	// appear, upload bounds how often a user may stream a (up to 1 GiB) build
+	// context. The keys are separate, so the lane's normal shape — create, then
+	// upload — is never blocked by its own throttle. Zero disables each lever
+	// (the same idiom as WakeCooldown); cmd/felis wires positive values.
+	SubmitCreateCooldown time.Duration
+	SubmitUploadCooldown time.Duration
+
 	// MaxRunningServers caps how many servers may be desired-Running cluster-wide
 	// (spec §9.1: the concurrency-上限 lever hanging on the same wake chokepoint as
 	// cooldown and autostartPolicy). Zero — the default — disables it: §9.2 wires
@@ -153,6 +167,9 @@ type API struct {
 	otpCooldownOnce sync.Once
 	otpCooldown     *cooldownLimiter
 
+	submitCooldownOnce sync.Once
+	submitCooldown     *cooldownLimiter
+
 	streamCapOnce sync.Once
 	streamCap     *streamLimiter
 }
@@ -199,6 +216,20 @@ func (a *API) otpLimiter() *cooldownLimiter {
 	return a.otpCooldown
 }
 
+// submitLimiter lazily builds a SEPARATE cooldown limiter for the user-modpack
+// submission lane, so its throttles never share state with the wake or OTP
+// keyspaces. One limiter backs both levers with prefixed keys (see the
+// submissionCreateKey/UploadKey constants), so create and upload never contend
+// with each other. Like the other cooldowns it is process-local; with multiple
+// api replicas the effective spacing is per-replica, the same accepted
+// KNOWN-LIMITATION the OTP resend throttle carries.
+func (a *API) submitLimiter() *cooldownLimiter {
+	a.submitCooldownOnce.Do(func() {
+		a.submitCooldown = &cooldownLimiter{now: a.now, last: map[string]time.Time{}}
+	})
+	return a.submitCooldown
+}
+
 // streamGate lazily builds the per-principal SSE stream cap bound to
 // MaxStreamsPerPrincipal. A zero cap yields a disabled limiter that admits every
 // stream, so a deployment (or test) that leaves it unset pays nothing.
@@ -259,13 +290,20 @@ type apiRoute struct {
 }
 
 // internalAPIRoutes is the internal face's served route table (spec §7, §14):
-// service-token auth, never Zero Trust. It carries both health probes.
+// service-token auth, never Zero Trust. It carries both health probes and the
+// metrics scrape.
 func (a *API) internalAPIRoutes() []apiRoute {
 	return []apiRoute{
 		{Method: "GET", Pattern: "/healthz", Public: true, h: a.handleHealthz},
 		{Method: "GET", Pattern: "/readyz", Public: true, h: a.handleReadyz},
+		// Prometheus scrape (felis_* collectors); public because a scrape carries
+		// no token, internal-only so it is never exposed off-cluster.
+		{Method: "GET", Pattern: "/metrics", Public: true, h: a.handleMetrics},
 
 		{Method: "GET", Pattern: "/api/v1/servers", h: a.handleListServers},
+		// The build Pod's context-fetch initContainer streams a submission's stored
+		// modpack through this route (build namespace cannot mount the uploads PVC).
+		{Method: "GET", Pattern: "/api/v1/internal/submissions/{id}/context", h: a.handleInternalSubmissionContext},
 		{Method: "POST", Pattern: "/api/v1/internal/servers/{name}/ready", h: a.handleReady},
 		{Method: "POST", Pattern: "/api/v1/internal/servers/{name}/join-event", h: a.handleJoinEvent},
 		// Domain-autostart (spec §9.1, §14): velocity drives the wake lever and polls
@@ -407,6 +445,7 @@ func (a *API) externalAPIRoutes() []apiRoute {
 		// owned), and restore is gated by owner-or-admin PLUS a former-owner match, so
 		// neither sits behind adminOnly.
 		{Method: "GET", Pattern: "/api/v1/backups", h: a.handleListBackups},
+		{Method: "GET", Pattern: "/api/v1/servers/{name}/jobs", h: a.handleServerJobs},
 		{Method: "POST", Pattern: "/api/v1/servers/{name}/restore-backup", h: a.handleRestoreBackup},
 		{Method: "POST", Pattern: "/api/v1/servers/{name}/backup", h: a.handleBackupNow},
 		// Server file editor: list / read / write a file in a STOPPED server's world
@@ -479,6 +518,11 @@ func (a *API) externalAPIRoutes() []apiRoute {
 		// App-tier and owner-scoped (the id must belong to the principal), exactly
 		// like the create/list routes above.
 		{Method: "POST", Pattern: "/api/v1/me/submissions/{id}/context", h: a.handleUploadSubmissionContext},
+		// Withdraw the caller's OWN pending submission: the row and its uploaded
+		// context are deleted, freeing the pending slot and storage budget. Same
+		// owner-scoping as the upload route — a reviewed submission is frozen (409)
+		// and another user's id is invisible (404).
+		{Method: "DELETE", Pattern: "/api/v1/me/submissions/{id}", h: a.handleWithdrawSubmission},
 		// Admin (Zero-Trust) tier: create / mutate spec / image admission. These gate
 		// on Principal.IsAdmin() inside the handler via the adminOnly wrapper, so the
 		// boundary is exercised even where the body is a later-phase stub.
@@ -508,6 +552,13 @@ func (a *API) externalAPIRoutes() []apiRoute {
 		{Method: "GET", Pattern: "/api/v1/submissions", Admin: true, h: a.handleListSubmissions},
 		{Method: "POST", Pattern: "/api/v1/submissions/{id}/approve", Admin: true, h: a.handleApproveSubmission},
 		{Method: "POST", Pattern: "/api/v1/submissions/{id}/reject", Admin: true, h: a.handleRejectSubmission},
+		// Retire a submission outright (row + uploaded context), any status. The
+		// lane's lifecycle valve: without it, rejected/consumed uploads accumulated
+		// on the uploads PVC forever — there is no other delete path.
+		{Method: "DELETE", Pattern: "/api/v1/submissions/{id}", Admin: true, h: a.handleDeleteSubmission},
+		// The reviewer's read path to the uploaded blob: the executed Dockerfile
+		// lives inside it, so approval would otherwise be blind.
+		{Method: "GET", Pattern: "/api/v1/submissions/{id}/context", Admin: true, h: a.handleAdminSubmissionContext},
 		// Auto-update maintenance window (spec §B; decision core internal/updates).
 		// Admin-tier: it governs whether Felis may apply an update to itself, so setting
 		// it requires the admin Zero-Trust path, not a mere session. API+persistence

internal/api/api_test.go

@@ -38,8 +38,16 @@ type fakeRepo struct {
 	owners    map[string]string
 	ownersErr error
 	claimOK   map[string]bool // name -> claim succeeds; absent name -> ErrNotFound
-	audits    []AuditEntry
+	// claimQuotaRefuse simulates ClaimServer's atomic quota gate (audit #4)
-	joins     []string
+	// refusing a name whose advisory pre-check already passed.
+	claimQuotaRefuse map[string]bool
+	// serverResources / resourceUpdates mirror the cached resource columns:
+	// ServerResources is what the resize path reads (to preserve storage), and
+	// UpdateServerResources records the write for assertions.
+	serverResources map[string]ResourceSpec
+	resourceUpdates map[string]ResourceSpec
+	audits          []AuditEntry
+	joins           []string
 	// create-server seeding (spec §15)
 	seeded  map[string]bool   // name -> servers row exists
 	aliases map[string]string // subdomain -> bound server name
@@ -59,6 +67,10 @@ type fakeRepo struct {
 	staff    map[string]*StaffUser   // username -> staff login row
 	sessions map[string]*fakeSession // token_hash -> session
 	settings map[string][]byte       // key -> jsonb value
+	// failSessionUser / failGetSetting force those reads to fail with a generic
+	// (non-ErrNotFound) error, simulating a store outage for the 503 auth path.
+	failSessionUser error
+	failGetSetting  error
 	// player email OTPs (spec §B2). Keyed by row id; the verify path scans for the
 	// newest live (user, purpose) just as the PG query does.
 	otps map[string]*fakeEmailOTP
@@ -91,6 +103,10 @@ type fakeRepo struct {
 	// user admin fakes
 	seededUsers []seededUser
 	fakeQuotas  map[string]*QuotaView
+	// deletedIDs remembers soft-deleted user ids: DeleteUser drops the row from
+	// seededUsers (so listings hide it, mirroring the WHERE deleted_at IS NULL
+	// query), and this set keeps the account dead for the liveness guards.
+	deletedIDs map[string]bool
 	// pingErr, when non-nil, is returned by Ping to simulate DB liveness check
 	// failures in /readyz tests.
 	pingErr error
@@ -202,8 +218,9 @@ func newFakeRepo() *fakeRepo {
 		allowlist: map[string]map[string]bool{}, allowUUID: map[string]map[string]bool{},
 		mine:    map[string][]MyServerView{},
 		owners:  map[string]string{},
-		claimOK: map[string]bool{},
+		claimOK: map[string]bool{}, claimQuotaRefuse: map[string]bool{},
-		seeded:  map[string]bool{}, aliases: map[string]string{},
+		serverResources: map[string]ResourceSpec{}, resourceUpdates: map[string]ResourceSpec{},
+		seeded: map[string]bool{}, aliases: map[string]string{},
 		linkCodes: map[string]fakeLinkCode{}, links: map[string]string{},
 		linkAuthSource:    map[string]string{},
 		staff:             map[string]*StaffUser{},
@@ -220,6 +237,7 @@ func newFakeRepo() *fakeRepo {
 		discoverableChallenges: map[string]*fakeDiscoverableChallenge{},
 		fakeQuotas:             map[string]*QuotaView{},
 		migrations:             map[string]*fakeMigration{},
+		deletedIDs:             map[string]bool{},
 	}
 }
 
@@ -245,10 +263,13 @@ func (f *fakeRepo) QuotaCheck(_ context.Context, userID string, _ string, _ Reso
 	return f.QuotaAvailable(context.TODO(), userID)
 }
 
-func (f *fakeRepo) UpdateServerResources(_ context.Context, _ string, _, _, _ int) error { return nil }
+func (f *fakeRepo) UpdateServerResources(_ context.Context, name string, cpu, mem, stor int) error {
+	f.resourceUpdates[name] = ResourceSpec{CPUMilli: cpu, MemoryMB: mem, StorageMB: stor}
+	return nil
+}
 
-func (f *fakeRepo) ServerResources(_ context.Context, _ string) (ResourceSpec, error) {
+func (f *fakeRepo) ServerResources(_ context.Context, name string) (ResourceSpec, error) {
-	return ResourceSpec{}, nil
+	return f.serverResources[name], nil
 }
 func (f *fakeRepo) CreateLinkCode(_ context.Context, code, mcUUID, authSource string, expiresAt time.Time) error {
 	f.linkCodes[code] = fakeLinkCode{mcUUID: mcUUID, authSource: authSource, expiresAt: expiresAt}
@@ -266,7 +287,13 @@ func (f *fakeRepo) VerifyLinkCode(_ context.Context, userID, code string, now ti
 		return "", "", ErrLinkCodeInvalid
 	}
 	if existing, ok := f.links[rec.mcUUID]; ok && existing != userID {
-		return "", "", ErrConflict // do not consume another user's pending code
+		// A soft-deleted link's identity is unclaimed: the fresh in-game code lets a
+		// live caller take it over (mirrors PGRepo). Disabled-but-not-deleted stays a
+		// conflict — takeover there would bypass the lockout. Neither arm consumes
+		// the code.
+		if !f.seededDeleted(existing) {
+			return "", "", ErrConflict
+		}
 	}
 	f.links[rec.mcUUID] = userID
 	f.linkAuthSource[rec.mcUUID] = rec.authSource // copy/refresh, mirrors DO UPDATE
@@ -299,6 +326,9 @@ func (f *fakeRepo) RedeemPlayerBindCode(_ context.Context, newUserID, code strin
 				return "", "", "", ErrPlayerBindForbidden // staff must use op.console; do not consume
 			}
 		}
+		if f.seededDead(existing) {
+			return "", "", "", ErrPlayerAccountRetired // dead account; do not consume
+		}
 		delete(f.linkCodes, code)
 		return existing, rec.mcUUID, rec.authSource, nil
 	}
@@ -350,6 +380,13 @@ func (f *fakeRepo) VerifyEmailOTP(_ context.Context, userID, purpose, codeHash s
 		live.attempts++ // a typo costs an attempt but does not consume the code
 		return "", ErrOTPInvalid
 	}
+	// A DIFFERENT verified holder of the same address → ErrEmailTaken, code left
+	// live — mirrors PGRepo's guard + the users_verified_email_unique index.
+	for _, u := range f.staff {
+		if u.ID != userID && u.EmailVerified && strings.EqualFold(u.Email, live.email) {
+			return "", ErrEmailTaken
+		}
+	}
 	live.consumed = true
 	for _, u := range f.staff { // flip the user row verified (UPDATE users ...)
 		if u.ID == userID {
@@ -595,7 +632,7 @@ func (f *fakeRepo) UUIDInAllowlist(_ context.Context, n, uuid string) (bool, err
 	return f.allowUUID[n][uuid], nil
 }
 func (f *fakeRepo) UserByMCUUID(_ context.Context, uuid string) (string, error) {
-	if u, ok := f.links[uuid]; ok {
+	if u, ok := f.links[uuid]; ok && !f.seededDead(u) {
 		return u, nil
 	}
 	return "", ErrNotFound
@@ -619,15 +656,15 @@ func (f *fakeRepo) IsUsernameBlacklisted(_ context.Context, mcUUID string) (bool
 }
 
 // IsProtectedAdminLink mirrors PGRepo's JOIN of account_links to users: linked,
-// auth_source 'thirdparty', and the linked user an admin — no password-hash test, so
+// auth_source 'thirdparty', and the linked user staff (admin OR owner) — no
-// an SSO Operator (role='admin', with no password) is protected like any other.
+// password-hash test, so an SSO Operator or the Owner is protected like any other.
 func (f *fakeRepo) IsProtectedAdminLink(_ context.Context, mcUUID string) (bool, error) {
 	userID, ok := f.links[mcUUID]
 	if !ok || f.linkAuthSource[mcUUID] != authSourceThirdParty {
 		return false, nil
 	}
 	for _, u := range f.staff {
-		if u.ID == userID && u.Role == "admin" {
+		if u.ID == userID && staffRole(u.Role) {
 			return true, nil
 		}
 	}
@@ -638,6 +675,9 @@ func (f *fakeRepo) ClaimServer(_ context.Context, n, u string) (bool, error) {
 	if !present {
 		return false, ErrNotFound
 	}
+	if ok && f.claimQuotaRefuse[n] {
+		return false, ErrQuotaExceeded // mirrors the atomic gate losing the race
+	}
 	return ok, nil
 }
 func (f *fakeRepo) RecordJoin(_ context.Context, n, uuid string) error {
@@ -768,12 +808,18 @@ func (f *fakeRepo) CreateSession(_ context.Context, tokenHash, userID string, ex
 	return nil
 }
 func (f *fakeRepo) SessionUser(_ context.Context, tokenHash string, now time.Time) (*SessionedUser, error) {
+	if f.failSessionUser != nil {
+		return nil, f.failSessionUser
+	}
 	s, ok := f.sessions[tokenHash]
 	if !ok || s.revoked || !s.expiresAt.After(now) {
 		return nil, ErrNotFound
 	}
 	for _, u := range f.staff {
 		if u.ID == s.userID {
+			if f.seededDead(u.ID) {
+				return nil, ErrNotFound
+			}
 			return &SessionedUser{
 				ID: u.ID, Email: u.Email, Role: u.Role,
 			}, nil
@@ -788,6 +834,9 @@ func (f *fakeRepo) RevokeSession(_ context.Context, tokenHash string) error {
 	return nil
 }
 func (f *fakeRepo) GetSetting(_ context.Context, key string) ([]byte, error) {
+	if f.failGetSetting != nil {
+		return nil, f.failGetSetting
+	}
 	if v, ok := f.settings[key]; ok {
 		return v, nil
 	}
@@ -863,11 +912,29 @@ func (f *fakeRepo) ListUsers(_ context.Context, opts ListUsersOpts) ([]UserView,
 }
 
 func (f *fakeRepo) UserDetail(_ context.Context, userID string) (*UserDetail, error) {
+	deletedAt := time.Unix(1_700_000_000, 0)
 	for _, su := range f.seededUsers {
 		if su.view.ID == userID {
 			return &su.detail, nil
 		}
 	}
+	// Legacy fixtures seeded only into f.staff are live accounts (nothing marked
+	// them disabled or deleted), so detail reads must resolve them too — the
+	// liveness guards (discoverable login, owner protection) treat "unknown" as a
+	// fault, and these fixtures are known.
+	for _, u := range f.staff {
+		if u.ID == userID {
+			if f.deletedIDs[userID] {
+				// A soft-deleted account still HAS a detail row; it is flagged, not gone.
+				return &UserDetail{UserView: UserView{
+					ID: u.ID, Username: u.Username, Role: u.Role, Disabled: true,
+				}, DeletedAt: &deletedAt}, nil
+			}
+			return &UserDetail{UserView: UserView{
+				ID: u.ID, Username: u.Username, Email: u.Email, Role: u.Role,
+			}}, nil
+		}
+	}
 	return nil, ErrNotFound
 }
 
@@ -904,6 +971,12 @@ func (f *fakeRepo) UpdateUser(_ context.Context, userID string, patch UpdateUser
 			f.seededUsers[i].detail.Username = *patch.Username
 		}
 		if patch.Email != nil {
+			// Changing the address voids the proof of it, exactly like PGRepo:
+			// only VerifyEmailOTP may assert a verified address.
+			if *patch.Email != f.seededUsers[i].view.Email {
+				f.seededUsers[i].view.EmailVerified = false
+				f.seededUsers[i].detail.EmailVerified = false
+			}
 			f.seededUsers[i].view.Email = *patch.Email
 			f.seededUsers[i].detail.Email = *patch.Email
 		}
@@ -921,6 +994,20 @@ func (f *fakeRepo) DeleteUser(_ context.Context, userID, _ string) error {
 	for i, su := range f.seededUsers {
 		if su.view.ID == userID {
 			f.seededUsers = append(f.seededUsers[:i], f.seededUsers[i+1:]...)
+			f.deletedIDs[userID] = true
+			// Mirror PGRepo: deletion severs the account's identity assets so the
+			// closed account keeps neither a login credential nor a MC-UUID claim.
+			for uuid, uid := range f.links {
+				if uid == userID {
+					delete(f.links, uuid)
+					delete(f.linkAuthSource, uuid)
+				}
+			}
+			for cid, cred := range f.passkeyCreds {
+				if cred.UserID == userID {
+					delete(f.passkeyCreds, cid)
+				}
+			}
 			return nil
 		}
 	}
@@ -1068,7 +1155,52 @@ func (f *fakeRepo) RedeemMigration(_ context.Context, targetUserID, codeHash str
 
 // ---- quota admin fakes ----
 
+// liveUserExists mirrors PGRepo.requireLiveUser: the admin quota/link fakes
+// only act on a live seeded row, so a unit test can drive the unknown-user 404
+// the real FK would otherwise turn into a 500.
+func (f *fakeRepo) liveUserExists(id string) bool {
+	for _, su := range f.seededUsers {
+		if su.view.ID == id && su.detail.DeletedAt == nil {
+			return true
+		}
+	}
+	return false
+}
+
+// seededDead mirrors PGRepo's liveness filters (audit #33): a seeded user that was
+// disabled or soft-deleted is dead for the login doors and session validation. A
+// fixture that was never seeded (legacy tests put it straight into f.staff) is
+// treated as live, matching the fakes' pre-existing behavior.
+func (f *fakeRepo) seededDead(id string) bool {
+	if f.deletedIDs[id] {
+		return true
+	}
+	for _, su := range f.seededUsers {
+		if su.view.ID == id {
+			return su.view.Disabled || su.detail.DeletedAt != nil
+		}
+	}
+	return false
+}
+
+// seededDeleted is the narrower liveness query: soft-deleted only (a disabled
+// account still holds its identity, mirroring VerifyLinkCode's takeover rule).
+func (f *fakeRepo) seededDeleted(id string) bool {
+	if f.deletedIDs[id] {
+		return true
+	}
+	for _, su := range f.seededUsers {
+		if su.view.ID == id {
+			return su.detail.DeletedAt != nil
+		}
+	}
+	return false
+}
+
 func (f *fakeRepo) GetQuotas(_ context.Context, userID string) (*QuotaView, error) {
+	if !f.liveUserExists(userID) {
+		return nil, ErrNotFound
+	}
 	v := &QuotaView{UserID: userID}
 	if f.fakeQuotas == nil {
 		return v, nil
@@ -1083,6 +1215,9 @@ func (f *fakeRepo) GetQuotas(_ context.Context, userID string) (*QuotaView, erro
 }
 
 func (f *fakeRepo) SetQuotas(_ context.Context, userID string, qi QuotaInput, _ string) (*QuotaView, error) {
+	if !f.liveUserExists(userID) {
+		return nil, ErrNotFound
+	}
 	if f.fakeQuotas == nil {
 		f.fakeQuotas = map[string]*QuotaView{}
 	}
@@ -1135,6 +1270,9 @@ func (f *fakeRepo) UnlinkAccount(_ context.Context, userID, mcUUID string) error
 }
 
 func (f *fakeRepo) LinkAccount(_ context.Context, userID, mcUUID, authSource string) error {
+	if !f.liveUserExists(userID) {
+		return ErrNotFound
+	}
 	if existing, ok := f.links[mcUUID]; ok && existing != userID {
 		return ErrConflict
 	}
@@ -1153,7 +1291,7 @@ func (f *fakeRepo) LinkAccount(_ context.Context, userID, mcUUID, authSource str
 // is indistinguishable from no account: both yield ErrNotFound.
 func (f *fakeRepo) UserByEmail(_ context.Context, email string) (*StaffUser, error) {
 	for _, u := range f.staff {
-		if u.EmailVerified && strings.EqualFold(u.Email, email) {
+		if u.EmailVerified && strings.EqualFold(u.Email, email) && !f.seededDead(u.ID) {
 			su := *u
 			return &su, nil
 		}
@@ -1313,6 +1451,7 @@ type fakeCluster struct {
 	desired   map[string]v1alpha1.DesiredState
 	created   map[string]CreateServerInput // name -> the validated input it was created from
 	patched   map[string]ServerSpecPatch   // name -> the validated spec patch it received
+	noWorld   map[string]bool              // server names modeled WITHOUT a world volume (never started / reaped)
 	createErr error
 	pingErr   error
 }
@@ -1320,7 +1459,7 @@ type fakeCluster struct {
 func newFakeCluster() *fakeCluster {
 	return &fakeCluster{byName: map[string]*ServerInfo{}, bySub: map[string]*ServerInfo{},
 		desired: map[string]v1alpha1.DesiredState{}, created: map[string]CreateServerInput{},
-		patched: map[string]ServerSpecPatch{}}
+		patched: map[string]ServerSpecPatch{}, noWorld: map[string]bool{}}
 }
 func (c *fakeCluster) GetServer(_ context.Context, n string) (*ServerInfo, error) {
 	if s, ok := c.byName[n]; ok {
@@ -1336,6 +1475,13 @@ func (c *fakeCluster) GetBySubdomain(_ context.Context, s string) (*ServerInfo,
 }
 func (c *fakeCluster) ListServers(_ context.Context) ([]ServerInfo, error) { return c.list, nil }
 func (c *fakeCluster) Ping(_ context.Context) error                        { return c.pingErr }
+
+// WorldVolumeExists models the world PVC: present unless the test named the
+// server in noWorld (never started / already reaped).
+func (c *fakeCluster) WorldVolumeExists(_ context.Context, n string) (bool, error) {
+	return !c.noWorld[n], nil
+}
+
 func (c *fakeCluster) SetDesiredState(_ context.Context, n string, s v1alpha1.DesiredState) error {
 	c.desired[n] = s
 	return nil
@@ -1671,6 +1817,44 @@ func TestFleetAdminRead(t *testing.T) {
 		}
 	})
 
+	t.Run("system services are marked read-only", func(t *testing.T) {
+		// The login gate and the lobby carry reserved names, so every per-server
+		// route rejects them; the fleet row must say "system" so the cockpit
+		// renders them without actions that would 400.
+		sysCl := newFakeCluster()
+		sysCl.list = []ServerInfo{
+			{Name: "login", Phase: "Running", Ready: true},
+			{Name: "lobby", Phase: "Running", Ready: true},
+			{Name: "survival", Phase: "Stopped"},
+		}
+		api := newTestAPI(newFakeRepo(), sysCl)
+		api.External = staticExternal{p: &Principal{UserID: "a1", Email: "[email protected]",
+			Role: "admin", ViaAdminAccess: true}}
+		w := do(api.ExternalHandler(), "GET", "/api/v1/fleet", "", nil)
+		if w.Code != http.StatusOK {
+			t.Fatalf("code = %d, want 200 (%s)", w.Code, w.Body.String())
+		}
+		var got struct {
+			Servers []struct {
+				Name   string `json:"name"`
+				System bool   `json:"system"`
+			} `json:"servers"`
+		}
+		if err := json.Unmarshal(w.Body.Bytes(), &got); err != nil {
+			t.Fatalf("body not JSON: %v", err)
+		}
+		byName := map[string]bool{}
+		for _, r := range got.Servers {
+			byName[r.Name] = r.System
+		}
+		if !byName["login"] || !byName["lobby"] {
+			t.Errorf("system flags = %+v, want login+lobby marked", byName)
+		}
+		if byName["survival"] {
+			t.Errorf("survival marked system; only platform services are")
+		}
+	})
+
 	t.Run("owner merges for claimed, absent for unclaimed", func(t *testing.T) {
 		repo := newFakeRepo()
 		// Only "survival" is claimed; "creative"/"skyblock" stay unowned.
@@ -1760,6 +1944,21 @@ func TestClaimStateMachine(t *testing.T) {
 			t.Fatalf("code = %d body %s", w.Code, w.Body.String())
 		}
 	})
+	t.Run("atomic gate refusal -> 403 quota_exceeded", func(t *testing.T) {
+		// The advisory pre-check passed, but ClaimServer's serialized re-check
+		// (audit #4) refuses: the caller must see the same 403, not a 500.
+		repo := newFakeRepo()
+		repo.linked["u1"] = true
+		repo.quota["u1"] = true
+		repo.claimOK["survival"] = true
+		repo.claimQuotaRefuse["survival"] = true
+		api := newTestAPI(repo, newFakeCluster())
+		api.External = staticExternal{p: user}
+		w := do(api.ExternalHandler(), "POST", "/api/v1/servers/survival/claim", "", nil)
+		if w.Code != http.StatusForbidden || decodeErr(t, w) != "quota_exceeded" {
+			t.Fatalf("code = %d body %s, want 403 quota_exceeded", w.Code, w.Body.String())
+		}
+	})
 	t.Run("already claimed -> 409", func(t *testing.T) {
 		repo := newFakeRepo()
 		repo.linked["u1"] = true
@@ -2195,3 +2394,38 @@ func TestAccessVerifier(t *testing.T) {
 		}
 	})
 }
+
+// TestSessionAuthOutageIs503Not401: a session-store outage must surface as 503
+// auth_unavailable, not a 401 that reads as "please log in again". Both failure
+// points are covered — the local_auth_enabled read and the session row read —
+// plus the regression that a genuinely missing session still answers 401.
+func TestSessionAuthOutageIs503Not401(t *testing.T) {
+	apiWith := func(repo *fakeRepo) *API {
+		a := newTestAPI(repo, newFakeCluster())
+		a.External = SessionAuth{Repo: repo, RootDomain: testRoot, AdminHostname: "op.console." + testRoot}
+		return a
+	}
+	cookie := map[string]string{"Cookie": sessionCookieName + "=any"}
+	outage := errors.New("dial tcp 10.0.0.5:5432: connect: connection refused")
+
+	repo := newFakeRepo()
+	repo.settings[LocalAuthEnabledKey] = []byte("true")
+	repo.failGetSetting = outage
+	if w := do(apiWith(repo).ExternalHandler(), "GET", "/api/v1/me", "", cookie); w.Code != http.StatusServiceUnavailable || decodeErr(t, w) != "auth_unavailable" {
+		t.Fatalf("settings read outage = %d body %s, want 503 auth_unavailable", w.Code, w.Body.String())
+	}
+
+	repo = newFakeRepo()
+	repo.settings[LocalAuthEnabledKey] = []byte("true")
+	repo.failSessionUser = outage
+	if w := do(apiWith(repo).ExternalHandler(), "GET", "/api/v1/me", "", cookie); w.Code != http.StatusServiceUnavailable || decodeErr(t, w) != "auth_unavailable" {
+		t.Fatalf("session read outage = %d body %s, want 503 auth_unavailable", w.Code, w.Body.String())
+	}
+
+	// Regression: fail-closed auth (missing/invalid session) stays a 401.
+	repo = newFakeRepo()
+	repo.settings[LocalAuthEnabledKey] = []byte("true")
+	if w := do(apiWith(repo).ExternalHandler(), "GET", "/api/v1/me", "", cookie); w.Code != http.StatusUnauthorized || decodeErr(t, w) != "unauthorized" {
+		t.Fatalf("missing session = %d body %s, want 401 unauthorized", w.Code, w.Body.String())
+	}
+}

internal/api/auth.go

@@ -17,13 +17,13 @@ type Principal struct {
 	UserID string
 	// Email is the audited actor identity (spec §14: audit actor = Access email).
 	Email string
-	// Role is "admin" or "user" (mirrors users.role).
+	// Role is "owner", "admin", or "user" (mirrors users.role).
 	Role string
 	// ViaAdminAccess is true only when the request arrived through an admin-graded
 	// path: the admin.* Zero-Trust hostname (Cloudflare Access, the remote face) OR
 	// a local session presented on the op.console host (SessionAuth, the
 	// passwordless face). Admin-tier operations require it in addition to
-	// Role=="admin" (spec §14: ZT is graded by operation). A role=admin session
+	// a staff role (spec §14: ZT is graded by operation). A staff session
 	// arriving on the player console (console.*) never sets it.
 	ViaAdminAccess bool
 	// EmailVerified mirrors users.email_verified. The lockdown middleware gates
@@ -49,7 +49,7 @@ func staffRole(role string) bool {
 }
 
 // IsAdmin reports whether the principal may perform admin-tier operations.
-// Both the role claim and the admin Access path are required: a role=admin
+// Both the role claim and the admin Access path are required: a staff
 // session arriving on panel.* must not bypass the Zero-Trust boundary.
 // An owner implicitly passes this check (the owner role is a superset of admin).
 func (p *Principal) IsAdmin() bool {

internal/api/cluster.go

@@ -81,6 +81,12 @@ type Cluster interface {
 
 	// GetServer reads one MinecraftServer's lifecycle view, or ErrNotFound.
 	GetServer(ctx context.Context, name string) (*ServerInfo, error)
+	// WorldVolumeExists reports whether the server's world PVC exists in the
+	// server namespace. A server that never started — or whose world the
+	// retention reaper already archived and deleted — has no claim, and a
+	// backup/restore Job would hang Pending on the missing volume with nothing
+	// ever recorded, so both handlers refuse those up front.
+	WorldVolumeExists(ctx context.Context, name string) (bool, error)
 	// GetBySubdomain finds the MinecraftServer whose spec.subdomain matches, or
 	// ErrNotFound.
 	GetBySubdomain(ctx context.Context, subdomain string) (*ServerInfo, error)

internal/api/errors.go

@@ -15,6 +15,13 @@ var (
 	ErrNotFound = errors.New("not found")
 	// ErrConflict means an atomic precondition failed (e.g. claim lost the race).
 	ErrConflict = errors.New("conflict")
+	// ErrQuotaExceeded means an ownership write would push the user over a quota
+	// cap (spec §9.3). ClaimServer — the atomic gate — returns it when a claim
+	// passes the handler's advisory pre-check but loses the serialized re-check
+	// (two concurrent claims by one user); handlers map it to a 403
+	// quota_exceeded, the same answer the pre-check gives, so the CONCURRENT case
+	// and the SEQUENTIAL case are indistinguishable to the caller.
+	ErrQuotaExceeded = errors.New("server quota exhausted")
 	// ErrLinkCodeInvalid means an account-link code is unknown or expired (spec
 	// §10). It is a client error (the verify endpoint exists; the code is bad), so
 	// handlers map it to 400, not 404.
@@ -44,14 +51,22 @@ var (
 	// consumed, or expired) — so handlers map it to 400, not 404.
 	ErrPasskeyChallengeInvalid = errors.New("passkey challenge invalid or expired")
 	// ErrPlayerBindForbidden means a public Bind-Code redemption resolved to a STAFF
-	// account (role=admin), which the player-console bootstrap refuses (console-tier
+	// account (admin or owner), which the player-console bootstrap refuses
-	// access model). Operators authenticate at op.console behind Zero Trust, never via
+	// (console-tier access model). Staff authenticate at op.console behind Zero Trust,
-	// the account-less console.<root_domain> door, so the public bootstrap provably
+	// never via the account-less console.<root_domain> door, so the public bootstrap
-	// never mints a session for an admin identity. It is distinct from ErrConflict so
+	// provably never mints a session for a staff identity. It is distinct from
-	// the handler answers 403 (wrong door) rather than 409 (already linked).
+	// ErrConflict so the handler answers 403 (wrong door) rather than 409.
 	ErrPlayerBindForbidden = errors.New("bind code belongs to a staff account")
+	// ErrPlayerAccountRetired means a Bind-Code redemption resolved to an account the
+	// platform has closed: an owner soft-deleted it, or it is disabled (locked out).
+	// Reusing the row would mint a fresh session for a dead account — the same
+	// resurrection the login doors refuse by resolving only live accounts — so the
+	// redeemer gets an explicit 403 instead. The code is NOT consumed, so re-enabling
+	// the account and retrying still works within the code's TTL.
+	ErrPlayerAccountRetired = errors.New("player account is retired or disabled")
 	// ErrEmailTaken means a verified email would collide with another account's
-	// already-verified address (spec §B email-first login foundation, migration 0010).
+	// already-verified address (spec §B email-first login foundation; the
+	// users_verified_email_unique index ships in migration 0020).
 	// VerifyEmailOTP returns it — WITHOUT consuming the code, since the address, not
 	// the code, is the problem — when a DIFFERENT user has already proven the same
 	// address case-insensitively. It is the clean, application-level counterpart of
@@ -89,8 +104,13 @@ func newError(status int, code, format string, a ...any) *apiError {
 // Common errors reused across handlers.
 var (
 	errUnauthorized = newError(http.StatusUnauthorized, "unauthorized", "authentication required")
-	errForbidden    = newError(http.StatusForbidden, "forbidden", "not permitted")
+	// errAuthUnavailable answers when the session store itself is unreachable
-	errBadRequest   = newError(http.StatusBadRequest, "bad_request", "invalid request")
+	// (Postgres down): an outage is not a credential verdict, so the caller gets
+	// 503 "retry" instead of a 401 that reads as "log in again".
+	errAuthUnavailable = newError(http.StatusServiceUnavailable, "auth_unavailable",
+		"authentication is temporarily unavailable; retry shortly")
+	errForbidden  = newError(http.StatusForbidden, "forbidden", "not permitted")
+	errBadRequest = newError(http.StatusBadRequest, "bad_request", "invalid request")
 )
 
 // writeJSON writes v as an indented JSON body with the given status.

internal/api/handlers_account_test.go

@@ -1,7 +1,9 @@
 package api
 
 import (
+	"context"
 	"encoding/json"
+	"errors"
 	"net/http"
 	"net/http/httptest"
 	"strings"
@@ -252,6 +254,58 @@ func TestLinkVerifyIdempotent(t *testing.T) {
 	}
 }
 
+// A link whose account was SOFT-DELETED is unclaimed: a fresh in-game code lets a
+// live account take it over (the migrated-source path — retire keeps the link but
+// kills the account), while a merely disabled holder keeps its identity so the
+// lockout cannot be re-linked away, and neither dead link has in-game standing
+// (UserByMCUUID reads it exactly like an unlinked UUID). Audit #33's in-game half.
+func TestLinkVerifyTakesOverDeletedLinkOnly(t *testing.T) {
+	ctx := context.Background()
+	const mcGone = "55555555-5555-5555-5555-555555555555"
+	const mcLocked = "66666666-6666-6666-6666-666666666666"
+	user := &Principal{UserID: "u-take", Email: "[email protected]", Role: "user"}
+	repo := newFakeRepo()
+	repo.seedUser(UserView{ID: "u-gone", Username: "gone", Email: "[email protected]", Role: "user"})
+	repo.seedUser(UserView{ID: "u-locked", Username: "locked", Email: "[email protected]", Role: "user"})
+	if err := repo.DeleteUser(ctx, "u-gone", "test"); err != nil {
+		t.Fatalf("DeleteUser: %v", err)
+	}
+	repo.links[mcGone] = "u-gone" // a retired source's link outlives the account
+
+	if _, err := repo.UserByMCUUID(ctx, mcGone); !errors.Is(err, ErrNotFound) {
+		t.Fatalf("UserByMCUUID(deleted link) = %v, want ErrNotFound (no in-game standing)", err)
+	}
+
+	repo.linkCodes["TAKEOVER"] = fakeLinkCode{mcUUID: mcGone, expiresAt: time.Unix(1_700_000_600, 0)}
+	api := newTestAPI(repo, newFakeCluster())
+	api.External = staticExternal{p: user}
+	if w := do(api.ExternalHandler(), "POST", "/api/v1/account/link/verify", `{"code":"TAKEOVER"}`, nil); w.Code != http.StatusOK {
+		t.Fatalf("takeover verify: code = %d, want 200 (%s)", w.Code, w.Body.String())
+	}
+	if repo.links[mcGone] != "u-take" {
+		t.Errorf("links[%s] = %q after takeover, want u-take", mcGone, repo.links[mcGone])
+	}
+
+	// A disabled (not deleted) holder keeps the identity: 409, code survives, link unmoved.
+	if err := repo.SetUserDisabled(ctx, "u-locked", true); err != nil {
+		t.Fatalf("disable: %v", err)
+	}
+	repo.links[mcLocked] = "u-locked"
+	repo.linkCodes["LOCKED12"] = fakeLinkCode{mcUUID: mcLocked, expiresAt: time.Unix(1_700_000_600, 0)}
+	if _, err := repo.UserByMCUUID(ctx, mcLocked); !errors.Is(err, ErrNotFound) {
+		t.Fatalf("UserByMCUUID(disabled link) = %v, want ErrNotFound", err)
+	}
+	if w := do(api.ExternalHandler(), "POST", "/api/v1/account/link/verify", `{"code":"LOCKED12"}`, nil); w.Code != http.StatusConflict {
+		t.Fatalf("takeover of a disabled holder: code = %d, want 409 (%s)", w.Code, w.Body.String())
+	}
+	if repo.links[mcLocked] != "u-locked" {
+		t.Errorf("disabled holder's link moved to %q", repo.links[mcLocked])
+	}
+	if _, ok := repo.linkCodes["LOCKED12"]; !ok {
+		t.Error("refused verify consumed the code")
+	}
+}
+
 // TestLinkAuthSourcePropagates proves auth_source survives the whole §10 flow: a
 // thirdparty source captured in-game at mint reaches the durable link and the
 // verify response — the value the web side can never originate itself.

internal/api/handlers_auth_email.go

@@ -241,7 +241,9 @@ func (a *API) handleLoginEmailVerify(w http.ResponseWriter, r *http.Request) {
 	// Code redeemed. Refuse staff here — never before the verify — so op.console keeps
 	// its Zero-Trust + in-game-approval gates and this public door provably yields only
 	// a role=user player session (mirrors handleBindRedeem's refuse-staff contract).
-	if u.Role == "admin" {
+	// Staff means anything above role=user: an admin OR the role=owner identity. The
+	// player door must yield only player sessions.
+	if u.Role != "user" {
 		writeError(w, r, newError(http.StatusForbidden, "staff_account",
 			"that account is staff; sign in at the operator console"))
 		return

internal/api/handlers_auth_email_test.go

@@ -1,7 +1,9 @@
 package api
 
 import (
+	"context"
 	"encoding/json"
+	"errors"
 	"net/http"
 	"net/http/httptest"
 	"testing"
@@ -510,12 +512,20 @@ func TestLoginEmailPurposeSeparation(t *testing.T) {
 // a wrong code for a staff address answers the same invalid_code as for anyone, and
 // the 403 costs the valid code (verify-then-refuse), so it cannot be farmed as an
 // is-this-address-staff oracle.
+// Staff means admin AND owner (migration 0011): both must be refused at the
+// player door, after the code proves mailbox control.
 func TestLoginEmailVerifyRefusesStaff(t *testing.T) {
+	for _, role := range []string{"admin", "owner"} {
+		t.Run(role, func(t *testing.T) { verifyStaffRefusedAtPlayerDoor(t, role) })
+	}
+}
+
+func verifyStaffRefusedAtPlayerDoor(t *testing.T, role string) {
 	repo := newFakeRepo()
 	repo.settings[LocalAuthEnabledKey] = []byte("true")
 	repo.staff["owner"] = &StaffUser{
 		ID: "a1", Username: "owner", Email: "[email protected]",
-		Role: "admin", EmailVerified: true,
+		Role: role, EmailVerified: true,
 	}
 	mailer := &captureMailer{}
 	api := newTestAPI(repo, newFakeCluster())
@@ -610,3 +620,94 @@ func TestLoginEmailStartFailedDeliveryReleasesCooldown(t *testing.T) {
 		t.Errorf("mailer calls = %d, want 2 (one failed, one delivered)", mailer.calls)
 	}
 }
+
+// A disabled or soft-deleted account is DEAD at every door: the pre-session
+// resolvers refuse it (uniformly, so the door stays no-oracle), a session that
+// was live a moment ago stops authenticating, DeleteUser severs the account's
+// passkeys and Minecraft links, and the bind door refuses to reuse the retired
+// identity instead of minting a session for it. Audit #33 found the opposite
+// live: a deleted user re-logged-in through the email door and GET /me answered
+// 200 — deletion and the disable lockout were both bypassable by logging in again.
+func TestDeadAccountsCannotLogInOrKeepSessions(t *testing.T) {
+	ctx := context.Background()
+	now := time.Unix(1_700_000_000, 0)
+
+	repo := newFakeRepo()
+	repo.settings[LocalAuthEnabledKey] = []byte("true")
+	repo.seedUser(UserView{ID: "u-dead", Username: "dead", Email: "[email protected]", Role: "user"})
+	repo.staff["dead"].EmailVerified = true
+	mailer := &captureMailer{}
+	api := newTestAPI(repo, newFakeCluster())
+	api.Mailer = mailer
+	eh := api.ExternalHandler()
+
+	// Control: alive — the door resolves the account and mails a real code, and a
+	// session minted for it authenticates.
+	if w := do(eh, "POST", "/api/v1/auth/email/start", `{"email":"[email protected]"}`, jsonHeader); w.Code != http.StatusAccepted {
+		t.Fatalf("alive start: code = %d, want 202 (%s)", w.Code, w.Body.String())
+	}
+	if mailer.calls != 1 {
+		t.Fatalf("alive start mailed %d codes, want 1", mailer.calls)
+	}
+	repo.sessions["h-live"] = &fakeSession{userID: "u-dead", expiresAt: now.Add(time.Hour)}
+	if _, err := repo.SessionUser(ctx, "h-live", now); err != nil {
+		t.Fatalf("live SessionUser: %v", err)
+	}
+
+	// Disabled: the start is neutral (no mail), verify refuses, session dies.
+	if err := repo.SetUserDisabled(ctx, "u-dead", true); err != nil {
+		t.Fatalf("disable: %v", err)
+	}
+	// The alive start's reservation must not mask the neutral branch: clear the
+	// throttle's window (test-only; the limiter itself is rebuilt lazily once).
+	lim := api.otpLimiter()
+	lim.mu.Lock()
+	lim.last = map[string]time.Time{}
+	lim.mu.Unlock()
+	if w := do(eh, "POST", "/api/v1/auth/email/start", `{"email":"[email protected]"}`, jsonHeader); w.Code != http.StatusAccepted {
+		t.Fatalf("disabled start: code = %d, want 202 neutral (%s)", w.Code, w.Body.String())
+	}
+	if mailer.calls != 1 {
+		t.Fatalf("disabled start mailed a code (%d calls) — a dead account must resolve to nothing", mailer.calls)
+	}
+	if w := do(eh, "POST", "/api/v1/auth/email/verify", `{"email":"[email protected]","code":"000000"}`, jsonHeader); w.Code != http.StatusBadRequest || decodeErr(t, w) != "invalid_code" {
+		t.Fatalf("disabled verify: code = %d body %s, want 400 invalid_code", w.Code, w.Body.String())
+	}
+	if _, err := repo.SessionUser(ctx, "h-live", now); !errors.Is(err, ErrNotFound) {
+		t.Fatalf("disabled SessionUser = %v, want ErrNotFound", err)
+	}
+
+	// Deleted: same refusals; assets severed (links released, passkeys dropped).
+	if err := repo.SetUserDisabled(ctx, "u-dead", false); err != nil {
+		t.Fatalf("re-enable: %v", err)
+	}
+	uuid := "11111111-2222-3333-4444-555555555555"
+	repo.links[uuid] = "u-dead"
+	repo.passkeyCreds["pk-1"] = PasskeyCredential{ID: "pk-1", UserID: "u-dead", CredentialID: "cred-1"}
+	if err := repo.DeleteUser(ctx, "u-dead", "test"); err != nil {
+		t.Fatalf("delete: %v", err)
+	}
+	if w := do(eh, "POST", "/api/v1/auth/email/verify", `{"email":"[email protected]","code":"000000"}`, jsonHeader); w.Code != http.StatusBadRequest || decodeErr(t, w) != "invalid_code" {
+		t.Fatalf("deleted verify: code = %d body %s, want 400 invalid_code", w.Code, w.Body.String())
+	}
+	if _, err := repo.SessionUser(ctx, "h-live", now); !errors.Is(err, ErrNotFound) {
+		t.Fatalf("deleted SessionUser = %v, want ErrNotFound", err)
+	}
+	if _, ok := repo.links[uuid]; ok {
+		t.Error("DeleteUser left the Minecraft link: the UUID stays claimed forever")
+	}
+	if _, ok := repo.passkeyCreds["pk-1"]; ok {
+		t.Error("DeleteUser left the passkey: a login credential outlives the account")
+	}
+
+	// The bind door refuses to reuse the retired identity (a stale link that
+	// predates the fix, or a username squatted by the deleted row).
+	repo.links[uuid] = "u-dead"
+	repo.linkCodes["CODE1234"] = fakeLinkCode{mcUUID: uuid, authSource: "mojang", expiresAt: now.Add(time.Hour)}
+	if _, _, _, err := repo.RedeemPlayerBindCode(ctx, "u-new", "CODE1234", now); !errors.Is(err, ErrPlayerAccountRetired) {
+		t.Fatalf("bind redeem onto a deleted account = %v, want ErrPlayerAccountRetired", err)
+	}
+	if _, ok := repo.linkCodes["CODE1234"]; !ok {
+		t.Error("refused redeem consumed the code; re-enabling the account must stay retryable within TTL")
+	}
+}

internal/api/handlers_backup_now_test.go

@@ -134,6 +134,22 @@ func TestBackupNow(t *testing.T) {
 		}
 	})
 
+	t.Run("no world volume -> 409 no_world_volume, no backup", func(t *testing.T) {
+		// A never-started (or reaped) server has no world PVC: the Job would hang
+		// Pending on the missing claim with nothing recorded, so the gate must
+		// refuse before the backuper is reached.
+		api, _, cl, backuper := mk()
+		cl.noWorld["survival"] = true
+		api.External = staticExternal{p: owner}
+		w := do(api.ExternalHandler(), "POST", path, "", nil)
+		if w.Code != http.StatusConflict || decodeErr(t, w) != "no_world_volume" {
+			t.Fatalf("code = %d body %s", w.Code, w.Body.String())
+		}
+		if backuper.calls != 0 {
+			t.Fatal("a world-less server must not reach the backuper")
+		}
+	})
+
 	t.Run("nil Backuper -> 503 backup_unavailable", func(t *testing.T) {
 		api, _, _, _ := mk()
 		api.Backuper = nil
@@ -240,6 +256,20 @@ func TestInternalBackup(t *testing.T) {
 		}
 	})
 
+	t.Run("no world volume -> 409 no_world_volume, no backup", func(t *testing.T) {
+		// The break-glass face shares enqueueBackup, so the world-volume gate must
+		// hold here too — this is the face the TUI's Sync picker drives.
+		api, _, cl, backuper := mk()
+		cl.noWorld["survival"] = true
+		w := do(api.InternalHandler(), "POST", path, "", jsonHeader)
+		if w.Code != http.StatusConflict || decodeErr(t, w) != "no_world_volume" {
+			t.Fatalf("code = %d body %s", w.Code, w.Body.String())
+		}
+		if backuper.calls != 0 {
+			t.Fatal("a world-less server must not reach the backuper")
+		}
+	})
+
 	t.Run("nil Backuper -> 503 backup_unavailable", func(t *testing.T) {
 		api, _, _, _ := mk()
 		api.Backuper = nil

internal/api/handlers_backups.go

@@ -9,6 +9,16 @@ import (
 	"felis.lolicon.best/internal/naming"
 )
 
+// errNoWorldVolume is the shared 409 for backup and restore when the server's
+// world PVC does not exist: the Job would only hang Pending on the missing
+// claim — invisible to the caller and to the backups list — so the handlers
+// refuse up front. Starting the server once (which creates the claim via the
+// StatefulSet volumeClaimTemplate) unlocks both ops.
+func errNoWorldVolume() error {
+	return newError(http.StatusConflict, "no_world_volume",
+		"this server has no world volume yet — start it once to create it, then retry")
+}
+
 // handleListBackups lists the world backups visible to the caller (spec §7 GET
 // /api/v1/backups; world_backups in §22). It is app-tier: an admin sees every
 // present backup; a regular user sees only the backups of worlds they formerly
@@ -154,6 +164,19 @@ func (a *API) handleRestoreBackup(w http.ResponseWriter, r *http.Request) {
 		return
 	}
 
+	// World-volume gate: the restore Job mounts the world PVC read-write to unpack
+	// the archive into it, so a missing claim means a Pod stuck Pending — a 202
+	// "restoring" with nothing ever written. Same refusal as the backup face
+	// (shared errNoWorldVolume): the operator starts the server once to create the
+	// claim, then restores into it.
+	if exists, err := a.Cluster.WorldVolumeExists(r.Context(), name); err != nil {
+		writeError(w, r, err)
+		return
+	} else if !exists {
+		writeError(w, r, errNoWorldVolume())
+		return
+	}
+
 	// Restorer is optional: when unwired the endpoint reports 503 rather than
 	// panicking, so the authorization boundary above is exercised even before the
 	// restore-Job executor is wired (see Restorer).
@@ -285,6 +308,18 @@ func (a *API) enqueueBackup(w http.ResponseWriter, r *http.Request, name string,
 		return
 	}
 
+	// World-volume gate: the Job mounts the world PVC by claim name, and a missing
+	// claim would leave its Pod Pending — a 202 "backing_up" with nothing ever
+	// recorded anywhere. A never-started or already-reaped server is refused with
+	// the same specificity as the stopped gate.
+	if exists, err := a.Cluster.WorldVolumeExists(r.Context(), name); err != nil {
+		writeError(w, r, err)
+		return
+	} else if !exists {
+		writeError(w, r, errNoWorldVolume())
+		return
+	}
+
 	// Backuper is optional: when unwired the endpoint reports 503 rather than
 	// panicking, so the authorization boundary above is exercised even before the
 	// backup-Job executor is wired (see Backuper).

internal/api/handlers_backups_test.go

@@ -269,6 +269,21 @@ func TestRestoreBackup(t *testing.T) {
 		}
 	})
 
+	t.Run("no world volume -> 409 no_world_volume, no restore", func(t *testing.T) {
+		// Restoring into a missing world PVC would leave the Job Pending on the
+		// missing claim — a 202 "restoring" that never writes anything.
+		api, _, cl, restorer := mk()
+		cl.noWorld["survival"] = true
+		api.External = staticExternal{p: owner}
+		w := do(api.ExternalHandler(), "POST", path, "", nil)
+		if w.Code != http.StatusConflict || decodeErr(t, w) != "no_world_volume" {
+			t.Fatalf("code = %d body %s", w.Code, w.Body.String())
+		}
+		if restorer.calls != 0 {
+			t.Fatal("a world-less server must not reach the restorer")
+		}
+	})
+
 	t.Run("nil Restorer -> 503 restore_unavailable", func(t *testing.T) {
 		api, _, _, _ := mk()
 		api.Restorer = nil

internal/api/handlers_email_otp.go

@@ -187,9 +187,11 @@ type emailOTPVerifyRequest struct {
 
 // handleEmailOTPVerify redeems a code for the caller (spec §B2, external app face).
 // Outcomes mirror the link-verify shape: an invalid/expired/mismatched code → 400
-// invalid_code, a locked code (too many wrong guesses) → 429 otp_locked, and on
+// invalid_code, a locked code (too many wrong guesses) → 429 otp_locked, an address
-// success the user's email is written and email_verified flips true. The verified
+// another account already proved → 409 email_taken (the code stays live — the
-// address is echoed so the panel can render it.
+// address, not the code, is the problem), and on success the user's email is
+// written and email_verified flips true. The verified address is echoed so the
+// panel can render it.
 func (a *API) handleEmailOTPVerify(w http.ResponseWriter, r *http.Request) {
 	p := principalFromContext(r.Context())
 	var req emailOTPVerifyRequest
@@ -211,6 +213,10 @@ func (a *API) handleEmailOTPVerify(w http.ResponseWriter, r *http.Request) {
 	case errors.Is(err, ErrOTPInvalid):
 		writeError(w, r, newError(http.StatusBadRequest, "invalid_code", "email code is invalid or expired"))
 		return
+	case errors.Is(err, ErrEmailTaken):
+		writeError(w, r, newError(http.StatusConflict, "email_taken",
+			"that email is already verified on another account; sign in with it or use another address"))
+		return
 	case err != nil:
 		writeError(w, r, err)
 		return

internal/api/handlers_email_otp_test.go

@@ -339,6 +339,20 @@ func TestEmailOTPVerifyRejections(t *testing.T) {
 			t.Fatalf("code = %d body %s, want 429 otp_locked", w.Code, w.Body.String())
 		}
 	})
+	t.Run("address proven elsewhere -> 409 email_taken, not consumed", func(t *testing.T) {
+		repo := newFakeRepo()
+		live(repo, "tk", otpCodeHash("123456"), time.Unix(1_700_000_600, 0), 0)
+		// Another account already proved the same address: login resolves accounts
+		// BY verified email, so the second proof must be refused.
+		repo.staff["other"] = &StaffUser{ID: "u2", Email: "[email protected]", EmailVerified: true}
+		w := do(mk(repo), "POST", "/api/v1/account/email/verify", `{"code":"123456"}`, nil)
+		if w.Code != http.StatusConflict || decodeErr(t, w) != "email_taken" {
+			t.Fatalf("code = %d body %s, want 409 email_taken", w.Code, w.Body.String())
+		}
+		if repo.otps["tk"].consumed {
+			t.Error("a taken address must not consume the code")
+		}
+	})
 }
 
 // TestEmailOTPBruteForceLockout drives the lockout end-to-end through the handler:

internal/api/handlers_files.go

@@ -210,6 +210,19 @@ func (a *API) authorizeFileOp(w http.ResponseWriter, r *http.Request) (string, b
 		return "", false
 	}
 
+	// World-volume gate, matching the backup/restore faces: the Job mounts the
+	// world PVC by claim name, so a server that has never started (or was already
+	// reaped) has no claim to mount and its Pod sits Pending until the executor's
+	// wait expires — a knowably impossible request answered by a 90s hang and a
+	// misleading 504. Refuse up front with the same specific 409.
+	if exists, err := a.Cluster.WorldVolumeExists(r.Context(), name); err != nil {
+		writeError(w, r, err)
+		return "", false
+	} else if !exists {
+		writeError(w, r, errNoWorldVolume())
+		return "", false
+	}
+
 	// Files is optional: when unwired the endpoints report 503 rather than
 	// panicking, so the authorization boundary above is exercised even before the
 	// file-Job executor is wired (see FileEditor).

internal/api/handlers_files_test.go

@@ -119,6 +119,47 @@ func TestFileEditorStoppedGate(t *testing.T) {
 	}
 }
 
+// TestFileEditorWorldVolumeGate pins the second physical gate: a server with no
+// world PVC (never started, or already reaped) has no claim for the Job to mount,
+// so its Pod would sit Pending until the executor's wait timed out — a 90s hang
+// and a misleading 504 files_timeout for a request that is knowably impossible.
+// All three routes must refuse BEFORE creating a Job, with the same specific 409
+// the backup/restore faces use.
+func TestFileEditorWorldVolumeGate(t *testing.T) {
+	owner := &Principal{UserID: "owner1", Email: "[email protected]", Role: "user"}
+
+	routes := []struct {
+		name   string
+		method string
+		path   string
+		body   string
+	}{
+		{"list", "GET", "/api/v1/servers/survival/files?path=config", ""},
+		{"read", "GET", "/api/v1/servers/survival/file?path=server.properties", ""},
+		{"write", "PUT", "/api/v1/servers/survival/file?path=server.properties", `{"content":"aGk="}`},
+	}
+
+	for _, rt := range routes {
+		t.Run(rt.name+" without a world volume -> 409 no_world_volume", func(t *testing.T) {
+			api, _, cl, files := mkFiles()
+			cl.noWorld["survival"] = true
+			api.External = staticExternal{p: owner}
+
+			var hdr map[string]string
+			if rt.body != "" {
+				hdr = jsonHeader
+			}
+			w := do(api.ExternalHandler(), rt.method, rt.path, rt.body, hdr)
+			if w.Code != http.StatusConflict || decodeErr(t, w) != "no_world_volume" {
+				t.Fatalf("code = %d body %s", w.Code, w.Body.String())
+			}
+			if files.calls != 0 {
+				t.Fatal("no claim to mount — the file Job must never be created")
+			}
+		})
+	}
+}
+
 // TestFileEditorAuthorization pins who may touch a world's files. It is the same
 // owner-or-admin rule the backup routes enforce, and it must hold on all three
 // routes — a read-only route leaking another owner's config (an RCON password

internal/api/handlers_hasjoined.go

@@ -142,7 +142,9 @@ func (a *API) handleHasJoined(w http.ResponseWriter, r *http.Request) {
 
 	// Canonicalize identity. A trusted (Mojang) source keeps its UUID; a self-asserted
 	// source is rewritten into felisAuthNS so it can never land in Mojang's UUID space
-	// nor onto another source's. An unparseable identity UUID is not trustworthy → reject.
+	// nor onto another source's. resolveHasJoined has already screened both shapes and
+	// skipped unusable ones as failed; the two guards below are the last line before
+	// anything leaves, kept even though nothing reaches them.
 	var canonical uuid.UUID
 	if src.Identity {
 		id, err := uuid.Parse(prof.ID)
@@ -154,8 +156,8 @@ func (a *API) handleHasJoined(w http.ResponseWriter, r *http.Request) {
 	} else {
 		// A third-party source is untrusted input, its name included: nothing stops a
 		// hostile or sloppy root from answering with "§4admin", an empty string, or 200
-		// characters, all of which this handler would otherwise relay straight into the
+		// characters, all of which must not be relayed straight into the proxy's player
-		// proxy's player list.
+		// list. Screened in resolveHasJoined; this is the last-line guard.
 		if !mcUsernameRe.MatchString(prof.Name) {
 			w.WriteHeader(http.StatusNoContent)
 			return
@@ -320,9 +322,10 @@ func lookupPremiumName(ctx context.Context, username string) (bool, error) {
 
 // resolveHasJoined queries each configured source in priority order and returns the
 // first that validates the session (200 with a profile). 204 is "not my player". Any other
-// outcome (unreachable, another status, a body that is not a profile) skips the source too,
+// outcome (unreachable, another status, a body that is not a profile, or a profile this
-// but is logged with its tag and reported as failed: otherwise a dead or mistyped source
+// multiplexer will not emit) skips the source too, but is logged with its tag and reported
-// looks exactly like a player it does not know, and nobody finds out.
+// as failed: otherwise a dead or mistyped source looks exactly like a player it does not
+// know, and nobody finds out.
 func (a *API) resolveHasJoined(ctx context.Context, username, serverID, ip string) (prof *sessionProfile, src AuthSource, failed bool) {
 	for _, src := range a.AuthSources {
 		u := src.URL + "?username=" + url.QueryEscape(username) + "&serverId=" + url.QueryEscape(serverID)
@@ -362,6 +365,23 @@ func (a *API) resolveHasJoined(ctx context.Context, username, serverID, ip strin
 			failed = true
 			continue
 		}
+		// Screen what a 200 is allowed to mean BEFORE it can win. A profile this
+		// multiplexer will not emit — an identity UUID that does not parse, a third-party
+		// name outside the Minecraft charset — is the same class as a body that is not a
+		// profile: skip, log, count as failed. Letting it win would stop the ladder on one
+		// sloppy root (every source behind it silently unreachable) and read as "nobody
+		// knows this player" to Velocity while a source had actually answered.
+		if src.Identity {
+			if _, perr := uuid.Parse(p.ID); perr != nil {
+				log.Printf("hasJoined: identity source %q answered 200 with an unparseable profile id %q", src.Tag, p.ID)
+				failed = true
+				continue
+			}
+		} else if !mcUsernameRe.MatchString(p.Name) {
+			log.Printf("hasJoined: source %q answered 200 with an unusable profile name %q", src.Tag, p.Name)
+			failed = true
+			continue
+		}
 		return &p, src, failed
 	}
 	return nil, AuthSource{}, failed

internal/api/handlers_hasjoined_test.go

@@ -2,12 +2,14 @@ package api
 
 import (
 	"bufio"
+	"bytes"
 	"context"
 	"encoding/hex"
 	"encoding/json"
 	"errors"
 	"fmt"
 	"io"
+	"log"
 	"net"
 	"net/http"
 	"net/http/httptest"
@@ -337,14 +339,26 @@ func TestHasJoined(t *testing.T) {
 		}
 	})
 
-	// Mojang is trusted for its UUIDs, which is exactly why one that does not parse must not
+	// Mojang is trusted for its UUIDs, which is exactly why one that does not parse must
-	// be emitted as some default: every such login would share the nil UUID.
+	// not be emitted as some default: every such login would share the nil UUID. The
-	t.Run("identity source with an unparseable id -> 204", func(t *testing.T) {
+	// unusable 200 is surfaced like every other bad answer — skipped, logged, and 503 when
+	// nothing else validates — instead of reading as "no such session".
+	t.Run("identity source with an unparseable id -> skipped and surfaced", func(t *testing.T) {
 		mojang := fakeYgg(t, "not-a-uuid", "Notch")
 		api := newTestAPI(newFakeRepo(), newFakeCluster())
 		api.AuthSources = []AuthSource{{Tag: "mojang", URL: mojang.URL, Identity: true}}
-		if w := getHasJoined(api.InternalHandler(), "Notch", "abc"); w.Code != http.StatusNoContent {
+
-			t.Fatalf("code = %d, want 204 (%q)", w.Code, w.Body.String())
+		var buf bytes.Buffer
+		prev := log.Writer()
+		log.SetOutput(&buf)
+		w := getHasJoined(api.InternalHandler(), "Notch", "abc")
+		log.SetOutput(prev)
+
+		if w.Code != http.StatusServiceUnavailable {
+			t.Fatalf("code = %d, want 503 (%q)", w.Code, w.Body.String())
+		}
+		if !strings.Contains(buf.String(), "unparseable profile id") {
+			t.Fatalf("the skip was not logged: %q", buf.String())
 		}
 	})
 
@@ -559,16 +573,51 @@ func TestHasJoinedPremiumNameRename(t *testing.T) {
 		}
 	})
 
-	// A third-party root is untrusted input, its name field included.
+	// A third-party root is untrusted input, its name field included. An unusable name is
-	t.Run("hostile upstream name -> 204", func(t *testing.T) {
+	// refused AND surfaced: skipped like every other bad answer (logged, counted as
+	// failed), so it can never be relayed and can never be mistaken for "no such player".
+	t.Run("hostile upstream name is skipped, logged and surfaced", func(t *testing.T) {
 		stubMojangNames(t)
 		for _, bad := range []string{"§4admin", "not a name", "ab", strings.Repeat("x", 17), ""} {
 			third := fakeYgg(t, notchMojangID, bad)
 			api := newTestAPI(newFakeRepo(), newFakeCluster())
 			api.AuthSources = []AuthSource{{Tag: "littleskin", Prefix: "LS", URL: third.URL, Identity: false}}
-			if w := getHasJoined(api.InternalHandler(), "Notch", "abc"); w.Code != http.StatusNoContent {
+
-				t.Fatalf("upstream name %q: code = %d, want 204 (%q)", bad, w.Code, w.Body.String())
+			var buf bytes.Buffer
+			prev := log.Writer()
+			log.SetOutput(&buf)
+			w := getHasJoined(api.InternalHandler(), "Notch", "abc")
+			log.SetOutput(prev)
+
+			if w.Code != http.StatusServiceUnavailable {
+				t.Fatalf("upstream name %q: code = %d, want 503 (%q)", bad, w.Code, w.Body.String())
+			}
+			if !strings.Contains(buf.String(), "unusable profile name") {
+				t.Fatalf("upstream name %q: the skip was not logged: %q", bad, buf.String())
 			}
 		}
 	})
+
+	// The reason the skip must CONTINUE the ladder rather than stop it: a broken root early
+	// in the list would otherwise silently block a login its later source would have
+	// validated (live on the audit box, the bad name produced a silent 204 and the valid
+	// source was never asked).
+	t.Run("hostile name cannot block a later source", func(t *testing.T) {
+		stubMojangNames(t)
+		bad := fakeYgg(t, notchMojangID, "§4admin")
+		good := fakeYgg(t, notchMojangID, "Notch")
+		api := newTestAPI(newFakeRepo(), newFakeCluster())
+		api.AuthSources = []AuthSource{
+			{Tag: "broken", Prefix: "BR", URL: bad.URL, Identity: false},
+			{Tag: "littleskin", Prefix: "LS", URL: good.URL, Identity: false},
+		}
+		w := getHasJoined(api.InternalHandler(), "Notch", "abc")
+		if w.Code != http.StatusOK {
+			t.Fatalf("code = %d, want 200 from the second source (%q)", w.Code, w.Body.String())
+		}
+		p := profileOf(t, w)
+		if want := undashed(uuid.NewMD5(felisAuthNS, []byte("littleskin:"+notchMojangID))); p.ID != want {
+			t.Fatalf("id = %q, want the second source's canonical %q", p.ID, want)
+		}
+	})
 }

internal/api/handlers_internal.go

@@ -240,6 +240,12 @@ func (a *API) handleInternalClaim(w http.ResponseWriter, r *http.Request) {
 	// 412 above; a lost race (0 rows) is 409.
 	claimed, err := a.Repo.ClaimServer(r.Context(), name, userID)
 	if err != nil {
+		// Same atomic quota gate as the external face (audit #4): the concurrent
+		// loser gets the sequential 403, never an over-provisioned tenant.
+		if errors.Is(err, ErrQuotaExceeded) {
+			writeError(w, r, newError(http.StatusForbidden, "quota_exceeded", "server quota exhausted"))
+			return
+		}
 		a.writeLookupError(w, r, err)
 		return
 	}

internal/api/handlers_logstream_test.go

@@ -486,8 +486,9 @@ func TestServerConsoleDisconnectTeardown(t *testing.T) {
 	}}
 	api.Logs = streamer
 
+	rec := httptest.NewRecorder()
+	w := &firstDataWriter{ResponseWriter: rec, data: make(chan struct{})}
 	r := httptest.NewRequest("GET", "/api/v1/servers/survival/console", nil).WithContext(ctx)
-	w := httptest.NewRecorder()
 
 	done := make(chan struct{})
 	go func() {
@@ -502,6 +503,15 @@ func TestServerConsoleDisconnectTeardown(t *testing.T) {
 	case <-time.After(2 * time.Second):
 		t.Fatal("relay never read the first log line")
 	}
+	// The first read is not enough: the relay still has to WRITE the event, and
+	// cancelling in that gap raced the write against the teardown (a live CI flake
+	// left the body empty). Wait for the write itself — the signal is closed after
+	// the recorder's Write returns, so the body read below happens-after it.
+	select {
+	case <-w.data:
+	case <-time.After(2 * time.Second):
+		t.Fatal("relay never wrote the first data event to the response")
+	}
 	cancel()
 
 	// The handler must return promptly...
@@ -525,8 +535,32 @@ func TestServerConsoleDisconnectTeardown(t *testing.T) {
 	default:
 		t.Fatal("StreamLogs did not receive the cancellable request context (gotCtx not Done)")
 	}
-	if !strings.Contains(w.Body.String(), "data: boot progress 50%\n\n") {
+	if !strings.Contains(rec.Body.String(), "data: boot progress 50%\n\n") {
-		t.Fatalf("expected the first event before disconnect, got %q", w.Body.String())
+		t.Fatalf("expected the first event before disconnect, got %q", rec.Body.String())
+	}
+}
+
+// firstDataWriter signals once the relay has written a `data:` event into the
+// wrapped recorder, so a test can disconnect only after the event is actually
+// observable in the body — inspecting p per Write is race-free because the relay
+// is the lone writer and each relay event is a single Write.
+type firstDataWriter struct {
+	http.ResponseWriter
+	data chan struct{}
+	once sync.Once
+}
+
+func (b *firstDataWriter) Write(p []byte) (int, error) {
+	n, err := b.ResponseWriter.Write(p)
+	if strings.HasPrefix(string(p), "data:") {
+		b.once.Do(func() { close(b.data) })
+	}
+	return n, err
+}
+
+func (b *firstDataWriter) Flush() {
+	if f, ok := b.ResponseWriter.(http.Flusher); ok {
+		f.Flush()
 	}
 }
 

internal/api/handlers_onboard.go

@@ -38,11 +38,11 @@ import (
 // in-game identity is the root of trust, so re-minting a code always re-grants a
 // session even after email/passkey are bound. "登录并非强制，但没登录什么都干不了".
 //
-// op.console stays behind Zero Trust. A code whose UUID belongs to STAFF (role=admin)
+// op.console stays behind Zero Trust. A code whose UUID belongs to STAFF (admin or
-// is refused here (ErrPlayerBindForbidden → 403), so the public bootstrap provably
+// owner) is refused here (ErrPlayerBindForbidden → 403), so the public bootstrap
-// never mints a session for an admin identity — the sole tier it yields is a role=user
+// provably never mints a session for a staff identity — the sole tier it yields is a
-// player session, host-only to console.<root_domain> (never sent to op.console) and
+// role=user player session, host-only to console.<root_domain> (never sent to
-// carrying ViaAdminAccess=false. "op.console 必须得 Auth".
+// op.console) and carrying ViaAdminAccess=false. "op.console 必须得 Auth".
 
 // newUserID returns an opaque random user id (128 bits, hex), matching the shape of
 // the ids break-glass mints for staff rows.
@@ -109,6 +109,14 @@ func (a *API) handleBindRedeem(w http.ResponseWriter, r *http.Request) {
 		writeError(w, r, newError(http.StatusForbidden, "staff_account",
 			"that Minecraft account belongs to staff; sign in at the operator console"))
 		return
+	case errors.Is(err, ErrPlayerAccountRetired):
+		// The linked Felis account is disabled or soft-deleted: the door refuses to
+		// reuse it, because minting a session here would resurrect the account the
+		// owner just retired (audit #33). The code survives, so re-enabling the
+		// account and retrying within its TTL still works.
+		writeError(w, r, newError(http.StatusForbidden, "account_retired",
+			"this Minecraft account's Felis account is disabled or deleted; contact the operator"))
+		return
 	case err != nil:
 		writeError(w, r, err)
 		return

internal/api/handlers_op_login.go

@@ -28,7 +28,7 @@ import (
 //     column keeps it from ever colliding with a console login_email or onboard code).
 //   - An in-game vouch — an already-trusted admin who is ONLINE approves the pending
 //     request via velocity's /felis command (internal approve). The API's own user
-//     table is the sole authority: only a UUID linked to a role=admin account may
+//     table is the sole authority: only a UUID linked to a staff account may
 //     approve (velocity's command runs for any player and relies on this check).
 //
 // finish mints the session only when BOTH have landed. Neither factor alone — a mailed
@@ -137,9 +137,10 @@ func (a *API) handleOpLoginStart(w http.ResponseWriter, r *http.Request) {
 		writeError(w, r, err)
 		return
 	}
-	// op.console is the STAFF door: a non-admin who typed their address here (they belong
+	// op.console is the STAFF door: a player who typed their address here (they belong
 	// on console.<root_domain>) gets the neutral response, never a request or a code.
-	if u.Role != "admin" {
+	// Staff means admin OR owner — the Owner is the primary op.console user.
+	if !staffRole(u.Role) {
 		neutral()
 		return
 	}
@@ -290,15 +291,15 @@ func (a *API) handleOpLoginFinish(w http.ResponseWriter, r *http.Request) {
 		writeError(w, r, err)
 		return
 	}
-	// Load the staff account for the session + response. Re-assert admin as defence in
+	// Load the staff account for the session + response. Re-assert staff as defence in
-	// depth: only admins ever get a request minted, but the session must never be issued
+	// depth: only staff ever get a request minted, but the session must never be issued
-	// to a non-admin identity even if the row were somehow otherwise.
+	// to a non-staff identity even if the row were somehow otherwise.
 	u, err := a.Repo.UserByID(r.Context(), loginReq.UserID)
 	if err != nil {
 		writeError(w, r, err)
 		return
 	}
-	if u.Role != "admin" {
+	if !staffRole(u.Role) {
 		writeError(w, r, newError(http.StatusForbidden, "staff_account", "that account is not an operator"))
 		return
 	}
@@ -343,7 +344,7 @@ func (a *API) handleOpLoginPending(w http.ResponseWriter, r *http.Request) {
 
 // opLoginApproveRequest is the internal approve body: the online-mode UUID of the
 // in-game admin running /felis web op approve. The API resolves it to a linked account
-// and refuses unless that account is role=admin — this check against the API's
+// and refuses unless that account is staff (admin or owner) — this check against the API's
 // authoritative user table is the only gate; velocity's command itself is unprivileged.
 type opLoginApproveRequest struct {
 	ApproverUUID string `json:"approver_uuid"`
@@ -351,10 +352,10 @@ type opLoginApproveRequest struct {
 
 // handleOpLoginApprove records an in-game admin's vouch for a pending staff login
 // (internal face), supplying the second factor. It resolves the approver UUID to a
-// linked role=admin account (else 403), then flips the request approved. A missing or
+// linked staff account (admin or owner; else 403), then flips the request approved.
-// no-longer-pending request is 404. Self-approval is allowed: a staff member online as
+// A missing or no-longer-pending request is 404. Self-approval is allowed: a staff
-// their own admin identity supplies a genuine second factor (in-game session control)
+// member online as their own admin identity supplies a genuine second factor
-// distinct from the mailbox factor.
+// (in-game session control) distinct from the mailbox factor.
 func (a *API) handleOpLoginApprove(w http.ResponseWriter, r *http.Request) {
 	id := r.PathValue("id")
 	var req opLoginApproveRequest

internal/api/handlers_op_login_test.go

@@ -170,6 +170,44 @@ func TestOpLoginVertical(t *testing.T) {
 	}
 }
 
+// TestOpLoginOwnerAdmitted pins that the staff door admits the Owner (role=owner),
+// not just plain admins: the Owner is the primary op.console identity, so a
+// role check of "admin only" would strand it outside its own console.
+func TestOpLoginOwnerAdmitted(t *testing.T) {
+	repo := newFakeRepo()
+	repo.settings[LocalAuthEnabledKey] = []byte("true")
+	repo.staff["owner"] = &StaffUser{
+		ID: "o1", Username: "owner", Email: "[email protected]",
+		Role: "owner", EmailVerified: true,
+	}
+	repo.links[opUUID] = "o1"
+	mailer := &captureMailer{}
+	api := newTestAPI(repo, newFakeCluster())
+	api.Mailer = mailer
+	eh := api.ExternalHandler()
+	ih := api.InternalHandler()
+
+	w := startOp(eh, "[email protected]")
+	if w.Code != http.StatusAccepted {
+		t.Fatalf("owner start: code = %d, want 202 (%s)", w.Code, w.Body.String())
+	}
+	reqID, _ := acctBody(t, w)["request_id"].(string)
+	if reqID == "" || mailer.calls != 1 || len(repo.opLogins) != 1 {
+		t.Fatalf("owner start must mint a request + mail a code: req=%q mails=%d rows=%d",
+			reqID, mailer.calls, len(repo.opLogins))
+	}
+	if w := approveOp(ih, reqID, opUUID); w.Code != http.StatusOK {
+		t.Fatalf("approve: code = %d (%s)", w.Code, w.Body.String())
+	}
+	w = finishOp(eh, reqID, mailer.code)
+	if w.Code != http.StatusOK {
+		t.Fatalf("owner finish: code = %d body %s, want 200", w.Code, w.Body.String())
+	}
+	if vb := acctBody(t, w); vb["role"] != "owner" {
+		t.Fatalf("finish role = %v, want owner", vb["role"])
+	}
+}
+
 // TestOpLoginStartNeutral pins the start-side anti-enumeration contract: op.console is
 // the STAFF door, so a non-admin account AND an unknown address both get a 202 carrying
 // a request_id + expires_at, mint/mail nothing, and still burn the per-recipient

internal/api/handlers_passkey_discoverable.go

@@ -148,6 +148,13 @@ func (a *API) handlePasskeyLoginDiscoverableFinish(w http.ResponseWriter, r *htt
 		if err != nil {
 			return PasskeyUser{}, err
 		}
+		// A disabled or soft-deleted account must not complete a login even when it
+		// still holds a credential (UserByID is an unfiltered lookup shared with admin
+		// reads, so the liveness check lives here, at the door). Fail closed with the
+		// same opaque outcome as an unknown handle (audit #33).
+		if d, err := a.Repo.UserDetail(r.Context(), u.ID); err != nil || d.Disabled || d.DeletedAt != nil {
+			return PasskeyUser{}, ErrNotFound
+		}
 		creds, err := a.Repo.PasskeyCredentialsForUser(r.Context(), u.ID)
 		if err != nil {
 			return PasskeyUser{}, err

internal/api/handlers_patch_test.go

@@ -127,6 +127,26 @@ func TestPatchServerMemoryOverride(t *testing.T) {
 	}
 }
 
+// TestPatchServerPreservesStorageCache pins the storage dimension of the quota
+// aggregate: a resources patch cannot change storage, so the cached storage
+// must survive it — passing 0 would silently zero the owner's aggregate (the
+// cached columns are QuotaCheck's only input) from that patch onward.
+func TestPatchServerPreservesStorageCache(t *testing.T) {
+	api, repo, _, _ := newPatchAPI()
+	repo.byName["survival"].OwnerID = "u1"
+	repo.quota["u1"] = true
+	repo.serverResources["survival"] = ResourceSpec{StorageMB: 10240}
+
+	w := patchSurvival(api, `{"memory":"2Gi","resources":{"memory":"4Gi","cpu":"2"}}`)
+	if w.Code != http.StatusOK {
+		t.Fatalf("code = %d, want 200 (%s)", w.Code, w.Body.String())
+	}
+	got := repo.resourceUpdates["survival"]
+	if got.CPUMilli != 2000 || got.MemoryMB != 4096 || got.StorageMB != 10240 {
+		t.Fatalf("resource cache = %+v, want cpu 2000 / mem 4096 / storage preserved 10240", got)
+	}
+}
+
 // TestPatchServerRejections is the validation matrix: each malformed request is
 // rejected with the right status and stable error code, and (critically) NOTHING
 // reaches the cluster on a rejection — the analog of create's "no CRD written".

internal/api/handlers_user.go

@@ -142,6 +142,13 @@ func (a *API) handleClaim(w http.ResponseWriter, r *http.Request) {
 	// ③ atomic claim
 	claimed, err := a.Repo.ClaimServer(r.Context(), name, p.UserID)
 	if err != nil {
+		// The atomic gate re-checks quota under the per-user lock (audit #4): a
+		// concurrent claim that spent the last slot surfaces here, with the same
+		// 403 the pre-check gives sequentially.
+		if errors.Is(err, ErrQuotaExceeded) {
+			writeError(w, r, newError(http.StatusForbidden, "quota_exceeded", "server quota exhausted"))
+			return
+		}
 		a.writeLookupError(w, r, err)
 		return
 	}
@@ -252,7 +259,8 @@ func (a *API) handleFleet(w http.ResponseWriter, r *http.Request) {
 	owners, _ := a.Repo.ServerOwners(r.Context())
 	views := make([]fleetServerView, len(servers))
 	for i, s := range servers {
-		views[i] = fleetServerView{ServerInfo: s, Owner: owners[s.Name]}
+		views[i] = fleetServerView{ServerInfo: s, Owner: owners[s.Name],
+			System: naming.IsSystemServer(s.Name)}
 	}
 	writeJSON(w, http.StatusOK, map[string]any{"servers": views})
 }
@@ -260,13 +268,18 @@ func (a *API) handleFleet(w http.ResponseWriter, r *http.Request) {
 // fleetServerView is one row of the SysAdmin cockpit's fleet read: the CRD
 // lifecycle view (ServerInfo, §1 authority) with the owner's display identity
 // joined alongside. The embed keeps every lifecycle field flat in the JSON so the
-// shape is a strict superset of ServerInfo; Owner is the only addition.
+// shape is a strict superset of ServerInfo.
 type fleetServerView struct {
 	ServerInfo
 	// Owner is the claiming user's display identity (email, or username when the
 	// address is absent), or "" when the server is unclaimed or the best-effort
 	// owner lookup failed — the cockpit renders "" as "unclaimed".
 	Owner string `json:"owner,omitempty"`
+	// System marks a platform-provisioned system service (the login gate and the
+	// lobby, naming.IsSystemServer). Their names are reserved, so every per-server
+	// API route rejects them — the cockpit must render them read-only rather than
+	// offer claim/wake/stop/console actions that would answer 400.
+	System bool `json:"system,omitempty"`
 }
 
 // createServerRequest is the structured §15 create-server form. This is the
@@ -759,7 +772,19 @@ func (a *API) handlePatchServer(w http.ResponseWriter, r *http.Request) {
 			a.writeLookupError(w, r, err)
 			return
 		}
-		_ = a.Repo.UpdateServerResources(r.Context(), name, newCPU, newMemMB, 0)
+		// A resource patch cannot change storage, so its cached contribution must
+		// be preserved: passing 0 would silently zero the storage dimension of the
+		// owner's four-cap aggregate (the cached columns are its only input).
+		storMB := 0
+		if rec != nil {
+			cur, err := a.Repo.ServerResources(r.Context(), name)
+			if err != nil {
+				writeError(w, r, err)
+				return
+			}
+			storMB = cur.StorageMB
+		}
+		_ = a.Repo.UpdateServerResources(r.Context(), name, newCPU, newMemMB, storMB)
 	} else {
 		if err := a.Cluster.PatchServerSpec(r.Context(), name, patch); err != nil {
 			a.writeLookupError(w, r, err)

internal/api/handlers_users.go

@@ -140,6 +140,18 @@ func (a *API) handlePatchUser(w http.ResponseWriter, r *http.Request) {
 		return
 	}
 
+	// Owner protection (migration 0011): the owner row is the one identity the
+	// panel may never demote — only the local break-glass console resets it.
+	// Username/email edits on it stay allowed. A failed detail read falls through;
+	// UpdateUser then answers the real 404.
+	if body.Role != nil && *body.Role != "owner" {
+		if d, err := a.Repo.UserDetail(r.Context(), id); err == nil && d.Role == "owner" {
+			writeError(w, r, newError(http.StatusForbidden, "forbidden",
+				"the owner account's role cannot be changed from the panel"))
+			return
+		}
+	}
+
 	if body.Username != nil {
 		if err := validateUsername(*body.Username); err != nil {
 			writeError(w, r, err)
@@ -186,6 +198,14 @@ func (a *API) handleDeleteUser(w http.ResponseWriter, r *http.Request) {
 		return
 	}
 
+	// Same owner protection as the role guard above: only break-glass retires the
+	// owner identity. A failed detail read falls through to the real 404.
+	if d, err := a.Repo.UserDetail(r.Context(), id); err == nil && d.Role == "owner" {
+		writeError(w, r, newError(http.StatusForbidden, "forbidden",
+			"the owner account cannot be deleted from the panel"))
+		return
+	}
+
 	if err := a.Repo.DeleteUser(r.Context(), id, p.Email); err != nil {
 		if errors.Is(err, ErrNotFound) {
 			writeError(w, r, newError(http.StatusNotFound, "not_found", "user not found"))
@@ -222,6 +242,17 @@ func (a *API) handleDisableUser(w http.ResponseWriter, r *http.Request) {
 		return
 	}
 
+	// Owner protection (migration 0011): disabling locks the owner out and revokes
+	// its sessions — effectively a demotion, so the panel refuses it; only
+	// break-glass touches the owner identity. Re-enabling stays allowed.
+	if body.Disabled {
+		if d, err := a.Repo.UserDetail(r.Context(), id); err == nil && d.Role == "owner" {
+			writeError(w, r, newError(http.StatusForbidden, "forbidden",
+				"the owner account cannot be disabled from the panel"))
+			return
+		}
+	}
+
 	if err := a.Repo.SetUserDisabled(r.Context(), id, body.Disabled); err != nil {
 		if errors.Is(err, ErrNotFound) {
 			writeError(w, r, newError(http.StatusNotFound, "not_found", "user not found"))
@@ -250,6 +281,10 @@ func (a *API) handleGetQuotas(w http.ResponseWriter, r *http.Request) {
 	}
 	v, err := a.Repo.GetQuotas(r.Context(), id)
 	if err != nil {
+		if errors.Is(err, ErrNotFound) {
+			writeError(w, r, newError(http.StatusNotFound, "not_found", "user not found"))
+			return
+		}
 		writeError(w, r, err)
 		return
 	}
@@ -446,6 +481,10 @@ func (a *API) handleLinkAccount(w http.ResponseWriter, r *http.Request) {
 				"this UUID is already linked to a different user"))
 			return
 		}
+		if errors.Is(err, ErrNotFound) {
+			writeError(w, r, newError(http.StatusNotFound, "not_found", "user not found"))
+			return
+		}
 		writeError(w, r, err)
 		return
 	}

internal/api/handlers_users_test.go

@@ -0,0 +1,99 @@
+package api
+
+import (
+	"net/http"
+	"testing"
+)
+
+// The owner row is the one identity the panel may never demote, delete, or
+// disable (migration 0011) — only the local break-glass console resets it.
+// These guards became load-bearing the moment provisioning started actually
+// writing role='owner'; before that the owner tier was simply unreachable, so
+// nothing could reach them.
+func TestOwnerAccountProtectedFromPanelMutations(t *testing.T) {
+	owner := &Principal{UserID: "usr-root", Role: "owner", ViaAdminAccess: true}
+	repo := newFakeRepo()
+	repo.seedUser(UserView{ID: "usr-root", Username: "root", Role: "owner"})
+	repo.seedUser(UserView{ID: "usr-owner2", Username: "spare-owner", Role: "owner"})
+	repo.seedUser(UserView{ID: "u2", Username: "alice", Role: "user"})
+	api := newTestAPI(repo, newFakeCluster())
+	api.External = staticExternal{p: owner}
+	eh := api.ExternalHandler()
+
+	t.Run("role change refused", func(t *testing.T) {
+		w := do(eh, "PATCH", "/api/v1/users/usr-owner2", `{"role":"admin"}`, jsonHeader)
+		if w.Code != http.StatusForbidden {
+			t.Fatalf("demote owner: code = %d body %s, want 403", w.Code, w.Body.String())
+		}
+	})
+	t.Run("delete refused", func(t *testing.T) {
+		w := do(eh, "DELETE", "/api/v1/users/usr-owner2", "", nil)
+		if w.Code != http.StatusForbidden {
+			t.Fatalf("delete owner: code = %d body %s, want 403", w.Code, w.Body.String())
+		}
+	})
+	t.Run("disable refused", func(t *testing.T) {
+		w := do(eh, "POST", "/api/v1/users/usr-owner2/disable", `{"disabled":true}`, jsonHeader)
+		if w.Code != http.StatusForbidden {
+			t.Fatalf("disable owner: code = %d body %s, want 403", w.Code, w.Body.String())
+		}
+	})
+	t.Run("email edits on an owner stay allowed", func(t *testing.T) {
+		w := do(eh, "PATCH", "/api/v1/users/usr-owner2", `{"email":"[email protected]"}`, jsonHeader)
+		if w.Code != http.StatusOK {
+			t.Fatalf("edit owner email: code = %d body %s, want 200", w.Code, w.Body.String())
+		}
+	})
+	t.Run("control: a normal user can still be promoted", func(t *testing.T) {
+		w := do(eh, "PATCH", "/api/v1/users/u2", `{"role":"admin"}`, jsonHeader)
+		if w.Code != http.StatusOK {
+			t.Fatalf("promote user: code = %d body %s, want 200", w.Code, w.Body.String())
+		}
+	})
+}
+
+// The user-scoped admin sub-resources (quotas, account links) must answer 404
+// for an unknown user id. Before the requireLiveUser guard the quota upsert and
+// the link insert reached the users(id) foreign key and surfaced as an opaque
+// 500 (found live against the drill cluster, audit #30), and the quotas read
+// answered a zero-value "unlimited" view as if the id existed.
+func TestAdminSubresourcesRequireLiveUser(t *testing.T) {
+	owner := &Principal{UserID: "usr-root", Role: "owner", ViaAdminAccess: true}
+	repo := newFakeRepo()
+	repo.seedUser(UserView{ID: "usr-root", Username: "root", Role: "owner"})
+	repo.seedUser(UserView{ID: "u2", Username: "alice", Role: "user"})
+	api := newTestAPI(repo, newFakeCluster())
+	api.External = staticExternal{p: owner}
+	eh := api.ExternalHandler()
+
+	t.Run("quotas read of an unknown user", func(t *testing.T) {
+		w := do(eh, "GET", "/api/v1/users/usr-nope/quotas", "", nil)
+		if w.Code != http.StatusNotFound {
+			t.Fatalf("code = %d body %s, want 404", w.Code, w.Body.String())
+		}
+	})
+	t.Run("quotas write of an unknown user", func(t *testing.T) {
+		w := do(eh, "PUT", "/api/v1/users/usr-nope/quotas", `{"max_servers":1}`, jsonHeader)
+		if w.Code != http.StatusNotFound {
+			t.Fatalf("code = %d body %s, want 404", w.Code, w.Body.String())
+		}
+	})
+	t.Run("account link of an unknown user", func(t *testing.T) {
+		w := do(eh, "POST", "/api/v1/users/usr-nope/links",
+			`{"mc_uuid":"22222222-3333-4444-5555-666666666666"}`, jsonHeader)
+		if w.Code != http.StatusNotFound {
+			t.Fatalf("code = %d body %s, want 404", w.Code, w.Body.String())
+		}
+	})
+	t.Run("control: a live user still accepts both writes", func(t *testing.T) {
+		w := do(eh, "PUT", "/api/v1/users/u2/quotas", `{"max_servers":2}`, jsonHeader)
+		if w.Code != http.StatusOK {
+			t.Fatalf("set quotas: code = %d body %s, want 200", w.Code, w.Body.String())
+		}
+		w = do(eh, "POST", "/api/v1/users/u2/links",
+			`{"mc_uuid":"22222222-3333-4444-5555-666666666677"}`, jsonHeader)
+		if w.Code != http.StatusOK {
+			t.Fatalf("link: code = %d body %s, want 200", w.Code, w.Body.String())
+		}
+	})
+}

internal/api/images.go

@@ -15,6 +15,11 @@ import (
 // admin-tier (Zero Trust), enforced by adminOnly before these handlers run.
 type ImageBuilder interface {
 	Submit(ctx context.Context, req build.Request) (*build.Build, error)
+	// Get reads one build row without reconciling it — the read-only lookup the
+	// submission views use to surface a build's outcome to its submitter (the
+	// /images/build routes are admin-tier). State advance belongs to the
+	// reconcile loop (Sync/SyncAll), so a list render never touches the cluster.
+	Get(ctx context.Context, id string) (*build.Build, error)
 	// Sync reconciles a build against its Job and returns the current view, so a
 	// GET doubles as the reconcile tick (idempotent on terminal builds).
 	Sync(ctx context.Context, id string) (*build.Build, error)

internal/api/images_test.go

@@ -16,6 +16,8 @@ import (
 type fakeBuilder struct {
 	submitted   *build.Request
 	submitErr   error
+	getBuilds   map[string]*build.Build
+	getErr      error
 	syncErr     error
 	cancelErr   error
 	addedRef    string
@@ -40,6 +42,17 @@ func (f *fakeBuilder) Submit(_ context.Context, req build.Request) (*build.Build
 		RequestedBy: req.RequestedBy}, nil
 }
 
+func (f *fakeBuilder) Get(_ context.Context, id string) (*build.Build, error) {
+	f.lastBuildID = id
+	if f.getErr != nil {
+		return nil, f.getErr
+	}
+	if b, ok := f.getBuilds[id]; ok {
+		return b, nil
+	}
+	return nil, build.ErrNotFound
+}
+
 func (f *fakeBuilder) Sync(_ context.Context, id string) (*build.Build, error) {
 	f.lastBuildID = id
 	if f.syncErr != nil {

internal/api/jobstatus.go

@@ -0,0 +1,65 @@
+package api
+
+import (
+	"context"
+	"net/http"
+	"time"
+
+	"felis.lolicon.best/internal/naming"
+)
+
+// AsyncJob is the observable outcome of one asynchronous world operation. The API
+// only ENQUEUES backup/restore Jobs — the work, and any failure, happens in the
+// cluster — so without this projection a failed job left its only trace in a Job
+// object an operator with kubectl could read. The route is the API-side outlet.
+type AsyncJob struct {
+	Name       string    `json:"name"`
+	Kind       string    `json:"kind"`  // "backup" | "restore"
+	State      string    `json:"state"` // "running" | "succeeded" | "failed"
+	Message    string    `json:"message,omitempty"`
+	StartedAt  time.Time `json:"started_at,omitempty"`
+	FinishedAt time.Time `json:"finished_at,omitempty"`
+}
+
+// JobStatusReader reads the newest backup/restore Jobs for a server, newest
+// first. Optional like Restorer/Backuper: when nil the route answers 503.
+type JobStatusReader interface {
+	LatestJobs(ctx context.Context, serverName string) ([]AsyncJob, error)
+}
+
+// handleServerJobs serves GET /api/v1/servers/{name}/jobs — the latest async world
+// operations for one server, so a 202 that later failed is visible without kubectl.
+// Authorization mirrors the backup/restore gates' front half (owner-or-admin); the
+// message is free-form Job text and can name paths the owner already sees through
+// the file editor.
+func (a *API) handleServerJobs(w http.ResponseWriter, r *http.Request) {
+	p := principalFromContext(r.Context())
+	name := r.PathValue("name")
+	if err := naming.ValidateServerName(name); err != nil {
+		writeError(w, r, newError(http.StatusBadRequest, "bad_name", "invalid server name: %v", err))
+		return
+	}
+	rec, err := a.Repo.ServerByName(r.Context(), name)
+	if err != nil {
+		a.writeLookupError(w, r, err)
+		return
+	}
+	if !a.isOwnerOrAdmin(p, rec) {
+		writeError(w, r, errForbidden)
+		return
+	}
+	if a.JobStatus == nil {
+		writeError(w, r, newError(http.StatusServiceUnavailable, "jobs_unavailable",
+			"job status is not configured"))
+		return
+	}
+	jobs, err := a.JobStatus.LatestJobs(r.Context(), name)
+	if err != nil {
+		writeError(w, r, err)
+		return
+	}
+	if jobs == nil {
+		jobs = []AsyncJob{}
+	}
+	writeJSON(w, http.StatusOK, map[string]any{"server": name, "jobs": jobs})
+}

internal/api/jobstatus_test.go

@@ -0,0 +1,147 @@
+package api
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"net/http"
+	"testing"
+	"time"
+
+	batchv1 "k8s.io/api/batch/v1"
+	corev1 "k8s.io/api/core/v1"
+	metav1 "k8s.io/apimachinery/pkg/apis/meta/v1"
+)
+
+type fakeJobStatus struct {
+	jobs []AsyncJob
+	err  error
+	got  string
+}
+
+func (f *fakeJobStatus) LatestJobs(_ context.Context, server string) ([]AsyncJob, error) {
+	f.got = server
+	return f.jobs, f.err
+}
+
+// TestServerJobsHandler covers GET /api/v1/servers/{name}/jobs: the owner sees the
+// recorded outcomes, strangers are refused, a nil reader is an honest 503, and a
+// reader error surfaces as 500.
+func TestServerJobsHandler(t *testing.T) {
+	owner := &Principal{UserID: "owner1", Email: "[email protected]", Role: "user"}
+	stranger := &Principal{UserID: "other", Email: "[email protected]", Role: "user"}
+
+	mk := func(reader JobStatusReader) (*API, *fakeJobStatus) {
+		repo := newFakeRepo()
+		repo.byName["survival"] = &ServerRecord{Name: "survival", OwnerID: "owner1"}
+		a := newTestAPI(repo, newFakeCluster())
+		a.External = staticExternal{p: owner}
+		var fjs *fakeJobStatus
+		if reader != nil {
+			a.JobStatus = reader
+		}
+		if fj, ok := reader.(*fakeJobStatus); ok {
+			fjs = fj
+		}
+		return a, fjs
+	}
+
+	t.Run("owner sees failed job", func(t *testing.T) {
+		fjs := &fakeJobStatus{jobs: []AsyncJob{{Name: "restore-survival", Kind: "restore", State: "failed", Message: "backoff limit exceeded"}}}
+		a, _ := mk(fjs)
+		w := do(a.ExternalHandler(), "GET", "/api/v1/servers/survival/jobs", "", nil)
+		if w.Code != http.StatusOK {
+			t.Fatalf("code = %d, want 200 (%s)", w.Code, w.Body.String())
+		}
+		var resp struct {
+			Server string     `json:"server"`
+			Jobs   []AsyncJob `json:"jobs"`
+		}
+		if err := json.Unmarshal(w.Body.Bytes(), &resp); err != nil {
+			t.Fatalf("bad JSON: %v", err)
+		}
+		if resp.Server != "survival" || len(resp.Jobs) != 1 || resp.Jobs[0].State != "failed" {
+			t.Fatalf("unexpected payload %+v", resp)
+		}
+		if fjs.got != "survival" {
+			t.Fatalf("reader asked for %q", fjs.got)
+		}
+	})
+
+	t.Run("stranger -> 403", func(t *testing.T) {
+		fjs := &fakeJobStatus{}
+		a, _ := mk(fjs)
+		a.External = staticExternal{p: stranger}
+		if w := do(a.ExternalHandler(), "GET", "/api/v1/servers/survival/jobs", "", nil); w.Code != http.StatusForbidden {
+			t.Fatalf("code = %d, want 403", w.Code)
+		}
+	})
+
+	t.Run("nil reader -> 503", func(t *testing.T) {
+		a, _ := mk(nil)
+		if w := do(a.ExternalHandler(), "GET", "/api/v1/servers/survival/jobs", "", nil); w.Code != http.StatusServiceUnavailable {
+			t.Fatalf("code = %d, want 503", w.Code)
+		}
+	})
+
+	t.Run("reader error -> 500", func(t *testing.T) {
+		fjs := &fakeJobStatus{err: errors.New("apiserver down")}
+		a, _ := mk(fjs)
+		if w := do(a.ExternalHandler(), "GET", "/api/v1/servers/survival/jobs", "", nil); w.Code != http.StatusInternalServerError {
+			t.Fatalf("code = %d, want 500", w.Code)
+		}
+	})
+
+	t.Run("unknown server -> 404", func(t *testing.T) {
+		a, _ := mk(&fakeJobStatus{})
+		if w := do(a.ExternalHandler(), "GET", "/api/v1/servers/ghost/jobs", "", nil); w.Code != http.StatusNotFound {
+			t.Fatalf("code = %d, want 404", w.Code)
+		}
+	})
+}
+
+// TestJobToAsyncJob pins the Job→AsyncJob projection: kinds come from managed-by,
+// terminal conditions decide state, and unknown owners are dropped.
+func TestJobToAsyncJob(t *testing.T) {
+	start := metav1.NewTime(time.Date(2026, 9, 22, 10, 0, 0, 0, time.UTC))
+	done := metav1.NewTime(time.Date(2026, 9, 22, 10, 5, 0, 0, time.UTC))
+
+	failed := &batchv1.Job{
+		ObjectMeta: metav1.ObjectMeta{Name: "restore-survival", Labels: map[string]string{jobManagedByLabel: jobManagedByRestore}},
+		Status: batchv1.JobStatus{
+			StartTime: &start,
+			Conditions: []batchv1.JobCondition{{
+				Type: batchv1.JobFailed, Status: corev1.ConditionTrue,
+				Reason: "BackoffLimitExceeded", Message: "Job has reached the specified backoff limit",
+				LastTransitionTime: done,
+			}},
+		},
+	}
+	aj, ok := jobToAsyncJob(failed)
+	if !ok || aj.Kind != "restore" || aj.State != "failed" || aj.Name != "restore-survival" {
+		t.Fatalf("failed job projection = %+v ok=%v", aj, ok)
+	}
+	if !aj.FinishedAt.Equal(done.Time) || !aj.StartedAt.Equal(start.Time) {
+		t.Fatalf("timestamps = %+v", aj)
+	}
+
+	complete := &batchv1.Job{
+		ObjectMeta: metav1.ObjectMeta{Name: "backup-survival-1", Labels: map[string]string{jobManagedByLabel: jobManagedByBackup}},
+		Status: batchv1.JobStatus{Conditions: []batchv1.JobCondition{{
+			Type: batchv1.JobComplete, Status: corev1.ConditionTrue, LastTransitionTime: done,
+		}}},
+	}
+	if aj, ok := jobToAsyncJob(complete); !ok || aj.Kind != "backup" || aj.State != "succeeded" {
+		t.Fatalf("complete job projection = %+v ok=%v", aj, ok)
+	}
+
+	running := &batchv1.Job{ObjectMeta: metav1.ObjectMeta{Labels: map[string]string{jobManagedByLabel: jobManagedByBackup}}}
+	if aj, ok := jobToAsyncJob(running); !ok || aj.State != "running" {
+		t.Fatalf("running job projection = %+v ok=%v", aj, ok)
+	}
+
+	foreign := &batchv1.Job{ObjectMeta: metav1.ObjectMeta{Labels: map[string]string{jobManagedByLabel: "someone-else"}}}
+	if _, ok := jobToAsyncJob(foreign); ok {
+		t.Fatal("foreign job must be dropped")
+	}
+}

internal/api/k8scluster.go

@@ -43,6 +43,22 @@ func (k *K8sCluster) GetServer(ctx context.Context, name string) (*ServerInfo, e
 	return serverInfo(&ms), nil
 }
 
+// WorldVolumeExists reads the world PVC the operator's StatefulSet
+// volumeClaimTemplate creates (naming.WorldPVCName — the same name the backup
+// and restore Jobs mount), so existence here is exactly existence at Job mount
+// time. NotFound is (false, nil): the caller refuses with a specific 409.
+func (k *K8sCluster) WorldVolumeExists(ctx context.Context, name string) (bool, error) {
+	var pvc corev1.PersistentVolumeClaim
+	err := k.c.Get(ctx, types.NamespacedName{Namespace: k.namespace, Name: naming.WorldPVCName(name)}, &pvc)
+	if apierrors.IsNotFound(err) {
+		return false, nil
+	}
+	if err != nil {
+		return false, err
+	}
+	return true, nil
+}
+
 func (k *K8sCluster) GetBySubdomain(ctx context.Context, subdomain string) (*ServerInfo, error) {
 	var list v1alpha1.MinecraftServerList
 	if err := k.c.List(ctx, &list, client.InNamespace(k.namespace)); err != nil {

internal/api/k8sjobstatus.go

@@ -0,0 +1,92 @@
+package api
+
+import (
+	"context"
+	"sort"
+
+	batchv1 "k8s.io/api/batch/v1"
+	corev1 "k8s.io/api/core/v1"
+	"sigs.k8s.io/controller-runtime/pkg/client"
+)
+
+// Labels the backup/restore executors apply (internal/backupjob and
+// internal/restore keep their own copies). Literals on purpose: api defines the
+// read seam and must not import the executors — their dependency direction is
+// "executors implement api's interfaces", never the reverse.
+const (
+	jobServerLabel    = "felis.lolicon.best/server"
+	jobManagedByLabel = "app.kubernetes.io/managed-by"
+
+	jobManagedByBackup  = "felis-backup"
+	jobManagedByRestore = "felis-restore"
+)
+
+// K8sJobStatus reads the async Jobs the executors created, by the server label
+// both apply.
+type K8sJobStatus struct {
+	c         client.Client
+	namespace string
+}
+
+// NewK8sJobStatus builds the reader over the cluster client and the namespace the
+// server workloads (and their Jobs) live in.
+func NewK8sJobStatus(c client.Client, namespace string) *K8sJobStatus {
+	return &K8sJobStatus{c: c, namespace: namespace}
+}
+
+// LatestJobs lists this server's backup/restore Jobs newest-first, capped so a
+// long history cannot balloon the response. Jobs the label selector catches but
+// another component created (unknown managed-by) are dropped.
+func (k *K8sJobStatus) LatestJobs(ctx context.Context, serverName string) ([]AsyncJob, error) {
+	var list batchv1.JobList
+	if err := k.c.List(ctx, &list, client.InNamespace(k.namespace),
+		client.MatchingLabels{jobServerLabel: serverName}); err != nil {
+		return nil, err
+	}
+	out := make([]AsyncJob, 0, len(list.Items))
+	for i := range list.Items {
+		if aj, ok := jobToAsyncJob(&list.Items[i]); ok {
+			out = append(out, aj)
+		}
+	}
+	sort.Slice(out, func(i, j int) bool { return out[i].StartedAt.After(out[j].StartedAt) })
+	if len(out) > 20 {
+		out = out[:20]
+	}
+	return out, nil
+}
+
+// jobToAsyncJob projects one Job onto its kind/state/message. Complete condition →
+// succeeded, Failed → failed with its reason (Job conditions carry the generic
+// "backoff limit exceeded" text; the pod log holds the underlying error), anything
+// else is still running.
+func jobToAsyncJob(j *batchv1.Job) (AsyncJob, bool) {
+	kind := ""
+	switch j.Labels[jobManagedByLabel] {
+	case jobManagedByBackup:
+		kind = "backup"
+	case jobManagedByRestore:
+		kind = "restore"
+	default:
+		return AsyncJob{}, false
+	}
+	aj := AsyncJob{Name: j.Name, Kind: kind, State: "running", StartedAt: j.CreationTimestamp.Time}
+	if j.Status.StartTime != nil {
+		aj.StartedAt = j.Status.StartTime.Time
+	}
+	for _, c := range j.Status.Conditions {
+		switch {
+		case c.Type == batchv1.JobComplete && c.Status == corev1.ConditionTrue:
+			aj.State = "succeeded"
+			aj.FinishedAt = c.LastTransitionTime.Time
+		case c.Type == batchv1.JobFailed && c.Status == corev1.ConditionTrue:
+			aj.State = "failed"
+			aj.FinishedAt = c.LastTransitionTime.Time
+			aj.Message = c.Message
+			if aj.Message == "" {
+				aj.Message = c.Reason
+			}
+		}
+	}
+	return aj, true
+}

internal/api/metrics.go

@@ -0,0 +1,34 @@
+package api
+
+import (
+	"net/http"
+
+	"felis.lolicon.best/internal/metrics"
+
+	"github.com/prometheus/client_golang/prometheus"
+	"github.com/prometheus/client_golang/prometheus/promhttp"
+)
+
+// apiMetricsHandler serves the felis_* collectors from a process-local registry.
+// The API process is the only producer of felis_image_build_failures_total (the
+// build reconcile loop lives in cmd/felis.reconcileBuilds), so this endpoint is
+// that counter's sole scrape path; the operator's :8080 carries the fleet
+// gauges instead.
+var apiMetricsHandler = newAPIMetricsHandler()
+
+func newAPIMetricsHandler() http.Handler {
+	reg := prometheus.NewRegistry()
+	// A fresh registry cannot already hold a collector, so Register's
+	// AlreadyRegistered tolerance arm never triggers here.
+	_ = metrics.Register(reg)
+	return promhttp.HandlerFor(reg, promhttp.HandlerOpts{})
+}
+
+// handleMetrics is mounted as a Public route on the internal face (the same
+// stance as the health probes): the internal listener is ClusterIP-only and a
+// Prometheus scrape carries no token. The external face never serves metrics.
+// See docs/troubleshooting.md §14 for what to scrape and the alert rules that
+// consume these series.
+func (a *API) handleMetrics(w http.ResponseWriter, r *http.Request) {
+	apiMetricsHandler.ServeHTTP(w, r)
+}

internal/api/metrics_test.go

@@ -0,0 +1,44 @@
+package api
+
+import (
+	"net/http"
+	"strings"
+	"testing"
+
+	"felis.lolicon.best/internal/metrics"
+)
+
+// TestMetricsEndpoint locks the scrape surface: the internal face serves the
+// felis_* collectors (the API process is felis_image_build_failures_total's
+// only producer), and the external face never serves metrics.
+func TestMetricsEndpoint(t *testing.T) {
+	// Give every collector a sample so the exposition carries all four families:
+	// an empty vector family is omitted from the text output entirely. The bumps
+	// stay inside this test binary — the collectors are per-process globals.
+	metrics.SyncServerGauge([]string{"Running"})
+	metrics.StartDurationSeconds.Observe(3)
+	metrics.ImageBuildFailuresTotal.Inc()
+	metrics.ReaperWorldsDeletedTotal.Inc()
+
+	a := &API{}
+
+	w := do(a.InternalHandler(), "GET", "/metrics", "", nil)
+	if w.Code != http.StatusOK {
+		t.Fatalf("internal GET /metrics = %d, want 200", w.Code)
+	}
+	body := w.Body.String()
+	for _, want := range []string{
+		`felis_servers_total{state="Running"} 1`,
+		"felis_start_duration_seconds_count 1",
+		"felis_image_build_failures_total",
+		"felis_reaper_worlds_deleted_total",
+	} {
+		if !strings.Contains(body, want) {
+			t.Errorf("metrics exposition missing %q", want)
+		}
+	}
+
+	if w := do(a.ExternalHandler(), "GET", "/metrics", "", nil); w.Code != http.StatusNotFound {
+		t.Fatalf("external GET /metrics = %d, want 404 (metrics stay internal)", w.Code)
+	}
+}

internal/api/middleware.go

@@ -4,6 +4,7 @@ import (
 	"context"
 	"crypto/rand"
 	"encoding/hex"
+	"errors"
 	"log"
 	"net/http"
 	"runtime/debug"
@@ -95,6 +96,11 @@ func (a *API) requireInternal(next http.Handler) http.Handler {
 func (a *API) requireExternal(next http.Handler) http.Handler {
 	return http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
 		p, err := a.External.Authenticate(r)
+		if errors.Is(err, errAuthBackend) {
+			// Session store unreachable — an outage, not a missing credential.
+			writeError(w, r, errAuthUnavailable)
+			return
+		}
 		if err != nil || p == nil {
 			writeError(w, r, errUnauthorized)
 			return
@@ -118,7 +124,8 @@ func (a *API) requireExternal(next http.Handler) http.Handler {
 
 // adminOnly gates an external-face handler on the admin Zero-Trust path. The
 // Access middleware has already authenticated; this enforces that admin-tier
-// operations both carry role=admin AND arrived via admin.* (spec §14).
+// operations both carry a staff role (admin or owner) AND arrived via admin.*
+// (spec §14).
 func (a *API) adminOnly(next http.HandlerFunc) http.HandlerFunc {
 	return func(w http.ResponseWriter, r *http.Request) {
 		if p := principalFromContext(r.Context()); !p.IsAdmin() {

internal/api/pgrepo.go

@@ -94,8 +94,13 @@ func (p *PGRepo) VerifyLinkCode(ctx context.Context, userID, code string, now ti
 		return "", "", err
 	}
 
-	// If this UUID is already linked, only the same user may re-verify (idempotent);
+	// If this UUID is already linked, only the same user may re-verify (idempotent).
-	// a different user is a conflict and must not consume the code.
+	// A different LIVE user is a conflict and must not consume the code. A link whose
+	// account was soft-deleted is the exception: the identity is unclaimed (the
+	// account is gone; e.g. a migrated source, whose retire keeps the link but is
+	// otherwise dead), and the fresh in-game code proves the caller still holds this
+	// UUID, so the live caller takes the link over. Disabled-but-not-deleted stays a
+	// conflict — taking over a locked account's identity would bypass the lockout.
 	var existingUser string
 	switch err := tx.QueryRowContext(ctx,
 		`SELECT user_id FROM account_links WHERE mc_uuid = $1`, mcUUID).Scan(&existingUser); {
@@ -105,7 +110,21 @@ func (p *PGRepo) VerifyLinkCode(ctx context.Context, userID, code string, now ti
 		return "", "", err
 	default:
 		if existingUser != userID {
-			return "", "", ErrConflict
+			var linkedDeleted bool
+			if err := tx.QueryRowContext(ctx,
+				`SELECT deleted_at IS NOT NULL FROM users WHERE id = $1`,
+				existingUser).Scan(&linkedDeleted); err != nil {
+				return "", "", err
+			}
+			if !linkedDeleted {
+				return "", "", ErrConflict
+			}
+			if _, err := tx.ExecContext(ctx,
+				`UPDATE account_links SET user_id = $1, auth_source = $2, verified_at = now()
+				  WHERE mc_uuid = $3`,
+				userID, authSource, mcUUID); err != nil {
+				return "", "", fmt.Errorf("take over retired link: %w", err)
+			}
 		}
 	}
 
@@ -159,14 +178,19 @@ func (p *PGRepo) RedeemPlayerBindCode(ctx context.Context, newUserID, code strin
 	// together), where both transactions reach the inserts before either commits.
 
 	// Create-or-fetch keyed on the verified UUID. An already-linked role='user' player
-	// is fetched (idempotent "log in via the game"); a role='admin' STAFF account is
+	// is fetched (idempotent "log in via the game"); any STAFF account (admin or
-	// refused (op.console only) BEFORE any consume, so the code survives; an unlinked
+	// owner, i.e. role != 'user') is refused (op.console only) BEFORE any consume, so
-	// UUID births a fresh role='user' player with a uuid-derived unique username.
+	// the code survives; a DISABLED or soft-deleted account is refused the same way
+	// (audit #33 — a dead account must not resurrect through the bind door); an
+	// unlinked UUID births a fresh role='user' player with a uuid-derived unique
+	// username.
 	userID := newUserID
 	var existingRole string
+	var disabled, deleted bool
 	switch err := tx.QueryRowContext(ctx,
-		`SELECT u.id, u.role::text FROM account_links al JOIN users u ON u.id = al.user_id WHERE al.mc_uuid = $1`,
+		`SELECT u.id, u.role::text, u.disabled, u.deleted_at IS NOT NULL
-		mcUUID).Scan(&userID, &existingRole); {
+		   FROM account_links al JOIN users u ON u.id = al.user_id WHERE al.mc_uuid = $1`,
+		mcUUID).Scan(&userID, &existingRole, &disabled, &deleted); {
 	case errors.Is(err, sql.ErrNoRows):
 		if _, err := tx.ExecContext(ctx,
 			`INSERT INTO users (id, username, role) VALUES ($1, $2, 'user')
@@ -175,15 +199,21 @@ func (p *PGRepo) RedeemPlayerBindCode(ctx context.Context, newUserID, code strin
 			return "", "", "", fmt.Errorf("create player: %w", err)
 		}
 		// Re-read by username so a cross-code race converges on the winner's row
-		// (our id was discarded by DO NOTHING) instead of a bare 500.
+		// (our id was discarded by DO NOTHING) instead of a bare 500 — and so a
+		// deleted row squatting on the username is refused rather than reused.
 		var role string
+		var dis, del bool
 		if err := tx.QueryRowContext(ctx,
-			`SELECT id, role::text FROM users WHERE username = $1`, mcUUID).Scan(&userID, &role); err != nil {
+			`SELECT id, role::text, disabled, deleted_at IS NOT NULL FROM users WHERE username = $1`,
+			mcUUID).Scan(&userID, &role, &dis, &del); err != nil {
 			return "", "", "", fmt.Errorf("create player: %w", err)
 		}
 		if role != "user" {
 			return "", "", "", ErrPlayerBindForbidden
 		}
+		if dis || del {
+			return "", "", "", ErrPlayerAccountRetired
+		}
 		if _, err := tx.ExecContext(ctx,
 			`INSERT INTO account_links (user_id, mc_uuid, auth_source) VALUES ($1, $2, $3)
 			 ON CONFLICT (mc_uuid) DO NOTHING`,
@@ -196,6 +226,9 @@ func (p *PGRepo) RedeemPlayerBindCode(ctx context.Context, newUserID, code strin
 		if existingRole != "user" {
 			return "", "", "", ErrPlayerBindForbidden // staff must use op.console
 		}
+		if disabled || deleted {
+			return "", "", "", ErrPlayerAccountRetired
+		}
 	}
 
 	if _, err := tx.ExecContext(ctx,
@@ -209,7 +242,7 @@ func (p *PGRepo) RedeemPlayerBindCode(ctx context.Context, newUserID, code strin
 }
 
 // CompleteOwnerSetup consumes an in-game link code, creates-or-promotes the bound
-// account to the passwordless Owner (role='admin'), enables local auth, and stores
+// account to the passwordless Owner (role='owner'), enables local auth, and stores
 // the one-time first-login token in one transaction. It is the `felis setup`
 // MC-bind path: the operator enters limbo, runs /link, and types the code here.
 // Unlike RedeemPlayerBindCode — which refuses an already-staff account so a game
@@ -240,16 +273,16 @@ func (p *PGRepo) CompleteOwnerSetup(ctx context.Context, newUserID, code string,
 	}
 
 	// Create-or-promote keyed on the verified UUID. An unlinked UUID births a fresh
-	// staff row (role='admin') with a uuid-derived username; an already-linked
+	// staff row (role='owner') with a uuid-derived username; an already-linked
-	// account is promoted to role='admin' in place (idempotent when it is already
+	// account is promoted to role='owner' in place (idempotent when it already is),
-	// staff), keeping its id and username. Setup elevates on purpose, so there is no
+	// keeping its id and username. Setup elevates on purpose, so there is no staff
-	// staff refusal here — that guard belongs to the player path only.
+	// refusal here — that guard belongs to the player path only.
 	userID := newUserID
 	switch err := tx.QueryRowContext(ctx,
 		`SELECT user_id FROM account_links WHERE mc_uuid = $1`, mcUUID).Scan(&userID); {
 	case errors.Is(err, sql.ErrNoRows):
 		if _, err := tx.ExecContext(ctx,
-			`INSERT INTO users (id, username, role) VALUES ($1, $2, 'admin')`,
+			`INSERT INTO users (id, username, role) VALUES ($1, $2, 'owner')`,
 			newUserID, mcUUID); err != nil {
 			return "", "", "", fmt.Errorf("create owner: %w", err)
 		}
@@ -263,7 +296,7 @@ func (p *PGRepo) CompleteOwnerSetup(ctx context.Context, newUserID, code string,
 		return "", "", "", err
 	default:
 		if _, err := tx.ExecContext(ctx,
-			`UPDATE users SET role = 'admin' WHERE id = $1`, userID); err != nil {
+			`UPDATE users SET role = 'owner' WHERE id = $1`, userID); err != nil {
 			return "", "", "", fmt.Errorf("promote owner: %w", err)
 		}
 	}
@@ -292,22 +325,10 @@ func (p *PGRepo) CompleteOwnerSetup(ctx context.Context, newUserID, code string,
 
 // QuotaAvailable treats a missing quota row or a NULL max_servers as unlimited;
 // otherwise it compares the live owned-server count against the cap (spec §9.3).
-//
+// It is the single-dimension convenience read; handlers use the four-dimension
-// KNOWN-LIMITATION (audit #4, quota TOCTOU): this check and ClaimServer are two
+// QuotaCheck. The former audit-#4 TOCTOU (check and claim in separate statements)
-// separate statements, not one transaction, so the count read here is not serialized
+// is closed inside ClaimServer, which re-runs the four-dimension gate under a
-// against a concurrent claim's UPDATE. Two claims by the same user for two DIFFERENT
+// per-user advisory lock in the SAME transaction as the ownership write.
-// ownerless servers can both read count < max_servers (under READ COMMITTED neither
-// sees the other's uncommitted UPDATE) and both succeed, leaving the user one server
-// over quota. Severity is low: it over-provisions the quota by a small margin under a
-// deliberate concurrent burst — it is NOT an authorization, ownership, or isolation
-// break (each server is still claimed atomically via UPDATE ... WHERE owner_id IS
-// NULL, so two users never share one server). Closing it needs Postgres transaction
-// semantics: wrap the count and a conditional UPDATE (gated on count < max_servers) in
-// one tx under pg_advisory_xact_lock(hashtext(user_id)) — or SERIALIZABLE with a retry
-// loop — folding the gate out of the two handlers (handleClaim and the internal UUID
-// claim) into a single repo method. That is INTEGRATION-dependent: it is verifiable
-// only against a real Postgres, not the hermetic fakeRepo suite, so it is documented
-// here rather than patched blind.
 func (p *PGRepo) QuotaAvailable(ctx context.Context, userID string) (bool, error) {
 	var maxServers sql.NullInt64
 	switch err := p.db.QueryRowContext(ctx,
@@ -333,8 +354,11 @@ func (p *PGRepo) QuotaAvailable(ctx context.Context, userID string) (bool, error
 // cached resources should be excluded ("" for a fresh claim where the row
 // doesn't exist yet). Four dimensions are checked: server count, CPU millicores,
 // memory MB, and storage MB. A NULL or missing quota row/column means unlimited
-// for that dimension. Like QuotaAvailable, the count check and the write are not
+// for that dimension. As a standalone read it is advisory — it backs the
-// serialized — see the QuotaAvailable TOCTOU docstring.
+// handler's fast-path 403 — while the AUTHORITATIVE gate for claims is the one
+// ClaimServer re-runs atomically; the resize path (server PATCH) keeps this
+// advisory shape because its write goes through the Kubernetes API, not this
+// transaction.
 func (p *PGRepo) QuotaCheck(ctx context.Context, userID string, excludeName string, incoming ResourceSpec) (bool, error) {
 	var maxServers, maxCPU, maxMem, maxStor sql.NullInt64
 	switch err := p.db.QueryRowContext(ctx,
@@ -347,8 +371,7 @@ func (p *PGRepo) QuotaCheck(ctx context.Context, userID string, excludeName stri
 		return false, err
 	}
 
-	var count int64
+	var count, cpuSum, memSum, storSum int64
-	var cpuSum, memSum, storSum sql.NullInt64
 	switch err := p.db.QueryRowContext(ctx,
 		`SELECT COUNT(*), COALESCE(SUM(cached_cpu_milli), 0), COALESCE(SUM(cached_memory_mb), 0), COALESCE(SUM(cached_storage_mb), 0)
 		 FROM servers WHERE owner_id = $1 AND deleted_at IS NULL AND name != $2`,
@@ -357,19 +380,7 @@ func (p *PGRepo) QuotaCheck(ctx context.Context, userID string, excludeName stri
 		return false, err
 	}
 
-	if maxServers.Valid && count >= maxServers.Int64 {
+	return quotaAllows(maxServers, maxCPU, maxMem, maxStor, count, cpuSum, memSum, storSum, incoming), nil
-		return false, nil
-	}
-	if maxCPU.Valid && cpuSum.Int64+int64(incoming.CPUMilli) > maxCPU.Int64 {
-		return false, nil
-	}
-	if maxMem.Valid && memSum.Int64+int64(incoming.MemoryMB) > maxMem.Int64 {
-		return false, nil
-	}
-	if maxStor.Valid && storSum.Int64+int64(incoming.StorageMB) > maxStor.Int64*1024 {
-		return false, nil
-	}
-	return true, nil
 }
 
 // UpdateServerResources updates the resource cache for a server after a spec
@@ -398,17 +409,67 @@ func (p *PGRepo) ServerResources(ctx context.Context, name string) (ResourceSpec
 
 // ClaimServer performs the atomic ownership transfer (spec §9.3). A missing
 // server is ErrNotFound; an existing-but-owned server yields claimed=false so the
-// handler can answer 409.
+// handler can answer 409; a claim that would push the user over any of the four
+// quota caps yields ErrQuotaExceeded (the handler's pre-check is a fast path,
+// this gate is the authoritative one). The whole decision — quota read,
+// per-owner aggregate, and the ownership UPDATE — runs in ONE transaction under
+// pg_advisory_xact_lock(hashtext(user_id)), so two concurrent claims by the same
+// user for two DIFFERENT ownerless servers serialize instead of both passing the
+// gate (audit #4); the row is additionally taken FOR UPDATE so concurrent claims
+// of the SAME server still resolve to exactly one winner.
 func (p *PGRepo) ClaimServer(ctx context.Context, name, userID string) (bool, error) {
-	var exists bool
+	tx, err := p.db.BeginTx(ctx, nil)
-	if err := p.db.QueryRowContext(ctx,
+	if err != nil {
-		`SELECT EXISTS(SELECT 1 FROM servers WHERE name = $1 AND deleted_at IS NULL)`, name).Scan(&exists); err != nil {
 		return false, err
 	}
-	if !exists {
+	defer tx.Rollback() //nolint:errcheck // no-op after commit
-		return false, ErrNotFound
+
+	// Serialize this user's claim lane: the aggregate read below and the
+	// ownership write must observe one consistent quota state. A hashtext
+	// collision across users merely serializes unrelated claims — never waives a
+	// cap.
+	if _, err := tx.ExecContext(ctx, `SELECT pg_advisory_xact_lock(hashtext($1))`, userID); err != nil {
+		return false, err
 	}
-	res, err := p.db.ExecContext(ctx,
+
+	var owned sql.NullString
+	var cpu, mem, stor int
+	switch err := tx.QueryRowContext(ctx,
+		`SELECT owner_id, cached_cpu_milli, cached_memory_mb, cached_storage_mb
+		 FROM servers WHERE name = $1 AND deleted_at IS NULL FOR UPDATE`,
+		name).Scan(&owned, &cpu, &mem, &stor); {
+	case errors.Is(err, sql.ErrNoRows):
+		return false, ErrNotFound
+	case err != nil:
+		return false, err
+	}
+	if owned.Valid {
+		return false, nil // already claimed → 409 at the handler
+	}
+
+	// The four-dimension gate, re-run inside the transaction. A missing quota
+	// row leaves every NullInt64 invalid → quotaAllows treats each dimension as
+	// unlimited, matching QuotaCheck.
+	var maxServers, maxCPU, maxMem, maxStor sql.NullInt64
+	if err := tx.QueryRowContext(ctx,
+		`SELECT max_servers, max_cpu_milli, max_memory_mb, max_storage_gb
+		 FROM quotas WHERE user_id = $1`, userID).Scan(
+		&maxServers, &maxCPU, &maxMem, &maxStor); err != nil && !errors.Is(err, sql.ErrNoRows) {
+		return false, err
+	}
+	var count, cpuSum, memSum, storSum int64
+	if err := tx.QueryRowContext(ctx,
+		`SELECT COUNT(*), COALESCE(SUM(cached_cpu_milli), 0), COALESCE(SUM(cached_memory_mb), 0), COALESCE(SUM(cached_storage_mb), 0)
+		 FROM servers WHERE owner_id = $1 AND deleted_at IS NULL AND name != $2`,
+		userID, name).Scan(&count, &cpuSum, &memSum, &storSum); err != nil {
+		return false, err
+	}
+	if !quotaAllows(maxServers, maxCPU, maxMem, maxStor, count, cpuSum, memSum, storSum,
+		ResourceSpec{CPUMilli: cpu, MemoryMB: mem, StorageMB: stor}) {
+		return false, ErrQuotaExceeded
+	}
+
+	res, err := tx.ExecContext(ctx,
 		`UPDATE servers SET owner_id = $2, claimed_at = now() WHERE name = $1 AND owner_id IS NULL AND deleted_at IS NULL`,
 		name, userID)
 	if err != nil {
@@ -418,7 +479,35 @@ func (p *PGRepo) ClaimServer(ctx context.Context, name, userID string) (bool, er
 	if err != nil {
 		return false, err
 	}
-	return n == 1, nil
+	if n != 1 {
+		return false, nil
+	}
+	if err := tx.Commit(); err != nil {
+		return false, err
+	}
+	return true, nil
+}
+
+// quotaAllows applies the four spec §9.3 caps to one per-owner aggregate plus
+// the incoming spec. Shared by QuotaCheck (the advisory pre-check) and
+// ClaimServer (the atomic gate) so the two can never drift. An invalid (NULL or
+// missing) cap means unlimited for that dimension; storage is compared in MB
+// against max_storage_gb × 1024.
+func quotaAllows(maxServers, maxCPU, maxMem, maxStor sql.NullInt64,
+	count, cpuSum, memSum, storSum int64, incoming ResourceSpec) bool {
+	if maxServers.Valid && count >= maxServers.Int64 {
+		return false
+	}
+	if maxCPU.Valid && cpuSum+int64(incoming.CPUMilli) > maxCPU.Int64 {
+		return false
+	}
+	if maxMem.Valid && memSum+int64(incoming.MemoryMB) > maxMem.Int64 {
+		return false
+	}
+	if maxStor.Valid && storSum+int64(incoming.StorageMB) > maxStor.Int64*1024 {
+		return false
+	}
+	return true
 }
 
 func (p *PGRepo) UserInAllowlist(ctx context.Context, name, userID string) (bool, error) {
@@ -443,11 +532,16 @@ func (p *PGRepo) UUIDInAllowlist(ctx context.Context, name, mcUUID string) (bool
 }
 
 // UserByMCUUID resolves a verified in-game UUID to its linked user_id (spec §10
-// account_links), or ErrNotFound when the UUID is not linked to any account.
+// account_links), or ErrNotFound when the UUID is not linked to any account. A
+// link whose account is dead reads the same as no link at all (audit #33), so the
+// in-game doors never act as a retired identity.
 func (p *PGRepo) UserByMCUUID(ctx context.Context, mcUUID string) (string, error) {
 	var userID string
 	switch err := p.db.QueryRowContext(ctx,
-		`SELECT user_id FROM account_links WHERE mc_uuid = $1`, mcUUID).Scan(&userID); {
+		`SELECT al.user_id FROM account_links al
+		   JOIN users u ON u.id = al.user_id
+		  WHERE al.mc_uuid = $1 AND u.disabled = false AND u.deleted_at IS NULL`,
+		mcUUID).Scan(&userID); {
 	case errors.Is(err, sql.ErrNoRows):
 		return "", ErrNotFound
 	case err != nil:
@@ -695,7 +789,9 @@ func (p *PGRepo) CreateEmailOTP(ctx context.Context, id, userID, email, codeHash
 // never silently accepted, and a hash mismatch costs an attempt (UPDATE attempts+1)
 // without consuming the code — a typo must not burn a still-valid code. On a match
 // the code is consumed and the user row is flipped verified, returning the proven
-// address. ErrOTPInvalid / ErrOTPLocked are the only domain errors.
+// address — unless a DIFFERENT account already proved the same address, which is
+// ErrEmailTaken with the code left unconsumed (the address, not the guess, is the
+// problem). ErrOTPInvalid / ErrOTPLocked / ErrEmailTaken are the only domain errors.
 func (p *PGRepo) VerifyEmailOTP(ctx context.Context, userID, purpose, codeHash string, now time.Time) (string, error) {
 	tx, err := p.db.BeginTx(ctx, nil)
 	if err != nil {
@@ -738,12 +834,35 @@ func (p *PGRepo) VerifyEmailOTP(ctx context.Context, userID, purpose, codeHash s
 		return "", ErrOTPInvalid
 	}
 
+	// A DIFFERENT account may not also prove this address: the pre-session login
+	// door resolves accounts BY verified email (UserByEmail), so a second verified
+	// holder would make the identity ambiguous. The code is NOT consumed and no
+	// attempt is charged — the address, not the guess, is the problem. The check is
+	// the application-level counterpart of the users_verified_email_unique index
+	// (migration 0020), which catches a cross-user race that passes this SELECT.
+	var taken bool
+	if err := tx.QueryRowContext(ctx,
+		`SELECT EXISTS (SELECT 1 FROM users
+		 WHERE lower(email) = lower($1) AND email_verified = true AND id <> $2)`,
+		email, userID).Scan(&taken); err != nil {
+		return "", fmt.Errorf("check verified-email uniqueness: %w", err)
+	}
+	if taken {
+		return "", ErrEmailTaken
+	}
+
 	if _, err := tx.ExecContext(ctx,
 		`UPDATE email_otps SET consumed_at = $2 WHERE id = $1`, id, now); err != nil {
 		return "", fmt.Errorf("consume otp: %w", err)
 	}
 	if _, err := tx.ExecContext(ctx,
 		`UPDATE users SET email = $2, email_verified = true WHERE id = $1`, userID, email); err != nil {
+		// Lost the race the guarded SELECT above cannot serialise: the index rejects
+		// the second write, and it reads as the same answer the sequential path gives.
+		// The rollback undoes the OTP consumption with it, so the code stays live.
+		if isUniqueViolation(err) {
+			return "", ErrEmailTaken
+		}
 		return "", fmt.Errorf("mark email verified: %w", err)
 	}
 	if err := tx.Commit(); err != nil {
@@ -833,17 +952,17 @@ func (p *PGRepo) IsUsernameBlacklisted(ctx context.Context, mcUUID string) (bool
 // who authenticates through the third-party Yggdrasil — the admin-on-Yggdrasil reclaim
 // exception (spec §B3). The EXISTS joins account_links to users on exactly three
 // conjuncts: the UUID is linked, that link authenticated via 'thirdparty', and the
-// linked user is an admin. It intentionally does not test HOW the account signs in:
+// linked user is staff (admin OR owner — the Owner is the one identity that must never
-// an Operator may authenticate via SSO (Cloudflare Access, §14) or any local
+// be displaced). It intentionally does not test HOW the account signs in: staff may
-// passwordless door and must be protected just the same — the sign-in method is
+// authenticate via SSO (Cloudflare Access, §14) or any local passwordless door and
-// orthogonal to "is staff" and "logs in via the Login Server". Keyed by UUID, the
+// must be protected just the same — the sign-in method is orthogonal to "is staff"
-// only identity velocity holds.
+// and "logs in via the Login Server". Keyed by UUID, the only identity velocity holds.
 func (p *PGRepo) IsProtectedAdminLink(ctx context.Context, mcUUID string) (bool, error) {
 	var ok bool
 	err := p.db.QueryRowContext(ctx,
 		`SELECT EXISTS(
 			SELECT 1 FROM account_links al JOIN users u ON u.id = al.user_id
-			WHERE al.mc_uuid = $1 AND al.auth_source = 'thirdparty' AND u.role = 'admin')`,
+			WHERE al.mc_uuid = $1 AND al.auth_source = 'thirdparty' AND u.role IN ('admin', 'owner'))`,
 		mcUUID).Scan(&ok)
 	return ok, err
 }
@@ -867,13 +986,13 @@ func (p *PGRepo) UserByUsername(ctx context.Context, username string) (*StaffUse
 	return &u, nil
 }
 
-// AdminExists reports whether any admin account already exists. It is the
+// AdminExists reports whether any staff account (admin or owner) already exists. It is the
 // break-glass console's bootstrap-vs-recovery switch: false means the typed
 // credential mints the first Owner (no prior identity to verify against), true
-// means the operator must identify against an existing admin for accountability.
+// means the operator must identify against an existing staff account for accountability.
 // It is not on the Repo interface because only the break-glass CLI consults it.
 func (p *PGRepo) AdminExists(ctx context.Context) (bool, error) {
-	const q = `SELECT 1 FROM users WHERE role = 'admin' LIMIT 1`
+	const q = `SELECT 1 FROM users WHERE role IN ('admin', 'owner') LIMIT 1`
 	var one int
 	switch err := p.db.QueryRowContext(ctx, q).Scan(&one); {
 	case errors.Is(err, sql.ErrNoRows):
@@ -901,29 +1020,55 @@ func (p *PGRepo) UserByID(ctx context.Context, id string) (*StaffUser, error) {
 }
 
 // UpsertOwner creates or resets the Owner account direct-to-Postgres (the
-// break-glass first-run / recovery path). role is forced to 'admin' — the
+// break-glass first-run / recovery path). role is forced to 'owner' — the
-// platform-level identity. On a username conflict the email is overwritten
+// platform-level identity above admin (migration 0011); every owner-tier route
-// while the existing id is preserved, so live sessions referencing it survive
+// and the panel's owner surfaces gate on exactly this role, so writing a plain
-// a reset. The account is passwordless by design. The empty email is stored
+// 'admin' here would silently strand them. On a username conflict the email is
-// as NULL (users.email is nullable).
+// overwritten while the existing id is preserved, so live sessions referencing
+// it survive a reset — and the role is re-asserted, which is also the documented
+// promotion path for a pre-0011 install whose Owner row is still 'admin'. The
+// account is passwordless by design. The empty email is stored as NULL
+// (users.email is nullable).
 func (p *PGRepo) UpsertOwner(ctx context.Context, id, username, email string) error {
 	_, err := p.db.ExecContext(ctx,
-		`INSERT INTO users (id, username, email, role) VALUES ($1, $2, NULLIF($3, ''), 'admin')
+		`INSERT INTO users (id, username, email, role) VALUES ($1, $2, NULLIF($3, ''), 'owner')
-		 ON CONFLICT (username) DO UPDATE SET email = EXCLUDED.email`,
+		 ON CONFLICT (username) DO UPDATE SET email = EXCLUDED.email, role = 'owner'`,
 		id, username, email)
 	return err
 }
 
+// OwnerUsername names the single active Owner seat, or "" when no owner exists.
+// It backs the console's single-seat guard: once a seat is occupied only that
+// username may be re-targeted (see cmd/felis provisionOwner), because a fresh
+// name would take the upsert's insert arm and mint a SECOND owner row that no
+// supported path can remove (the panel protects every owner row).
+func (p *PGRepo) OwnerUsername(ctx context.Context) (string, error) {
+	var name string
+	switch err := p.db.QueryRowContext(ctx,
+		`SELECT username FROM users WHERE role = 'owner' AND deleted_at IS NULL ORDER BY created_at LIMIT 1`).Scan(&name); {
+	case errors.Is(err, sql.ErrNoRows):
+		return "", nil
+	case err != nil:
+		return "", err
+	default:
+		return name, nil
+	}
+}
+
 // InsertOperator mints a NEW Operator (additional staff admin) account
 // direct-to-Postgres. role is forced to 'admin'. UNLIKE UpsertOwner this is
-// insert-only: a username conflict is left untouched and surfaces as a driver
+// insert-only: a username conflict leaves the existing row untouched and
-// error, so adding an Operator can never silently reset the Owner's or another
+// surfaces as ErrConflict — the console routes a rename off that sentinel — so
+// adding an Operator can never silently reset the Owner's or another
 // Operator's row. The account is passwordless by design. The empty email is
 // stored as NULL.
 func (p *PGRepo) InsertOperator(ctx context.Context, id, username, email string) error {
 	_, err := p.db.ExecContext(ctx,
 		`INSERT INTO users (id, username, email, role) VALUES ($1, $2, NULLIF($3, ''), 'admin')`,
 		id, username, email)
+	if err != nil && isUniqueViolation(err) {
+		return ErrConflict
+	}
 	return err
 }
 
@@ -939,9 +1084,14 @@ func (p *PGRepo) CreateSession(ctx context.Context, tokenHash, userID string, ex
 // SessionUser resolves a live (unrevoked, unexpired at now) session hash to its
 // user, or ErrNotFound.
 func (p *PGRepo) SessionUser(ctx context.Context, tokenHash string, now time.Time) (*SessionedUser, error) {
+	// The disabled/deleted filter is the belt to the doors' braces: even a session
+	// minted for an account that was alive a moment ago stops authenticating the
+	// instant the account is disabled or soft-deleted, so every authenticated route
+	// is fail-closed regardless of which door minted the cookie (audit #33).
 	const q = `SELECT u.id, COALESCE(u.email, ''), u.role::text, COALESCE(u.email_verified, false)
 		FROM sessions s JOIN users u ON u.id = s.user_id
-		WHERE s.token_hash = $1 AND s.revoked_at IS NULL AND s.expires_at > $2`
+		WHERE s.token_hash = $1 AND s.revoked_at IS NULL AND s.expires_at > $2
+		  AND u.disabled = false AND u.deleted_at IS NULL`
 	var u SessionedUser
 	switch err := p.db.QueryRowContext(ctx, q, tokenHash, now).Scan(
 		&u.ID, &u.Email, &u.Role, &u.EmailVerified); {
@@ -1397,7 +1547,15 @@ func (p *PGRepo) UpdateUser(ctx context.Context, userID string, patch UpdateUser
 	}
 	if patch.Email != nil {
 		argn++
-		sets = append(sets, fmt.Sprintf("email = NULLIF($%d, '')", argn))
+		// Changing the address voids any proof of it: only VerifyEmailOTP may assert
+		// a verified address (mirrors SetUserEmail's rationale — a fresh, unproven
+		// value must not keep a stale verified flag that would let the pre-session
+		// email login resolve the account). A no-op edit that passes the same value
+		// keeps the flag; the second expression reads the OLD row, so comparing
+		// there is exact.
+		sets = append(sets,
+			fmt.Sprintf("email = NULLIF($%d, '')", argn),
+			fmt.Sprintf("email_verified = (email_verified AND email IS NOT DISTINCT FROM NULLIF($%d, ''))", argn))
 		args = append(args, *patch.Email)
 	}
 	if patch.Role != nil {
@@ -1463,8 +1621,11 @@ func (p *PGRepo) userView(ctx context.Context, userID string) (*UserView, error)
 }
 
 // DeleteUser soft-deletes a user in one transaction: sets deleted_at, revokes
-// every live session, and releases every owned server. The row is preserved so
+// every live session, releases every owned server, and severs the account's
-// audit_logs.actor references survive.
+// identity assets (passkey credentials, Minecraft links) so a closed account
+// cannot keep a login credential or pin an in-game identity via
+// UNIQUE(mc_uuid)(audit #33). The user row is preserved so audit_logs.actor
+// references survive.
 func (p *PGRepo) DeleteUser(ctx context.Context, userID, _ string) error {
 	tx, err := p.db.BeginTx(ctx, nil)
 	if err != nil {
@@ -1497,6 +1658,18 @@ func (p *PGRepo) DeleteUser(ctx context.Context, userID, _ string) error {
 		return err
 	}
 
+	// Sever the login credentials and in-game bindings: a passkey is a standing
+	// login foothold and occupies credential_id UNIQUE, and an account_links row
+	// would keep the Minecraft UUID claimed forever, blocking any future binding.
+	if _, err := tx.ExecContext(ctx,
+		`DELETE FROM webauthn_credentials WHERE user_id = $1`, userID); err != nil {
+		return err
+	}
+	if _, err := tx.ExecContext(ctx,
+		`DELETE FROM account_links WHERE user_id = $1`, userID); err != nil {
+		return err
+	}
+
 	// Soft-delete the user row.
 	if _, err := tx.ExecContext(ctx,
 		`UPDATE users SET disabled = true, deleted_at = now() WHERE id = $1`,
@@ -1538,8 +1711,13 @@ func (p *PGRepo) SetUserDisabled(ctx context.Context, userID string, disabled bo
 // ---- quota admin ----
 
 // GetQuotas returns the quotas row for a user, or a zero-value view when no
-// row exists (meaning unlimited).
+// row exists (meaning unlimited). An unknown or soft-deleted user id is
+// ErrNotFound, never a zero-value "unlimited" answer — the admin sub-resource
+// routes all 404 on a user that has no live row.
 func (p *PGRepo) GetQuotas(ctx context.Context, userID string) (*QuotaView, error) {
+	if err := p.requireLiveUser(ctx, userID); err != nil {
+		return nil, err
+	}
 	const q = `SELECT user_id, max_servers, max_cpu_milli, max_memory_mb, max_storage_gb
 		FROM quotas WHERE user_id = $1`
 	v := QuotaView{UserID: userID}
@@ -1553,9 +1731,29 @@ func (p *PGRepo) GetQuotas(ctx context.Context, userID string) (*QuotaView, erro
 	return &v, nil
 }
 
+// requireLiveUser gates the user-scoped admin sub-resources (quotas, account
+// links) on a live users row. Without it a write would hit the user_id foreign
+// key and surface as an opaque 500, and the read would answer as if a
+// never-existed id did; every admin route answers ErrNotFound → 404 instead.
+func (p *PGRepo) requireLiveUser(ctx context.Context, userID string) error {
+	var ok bool
+	if err := p.db.QueryRowContext(ctx,
+		`SELECT EXISTS(SELECT 1 FROM users WHERE id = $1 AND deleted_at IS NULL)`,
+		userID).Scan(&ok); err != nil {
+		return err
+	}
+	if !ok {
+		return ErrNotFound
+	}
+	return nil
+}
+
 // SetQuotas upserts a quotas row. Nil fields are left unchanged; a non-nil
 // zero-value field clears the cap.
 func (p *PGRepo) SetQuotas(ctx context.Context, userID string, qi QuotaInput, setBy string) (*QuotaView, error) {
+	if err := p.requireLiveUser(ctx, userID); err != nil {
+		return nil, err
+	}
 	type col struct {
 		name  string
 		value *int
@@ -1660,6 +1858,9 @@ func (p *PGRepo) UnlinkAccount(ctx context.Context, userID, mcUUID string) error
 // NOTHING on the UNIQUE(mc_uuid) constraint, plus an idempotency check via
 // EXISTS).
 func (p *PGRepo) LinkAccount(ctx context.Context, userID, mcUUID, authSource string) error {
+	if err := p.requireLiveUser(ctx, userID); err != nil {
+		return err
+	}
 	// Check idempotency first: already linked to this user → success.
 	var exists bool
 	if err := p.db.QueryRowContext(ctx,
@@ -1880,8 +2081,13 @@ func (p *PGRepo) RedeemMigration(ctx context.Context, targetUserID, codeHash str
 func (p *PGRepo) UserByEmail(ctx context.Context, email string) (*StaffUser, error) {
 	// lower() on both sides honors the interface's case-insensitivity contract
 	// and matches the users_verified_email_unique index (lower(email)).
+	// Disabled and soft-deleted accounts are invisible here on purpose: every caller
+	// is a pre-session LOGIN door (console email/passkey, op-login, /auth/options),
+	// and a dead account must not be able to mint a session again — deletion and the
+	// disable lockout would otherwise be bypassable by simply logging in (audit #33).
 	const q = `SELECT id, username, COALESCE(email, ''), role::text, email_verified
-		FROM users WHERE lower(email) = lower($1) AND email_verified = true`
+		FROM users WHERE lower(email) = lower($1) AND email_verified = true
+		  AND disabled = false AND deleted_at IS NULL`
 	var u StaffUser
 	switch err := p.db.QueryRowContext(ctx, q, email).Scan(
 		&u.ID, &u.Username, &u.Email, &u.Role, &u.EmailVerified); {

internal/api/repo.go

@@ -77,10 +77,10 @@ type BackupRecord struct {
 }
 
 // StaffUser is the login-side projection of a users row (spec §B passwordless
-// auth). Owner/Operator are role=admin rows, minted by `felis setup` (MC link)
+// auth). The Owner is the role=owner row, minted by `felis setup` (MC link) and
-// and recovered by `felis breakGlass` (email OTP); players are role=user rows.
+// recovered by `felis breakGlass` (email OTP); Operators are role=admin;
-// There is no password column — staff authenticate via email-OTP / passkey +
+// players are role=user. There is no password column — staff authenticate via
-// in-game approve, never a password.
+// email-OTP / passkey + in-game approve, never a password.
 type StaffUser struct {
 	ID       string
 	Username string
@@ -207,9 +207,10 @@ type Repo interface {
 	//     return newUserID;
 	//   - the uuid is already linked to a role='user' player → return THAT user
 	//     (idempotent "log in via the game"), consuming the code;
-	//   - the uuid is linked to a role='admin' STAFF account → ErrPlayerBindForbidden
+	//   - the uuid is linked to a STAFF account (role != 'user', i.e. admin or owner)
-	//     WITHOUT consuming the code (operators use op.console behind Zero Trust; the
+	//     → ErrPlayerBindForbidden WITHOUT consuming the code (staff use op.console
-	//     public bootstrap never mints a session for an admin identity).
+	//     behind Zero Trust; the public bootstrap never mints a session for a staff
+	//     identity).
 	//
 	// Safe as an unauthenticated entrypoint because a Bind Code is minted internal-face
 	// only (CreateLinkCode), against an online-mode-verified UUID, short-TTL and
@@ -230,7 +231,11 @@ type Repo interface {
 	QuotaCheck(ctx context.Context, userID string, excludeName string, incoming ResourceSpec) (bool, error)
 	// ClaimServer atomically sets owner_id where it is currently NULL and returns
 	// whether a row changed. false means the server was already claimed (spec §9.3:
-	// 0 rows → 409).
+	// 0 rows → 409). The claim runs in one transaction that re-checks the four quota
+	// dimensions under pg_advisory_xact_lock(hashtext(user_id)), so it is the
+	// authoritative gate: a claim that would exceed a cap → ErrQuotaExceeded (403),
+	// and two concurrent claims by one user cannot both pass (audit #4).
+	// QuotaCheck remains the advisory pre-check for the handler's fast-path 403.
 	ClaimServer(ctx context.Context, name, userID string) (bool, error)
 	// UserInAllowlist reports whether the user's linked UUID is on the server
 	// allowlist (spec §9.4).
@@ -246,6 +251,10 @@ type Repo interface {
 	// (spec §10 account_links), or ErrNotFound when the UUID is not linked. The
 	// internal-face wake uses it to apply the owner bypass for a player known only
 	// by UUID; an unlinked UUID simply falls through to the autostartPolicy gate.
+	// Only a LIVE account resolves (audit #33): a link whose account is disabled or
+	// soft-deleted carries no standing on the in-game doors — claim, menu, wake
+	// authorization, op-login vouch and the QR link-status poll all read a dead
+	// account exactly like an unlinked UUID, never as a retired identity.
 	UserByMCUUID(ctx context.Context, mcUUID string) (userID string, err error)
 	// RecordJoin updates last_active_at, clears reaper warnings, and auto-appends
 	// the UUID to the allowlist (spec §7 join-event, §9.4).
@@ -309,7 +318,10 @@ type Repo interface {
 	// mismatch increments attempts and returns ErrOTPInvalid WITHOUT consuming the
 	// code (so a typo does not burn it). On a match the code is consumed and the
 	// user row is flipped to email=<the proven address>, email_verified=true; the
-	// proven email is returned. now is the API clock so expiry is testable.
+	// proven email is returned — unless a DIFFERENT account has already proven the
+	// same address, which is ErrEmailTaken with the code left unconsumed (the
+	// address, not the code, is the problem). now is the API clock so expiry is
+	// testable.
 	//
 	// This is the ONBOARDING primitive: verifying the code is the moment the address
 	// becomes proven, so the write is load-bearing. The pre-session LOGIN door must
@@ -460,12 +472,13 @@ type Repo interface {
 	// is staff logging in via the Login Server, not a Mojang squatter, so a
 	// Mojang-priority reclaim must never bar them. The predicate is exactly three
 	// conjuncts: the UUID is linked (account_links), that link authenticated via
-	// 'thirdparty' (auth_source), and the linked user is an admin (role='admin').
+	// 'thirdparty' (auth_source), and the linked user is staff — admin or owner
+	// (migration 0011), the Owner being the identity most in need of the exception.
 	// It deliberately does NOT ask HOW the staff account signs in: an Operator may
 	// authenticate via SSO (Cloudflare Access, IdP-agnostic per §14) or any local
 	// passwordless door, and must be protected all the same — the sign-in method is
 	// orthogonal to both "is staff" and "logs in via the Login Server". An unlinked
-	// UUID, a Mojang-sourced link, or a non-admin link all yield false, so the
+	// UUID, a Mojang-sourced link, or a player link all yield false, so the
 	// exception never broadens to ordinary thirdparty players (Mojang priority still
 	// displaces them) nor to Mojang-authenticated identities (who have no Login-Server
 	// name to protect). Keyed by UUID — the only identity velocity knows.
@@ -490,7 +503,7 @@ type Repo interface {
 	// (email_verified true), so a merely-asserted or unverified address never
 	// resolves to a session-mintable identity — an attacker cannot claim someone
 	// else's login by typing their email. Matching is on lower(email) to align with
-	// the users_verified_email_unique partial index (migration 0010), which
+	// the users_verified_email_unique partial index (migration 0020), which
 	// guarantees at most one verified row per normalized address, so the result is
 	// unambiguous. A player (role='user') row resolves too — email-first login is
 	// passwordless and role-agnostic here; the door that consumes this result decides
@@ -498,8 +511,10 @@ type Repo interface {
 	UserByEmail(ctx context.Context, email string) (*StaffUser, error)
 	// UpsertOwner creates or resets the single Owner account direct-to-Postgres
 	// (the `felis setup` / `felis breakGlass` recovery path). role is forced to
-	// 'admin'; on a username conflict the existing row's email is overwritten so
+	// 'owner' (migration 0011 — the tier every user-admin route gates on); on a
-	// a reset is idempotent. The account is passwordless by design.
+	// username conflict the existing row's email is overwritten and the role
+	// re-asserted, so a reset is idempotent and a pre-0011 'admin' Owner row is
+	// promoted. The account is passwordless by design.
 	UpsertOwner(ctx context.Context, id, username, email string) error
 	// CreateSession records a minted session: the sha-256 of the opaque cookie
 	// value, its owner, and its expiry (spec §B sessions). Only the hash is stored,

internal/api/session.go

@@ -7,6 +7,7 @@ import (
 	"encoding/base64"
 	"encoding/hex"
 	"encoding/json"
+	"errors"
 	"fmt"
 	"net"
 	"net/http"
@@ -145,14 +146,24 @@ func (s SessionAuth) Authenticate(r *http.Request) (*Principal, error) {
 	}
 
 	ctx := r.Context()
-	if !localAuthEnabled(ctx, s.Repo) {
+	enabled, err := localAuthEnabledStatus(ctx, s.Repo)
+	if err != nil {
+		// The session store is unreachable: this is an outage, not a verdict on
+		// the caller's credentials, so the middleware answers 503 rather than a
+		// misleading "please log in".
+		return nil, fmt.Errorf("%w: %v", errAuthBackend, err)
+	}
+	if !enabled {
 		// A cookie was presented but local auth is off: reject, never fall through.
 		return nil, fmt.Errorf("local auth disabled")
 	}
 
 	u, err := s.Repo.SessionUser(ctx, hashCookie(cookie.Value), s.now())
-	if err != nil {
+	switch {
+	case errors.Is(err, ErrNotFound):
 		return nil, fmt.Errorf("invalid session: %w", err)
+	case err != nil:
+		return nil, fmt.Errorf("%w: %v", errAuthBackend, err)
 	}
 	return &Principal{
 		UserID:         u.ID,
@@ -164,6 +175,11 @@ func (s SessionAuth) Authenticate(r *http.Request) (*Principal, error) {
 	}, nil
 }
 
+// errAuthBackend marks an authentication failure caused by the session store
+// being unreachable (e.g. Postgres down) rather than by a missing or invalid
+// credential. Middleware maps it to 503 so an outage is not misreported as 401.
+var errAuthBackend = errors.New("auth backend unavailable")
+
 // localAuthEnabled reports whether the runtime local_auth_enabled toggle is true.
 // A missing setting, a read error, or a non-true value all read as disabled — the
 // gate fails closed so local sessions are honored, and new ones minted, only on an
@@ -171,15 +187,28 @@ func (s SessionAuth) Authenticate(r *http.Request) (*Principal, error) {
 // (minting one) consult it, so the two never disagree about whether local auth is
 // live.
 func localAuthEnabled(ctx context.Context, repo Repo) bool {
+	enabled, _ := localAuthEnabledStatus(ctx, repo)
+	return enabled
+}
+
+// localAuthEnabledStatus is localAuthEnabled with the outage case kept apart: a
+// MISSING setting (ErrNotFound — never enabled) reads as (false, nil), while a
+// store read failure reads as (false, err) so SessionAuth can tell "local auth
+// is off" (401) from "the database is down" (503). An unreadable value still
+// fails closed as disabled — it is a config fault, not an outage.
+func localAuthEnabledStatus(ctx context.Context, repo Repo) (bool, error) {
 	raw, err := repo.GetSetting(ctx, LocalAuthEnabledKey)
-	if err != nil {
+	switch {
-		return false // ErrNotFound (never enabled) or a transient read error → closed
+	case errors.Is(err, ErrNotFound):
+		return false, nil
+	case err != nil:
+		return false, err
 	}
 	var enabled bool
 	if err := json.Unmarshal(raw, &enabled); err != nil {
-		return false
+		return false, nil
 	}
-	return enabled
+	return enabled, nil
 }
 
 // ensure SessionAuth satisfies ExternalAuth at compile time.

internal/api/submissions.go

@@ -6,6 +6,7 @@ import (
 	"io"
 	"net/http"
 
+	"felis.lolicon.best/internal/build"
 	"felis.lolicon.best/internal/submit"
 )
 
@@ -42,6 +43,18 @@ type SubmissionService interface {
 	// Reject is the admin's other verdict: pending_review -> rejected with a
 	// required reason; it starts no build.
 	Reject(ctx context.Context, id, reviewedBy, reason string) (*submit.Submission, error)
+	// Withdraw retracts the caller's OWN pending submission: the row and its
+	// uploaded context are deleted. A reviewed submission is frozen (409) and a
+	// submission the caller does not own reads back as 404, like the upload route.
+	Withdraw(ctx context.Context, id, submittedBy string) (*submit.Submission, error)
+	// Delete retires any submission outright (the admin lifecycle valve): the row
+	// and its uploaded context are removed, any status.
+	Delete(ctx context.Context, id string) (*submit.Submission, error)
+	// OpenContext returns the stored build-context blob for the internal
+	// context-fetch route: the build Pod's initContainer cannot mount the uploads
+	// PVC across namespaces and holds no object-store credentials, so it streams
+	// the blob from the API over the service-token-gated internal face instead.
+	OpenContext(ctx context.Context, id string) (io.ReadCloser, error)
 }
 
 // createSubmissionRequest is the POST /me/submissions body. The user
@@ -53,6 +66,15 @@ type createSubmissionRequest struct {
 	DisplayName string `json:"display_name"`
 }
 
+// The submission lane's two cooldown keys, prefixed into the shared submit
+// limiter's per-user keys. Create and upload are separate levers on purpose:
+// creating a submission and then immediately uploading its context is the lane's
+// normal shape, so one must never consume the other's window.
+const (
+	submissionCreateKey = "create:"
+	submissionUploadKey = "upload:"
+)
+
 // rejectSubmissionRequest is the POST /submissions/{id}/reject body. A reason is
 // required (the submit layer rejects an empty one with 400).
 type rejectSubmissionRequest struct {
@@ -68,6 +90,26 @@ func (a *API) handleCreateSubmission(w http.ResponseWriter, r *http.Request) {
 		return
 	}
 	p := principalFromContext(r.Context())
+	// Reserve the per-user create cooldown BEFORE the store write. Unlike the
+	// wake lever's allowed→record (whose real gate is the running cap and whose
+	// effect is idempotent), a create is a non-idempotent row insertion with no
+	// other bound on its rate, so a burst of truly concurrent creates must yield
+	// exactly one winner per window. The deferred rollback frees the window
+	// whenever the create fails — a 400 typo, a spent quota, a store error — so
+	// only a row that was actually recorded consumes it.
+	lim := a.submitLimiter()
+	reservedAt, ok := lim.reserve(submissionCreateKey+p.UserID, a.SubmitCreateCooldown)
+	if !ok {
+		writeError(w, r, newError(http.StatusTooManyRequests, "submission_cooldown",
+			"a submission was created recently; wait a moment before creating another"))
+		return
+	}
+	committed := false
+	defer func() {
+		if !committed {
+			lim.release(submissionCreateKey+p.UserID, reservedAt)
+		}
+	}()
 	var body createSubmissionRequest
 	if err := decodeJSON(w, r, &body); err != nil {
 		writeError(w, r, err)
@@ -81,6 +123,7 @@ func (a *API) handleCreateSubmission(w http.ResponseWriter, r *http.Request) {
 		writeSubmitError(w, r, err)
 		return
 	}
+	committed = true
 	a.audit(r, p.Email, "submission.create", sub.ID)
 	writeJSON(w, http.StatusCreated, sub)
 }
@@ -101,19 +144,43 @@ func (a *API) handleUploadSubmissionContext(w http.ResponseWriter, r *http.Reque
 		return
 	}
 	p := principalFromContext(r.Context())
+	// Reserve the per-user upload cooldown BEFORE streaming. The body is the
+	// expensive part (up to the 1 GiB blob cap), so without a reservation the
+	// throttle would never bound the resource it exists for: a caller could
+	// repeatedly start long uploads and abort them. Reserving also collapses the
+	// lane's parallel overshoot — a burst of concurrent uploads from one user
+	// yields exactly one admitted stream per replica. The rollback keeps a failed
+	// upload (aborted transfer, wrong format, spent quota) from burning the
+	// window, so a legit retry after a genuine failure is not punished.
+	lim := a.submitLimiter()
+	reservedAt, ok := lim.reserve(submissionUploadKey+p.UserID, a.SubmitUploadCooldown)
+	if !ok {
+		writeError(w, r, newError(http.StatusTooManyRequests, "submission_cooldown",
+			"an upload was accepted recently; wait a moment before uploading again"))
+		return
+	}
+	committed := false
+	defer func() {
+		if !committed {
+			lim.release(submissionUploadKey+p.UserID, reservedAt)
+		}
+	}()
 	id := r.PathValue("id")
 	sub, err := a.Submissions.UploadContext(r.Context(), id, p.UserID, r.Body)
 	if err != nil {
 		writeSubmitError(w, r, err)
 		return
 	}
+	committed = true
 	a.audit(r, p.Email, "submission.upload", sub.ID)
 	writeJSON(w, http.StatusOK, sub)
 }
 
 // handleMySubmissions lists the caller's own submissions (app-tier). It scopes
 // strictly to the principal's id; there is no parameter that could widen the
-// query to another user's uploads.
+// query to another user's uploads. Each row is enriched with its linked build's
+// outcome — this list is the only player-visible outlet for a build result, so
+// a failed build is not invisible to the person who submitted it.
 func (a *API) handleMySubmissions(w http.ResponseWriter, r *http.Request) {
 	if a.Submissions == nil {
 		writeError(w, r, errSubmissionsUnavailable)
@@ -125,12 +192,18 @@ func (a *API) handleMySubmissions(w http.ResponseWriter, r *http.Request) {
 		writeSubmitError(w, r, err)
 		return
 	}
-	writeJSON(w, http.StatusOK, map[string]any{"submissions": subs})
+	views, err := a.submissionViews(r.Context(), subs)
+	if err != nil {
+		writeSubmitError(w, r, err)
+		return
+	}
+	writeJSON(w, http.StatusOK, map[string]any{"submissions": views})
 }
 
 // handleListSubmissions is the admin review queue: every submission across all
 // users, newest first (admin-tier — it reads other users' uploads, so it gates
-// on the admin Zero-Trust path via adminOnly).
+// on the admin Zero-Trust path via adminOnly). Rows carry the same build
+// outcome enrichment as /me/submissions.
 func (a *API) handleListSubmissions(w http.ResponseWriter, r *http.Request) {
 	if a.Submissions == nil {
 		writeError(w, r, errSubmissionsUnavailable)
@@ -141,7 +214,49 @@ func (a *API) handleListSubmissions(w http.ResponseWriter, r *http.Request) {
 		writeSubmitError(w, r, err)
 		return
 	}
-	writeJSON(w, http.StatusOK, map[string]any{"submissions": subs})
+	views, err := a.submissionViews(r.Context(), subs)
+	if err != nil {
+		writeSubmitError(w, r, err)
+		return
+	}
+	writeJSON(w, http.StatusOK, map[string]any{"submissions": views})
+}
+
+// submissionView is one submission row enriched with its linked build's
+// outcome. The row itself is embedded unchanged, so the wire shape only gains
+// the two optional fields; they appear solely once a build has been linked
+// (BuildID set) and its record is still readable.
+type submissionView struct {
+	submit.Submission
+	BuildStatus string `json:"build_status,omitempty"`
+	BuildError  string `json:"build_error,omitempty"`
+}
+
+// submissionViews enriches each submission with its linked build's status via a
+// read-only Builder.Get — deliberately never Sync, because the 15s reconcile
+// loop owns state advance and rendering a list must not touch the cluster. A
+// submission with no linked build (never approved, or approved before the
+// hand-off could record the id), no Builder wired, or a build row that is gone
+// (ErrNotFound) renders without the extra fields; any other store failure is
+// returned so the handler reports it rather than silently dropping the outcome.
+func (a *API) submissionViews(ctx context.Context, subs []submit.Submission) ([]submissionView, error) {
+	views := make([]submissionView, len(subs))
+	for i, s := range subs {
+		views[i] = submissionView{Submission: s}
+		if a.Builder == nil || s.BuildID == "" {
+			continue
+		}
+		bld, err := a.Builder.Get(ctx, s.BuildID)
+		if errors.Is(err, build.ErrNotFound) {
+			continue
+		}
+		if err != nil {
+			return nil, err
+		}
+		views[i].BuildStatus = string(bld.Status)
+		views[i].BuildError = bld.Error
+	}
+	return views, nil
 }
 
 // handleApproveSubmission is the admin approve gate (admin-tier). The reviewer is
@@ -186,6 +301,48 @@ func (a *API) handleRejectSubmission(w http.ResponseWriter, r *http.Request) {
 	writeJSON(w, http.StatusOK, sub)
 }
 
+// handleWithdrawSubmission retracts the caller's own pending submission
+// (app-tier): the row and its uploaded context are deleted, freeing the pending
+// slot and the storage budget for a fresh submission. The submitter is the
+// principal, never the body; a submission the caller does not own is reported as
+// 404, so this endpoint cannot probe or clear another user's uploads, and a
+// reviewed submission is 409 (its build may already be consuming the context).
+func (a *API) handleWithdrawSubmission(w http.ResponseWriter, r *http.Request) {
+	if a.Submissions == nil {
+		writeError(w, r, errSubmissionsUnavailable)
+		return
+	}
+	p := principalFromContext(r.Context())
+	sub, err := a.Submissions.Withdraw(r.Context(), r.PathValue("id"), p.UserID)
+	if err != nil {
+		writeSubmitError(w, r, err)
+		return
+	}
+	a.audit(r, p.Email, "submission.withdraw", sub.ID)
+	writeJSON(w, http.StatusOK, sub)
+}
+
+// handleDeleteSubmission retires any submission outright (admin-tier): the row
+// and its uploaded context are removed, any status. This is the lane's lifecycle
+// valve — the only path that reclaims a rejected or consumed upload from the
+// uploads PVC. The reviewer identity goes to the audit event, not the (now
+// nonexistent) row. Deleting an approved submission whose build is still running
+// fails that build's context fetch; the admin has explicitly chosen to retire it.
+func (a *API) handleDeleteSubmission(w http.ResponseWriter, r *http.Request) {
+	if a.Submissions == nil {
+		writeError(w, r, errSubmissionsUnavailable)
+		return
+	}
+	p := principalFromContext(r.Context())
+	sub, err := a.Submissions.Delete(r.Context(), r.PathValue("id"))
+	if err != nil {
+		writeSubmitError(w, r, err)
+		return
+	}
+	a.audit(r, p.Email, "submission.delete", sub.ID)
+	writeJSON(w, http.StatusOK, sub)
+}
+
 // errSubmissionsUnavailable is returned when the approval lane is not configured
 // on this api instance (a nil Submissions service), so the admin/app boundary is
 // still exercised before the subsystem is wired in.
@@ -194,9 +351,10 @@ var errSubmissionsUnavailable = newError(http.StatusServiceUnavailable, "submiss
 
 // writeSubmitError maps submit-package errors onto HTTP status codes. Only the
 // business sentinels are client-facing: a validation failure is 400, a missing
-// submission is 404, an already-reviewed submission is 409, and an unconfigured
+// submission is 404, an already-reviewed submission is 409, a spent per-user
-// upload transport is 503 (the store this deployment set has no implemented
+// allowance is 403 (the same status the server-resource quota answers with), and
-// transport — an honest "not available here", not a client error). Everything
+// an unconfigured upload transport is 503 (the store this deployment set has no
+// implemented transport — an honest "not available here", not a client error). Everything
 // else — including a build.ErrInvalid raised by the pre-CAS build.Validate (a
 // platform registry/context MISCONFIGURATION, never client input, since every
 // build input is platform-derived) and a post-CAS Submit hand-off failure — is a
@@ -211,6 +369,11 @@ func writeSubmitError(w http.ResponseWriter, r *http.Request, err error) {
 	case errors.Is(err, submit.ErrAlreadyReviewed):
 		writeError(w, r, newError(http.StatusConflict, "already_reviewed",
 			"submission has already been reviewed"))
+	case errors.Is(err, submit.ErrQuotaExceeded):
+		writeError(w, r, newError(http.StatusForbidden, "submission_quota_exceeded",
+			"submission quota reached"))
+	case errors.Is(err, submit.ErrBlobNotFound):
+		writeError(w, r, newError(http.StatusNotFound, "not_found", "no context uploaded for this submission"))
 	case errors.Is(err, submit.ErrUploadsUnavailable):
 		writeError(w, r, newError(http.StatusServiceUnavailable, "uploads_unavailable",
 			"modpack upload transport is not configured"))
@@ -219,5 +382,69 @@ func writeSubmitError(w http.ResponseWriter, r *http.Request, err error) {
 	}
 }
 
+// handleInternalSubmissionContext streams a submission's stored build-context
+// tarball to the build Pod's `felis fetch-context` initContainer. It lives on the
+// internal face (service-token, no Zero Trust) because its only caller is
+// in-cluster infrastructure: the build Job runs in the build namespace, where it
+// can neither mount the uploads PVC nor hold object-store credentials, so the API
+// — which wrote the blob — is the transport. The blob is served verbatim; the
+// fetcher extracts it under a zip-slip guard, and Kaniko treats the result as
+// hostile regardless (spec §16).
+func (a *API) handleInternalSubmissionContext(w http.ResponseWriter, r *http.Request) {
+	rc, ok := a.openSubmissionContext(w, r)
+	if !ok {
+		return
+	}
+	streamSubmissionContext(w, rc)
+}
+
+// handleAdminSubmissionContext streams a submission's stored build-context
+// tarball to a reviewing admin (admin-tier). Review is only a real gate if the
+// reviewer can inspect what they approve: the executed Dockerfile lives INSIDE
+// this tarball (build/jobspec.go pins --dockerfile=Dockerfile), so without this
+// route the human gate could not see the recipe at all. The bytes are the same
+// ones the build Pod fetches over the internal face; the attachment disposition
+// makes the browser download the attacker-supplied archive, never render it.
+func (a *API) handleAdminSubmissionContext(w http.ResponseWriter, r *http.Request) {
+	rc, ok := a.openSubmissionContext(w, r)
+	if !ok {
+		return
+	}
+	w.Header().Set("Content-Disposition", `attachment; filename="context.tar.gz"`)
+	w.Header().Set("X-Content-Type-Options", "nosniff")
+	a.audit(r, principalFromContext(r.Context()).Email, "submission.context.download", r.PathValue("id"))
+	streamSubmissionContext(w, rc)
+}
+
+// openSubmissionContext resolves the build-context blob named in the request
+// path, mapping the submit-layer errors onto the shared submission statuses (a
+// missing blob is 404, an unwired transport 503). On failure the error response
+// is already written and the caller must return.
+func (a *API) openSubmissionContext(w http.ResponseWriter, r *http.Request) (io.ReadCloser, bool) {
+	if a.Submissions == nil {
+		writeError(w, r, errSubmissionsUnavailable)
+		return nil, false
+	}
+	rc, err := a.Submissions.OpenContext(r.Context(), r.PathValue("id"))
+	if err != nil {
+		writeSubmitError(w, r, err)
+		return nil, false
+	}
+	return rc, true
+}
+
+// streamSubmissionContext copies the blob to w verbatim and closes it. The
+// caller must have set every header already: the copy commits the response, so
+// a failure mid-stream can only truncate it.
+func streamSubmissionContext(w http.ResponseWriter, rc io.ReadCloser) {
+	defer rc.Close()
+	w.Header().Set("Content-Type", "application/gzip")
+	if _, err := io.Copy(w, rc); err != nil {
+		// The status is already committed; the client sees a truncated stream and
+		// the fetch fails on size/extract, so there is nothing left to write here.
+		return
+	}
+}
+
 // Compile-time proof that the production Manager satisfies the API interface.
 var _ SubmissionService = (*submit.Manager)(nil)

internal/api/submissions_test.go

@@ -7,8 +7,11 @@ import (
 	"fmt"
 	"io"
 	"net/http"
+	"strings"
 	"testing"
+	"time"
 
+	"felis.lolicon.best/internal/build"
 	"felis.lolicon.best/internal/submit"
 )
 
@@ -17,24 +20,32 @@ import (
 // what the handler forwarded (the point of the owner-scoping checks: the
 // submitter and reviewer must come from the principal, never the body).
 type fakeSubmissions struct {
-	created    *submit.CreateRequest
+	created     *submit.CreateRequest
-	createErr  error
+	createErr   error
-	uploadedID string
+	uploadedID  string
-	uploadedBy string
+	uploadedBy  string
-	uploadedN  int64
+	uploadedN   int64
-	uploadErr  error
+	uploadErr   error
-	listedBy   string
+	listedBy    string
-	byResult   []submit.Submission
+	byResult    []submit.Submission
-	byErr      error
+	byErr       error
-	listed     []submit.Submission
+	listed      []submit.Submission
-	listErr    error
+	listErr     error
-	approvedID string
+	approvedID  string
-	approvedBy string
+	approvedBy  string
-	approveErr error
+	approveErr  error
-	rejectedID string
+	rejectedID  string
-	rejectedBy string
+	rejectedBy  string
-	rejectReas string
+	rejectReas  string
-	rejectErr  error
+	rejectErr   error
+	withdrawnID string
+	withdrawBy  string
+	withdrawErr error
+	deletedID   string
+	deleteErr   error
+	openedID    string
+	openBody    string
+	openErr     error
 }
 
 func (f *fakeSubmissions) Create(_ context.Context, req submit.CreateRequest) (*submit.Submission, error) {
@@ -82,6 +93,32 @@ func (f *fakeSubmissions) Reject(_ context.Context, id, reviewedBy, reason strin
 	return &submit.Submission{ID: id, Status: submit.StatusRejected, ReviewedBy: reviewedBy, RejectReason: reason}, nil
 }
 
+func (f *fakeSubmissions) Withdraw(_ context.Context, id, submittedBy string) (*submit.Submission, error) {
+	f.withdrawnID, f.withdrawBy = id, submittedBy
+	if f.withdrawErr != nil {
+		return nil, f.withdrawErr
+	}
+	return &submit.Submission{ID: id, SubmittedBy: submittedBy, Status: submit.StatusPendingReview}, nil
+}
+
+func (f *fakeSubmissions) Delete(_ context.Context, id string) (*submit.Submission, error) {
+	f.deletedID = id
+	if f.deleteErr != nil {
+		return nil, f.deleteErr
+	}
+	return &submit.Submission{ID: id, Status: submit.StatusRejected}, nil
+}
+
+// openErr injects the OpenContext outcome; the body recorder lets the internal
+// route test assert byte-exact streaming and the 404 mapping.
+func (f *fakeSubmissions) OpenContext(_ context.Context, id string) (io.ReadCloser, error) {
+	f.openedID = id
+	if f.openErr != nil {
+		return nil, f.openErr
+	}
+	return io.NopCloser(strings.NewReader(f.openBody)), nil
+}
+
 // appSubAPI wires a submissions service behind an ordinary user principal (the
 // app tier — /me/submissions). [email protected] / .test are deliberately not the
 // deployment domain.
@@ -233,6 +270,69 @@ func TestUploadSubmissionContextWithoutServiceIs503(t *testing.T) {
 	}
 }
 
+// Withdraw retracts the caller's OWN pending submission: the submitter is the
+// principal (never the body), a reviewed submission is 409, and a foreign id is
+// 404 — the same posture as the upload route.
+func TestWithdrawSubmission(t *testing.T) {
+	t.Run("withdraws as the principal", func(t *testing.T) {
+		fs := &fakeSubmissions{}
+		w := do(appSubAPI(fs).ExternalHandler(), "DELETE", "/api/v1/me/submissions/sub-3", "", nil)
+		if w.Code != http.StatusOK {
+			t.Fatalf("code = %d, want 200 (%s)", w.Code, w.Body.String())
+		}
+		if fs.withdrawnID != "sub-3" || fs.withdrawBy != "user-7" {
+			t.Fatalf("withdraw forwarded (%q, %q), want (sub-3, user-7)", fs.withdrawnID, fs.withdrawBy)
+		}
+	})
+	t.Run("reviewed submission is 409", func(t *testing.T) {
+		fs := &fakeSubmissions{withdrawErr: submit.ErrAlreadyReviewed}
+		w := do(appSubAPI(fs).ExternalHandler(), "DELETE", "/api/v1/me/submissions/sub-3", "", nil)
+		if w.Code != http.StatusConflict {
+			t.Fatalf("code = %d, want 409 (%s)", w.Code, w.Body.String())
+		}
+		if got := decodeErr(t, w); got != "already_reviewed" {
+			t.Errorf("error code = %q, want already_reviewed", got)
+		}
+	})
+	t.Run("foreign or unknown id is 404", func(t *testing.T) {
+		fs := &fakeSubmissions{withdrawErr: submit.ErrNotFound}
+		w := do(appSubAPI(fs).ExternalHandler(), "DELETE", "/api/v1/me/submissions/sub-x", "", nil)
+		if w.Code != http.StatusNotFound {
+			t.Fatalf("code = %d, want 404 (%s)", w.Code, w.Body.String())
+		}
+	})
+	t.Run("no service is 503", func(t *testing.T) {
+		app := appSubAPI(nil)
+		app.Submissions = nil
+		w := do(app.ExternalHandler(), "DELETE", "/api/v1/me/submissions/sub-3", "", nil)
+		if w.Code != http.StatusServiceUnavailable {
+			t.Fatalf("code = %d, want 503 (%s)", w.Code, w.Body.String())
+		}
+	})
+}
+
+// The admin delete retires any submission and maps the lane's 404; the route's
+// admin gate itself is pinned by TestSubmissionAdminRoutesAreAdminOnly.
+func TestDeleteSubmissionAdmin(t *testing.T) {
+	t.Run("deletes the named row", func(t *testing.T) {
+		fs := &fakeSubmissions{}
+		w := do(adminSubAPI(fs).ExternalHandler(), "DELETE", "/api/v1/submissions/sub-8", "", nil)
+		if w.Code != http.StatusOK {
+			t.Fatalf("code = %d, want 200 (%s)", w.Code, w.Body.String())
+		}
+		if fs.deletedID != "sub-8" {
+			t.Fatalf("delete forwarded id %q, want sub-8", fs.deletedID)
+		}
+	})
+	t.Run("unknown is 404", func(t *testing.T) {
+		fs := &fakeSubmissions{deleteErr: submit.ErrNotFound}
+		w := do(adminSubAPI(fs).ExternalHandler(), "DELETE", "/api/v1/submissions/sub-x", "", nil)
+		if w.Code != http.StatusNotFound {
+			t.Fatalf("code = %d, want 404 (%s)", w.Code, w.Body.String())
+		}
+	})
+}
+
 // The "my uploads" list scopes strictly to the principal's id — there is no
 // parameter that could widen it to another user's submissions.
 func TestMySubmissionsScopesToPrincipal(t *testing.T) {
@@ -256,6 +356,68 @@ func TestMySubmissionsScopesToPrincipal(t *testing.T) {
 	}
 }
 
+// The "my uploads" list carries the linked build's outcome — for a submitter it
+// is the only visible outlet for a failed build (the /images/build routes are
+// admin-tier). A row with no linked build gains no build fields.
+func TestMySubmissionsCarriesBuildOutcome(t *testing.T) {
+	fs := &fakeSubmissions{byResult: []submit.Submission{
+		{ID: "sub-1", SubmittedBy: "user-7", Status: submit.StatusApproved, BuildID: "bld-9"},
+		{ID: "sub-2", SubmittedBy: "user-7", Status: submit.StatusPendingReview},
+	}}
+	api := appSubAPI(fs)
+	api.Builder = &fakeBuilder{getBuilds: map[string]*build.Build{
+		"bld-9": {ID: "bld-9", Status: build.StatusFailed, Error: "build job failed or scan found a CRITICAL CVE"},
+	}}
+	w := do(api.ExternalHandler(), "GET", "/api/v1/me/submissions", "", nil)
+	if w.Code != http.StatusOK {
+		t.Fatalf("code = %d, want 200 (%s)", w.Code, w.Body.String())
+	}
+	var got struct {
+		Submissions []struct {
+			ID          string `json:"id"`
+			BuildStatus string `json:"build_status"`
+			BuildError  string `json:"build_error"`
+		} `json:"submissions"`
+	}
+	if err := json.Unmarshal(w.Body.Bytes(), &got); err != nil {
+		t.Fatalf("body not JSON: %v", err)
+	}
+	if len(got.Submissions) != 2 {
+		t.Fatalf("submissions = %d, want 2", len(got.Submissions))
+	}
+	if got.Submissions[0].BuildStatus != "failed" || got.Submissions[0].BuildError == "" {
+		t.Errorf("sub-1 outcome = %+v, want failed with the error text", got.Submissions[0])
+	}
+	if got.Submissions[1].BuildStatus != "" || got.Submissions[1].BuildError != "" {
+		t.Errorf("sub-2 outcome = %+v, want no build fields without a linked build", got.Submissions[1])
+	}
+}
+
+// A linked build whose row is gone renders as "no outcome" rather than failing
+// the whole list; any other lookup failure must surface, never be swallowed.
+func TestMySubmissionsBuildLookupSemantics(t *testing.T) {
+	// Missing row (ErrNotFound): 200 with no build fields.
+	fs := &fakeSubmissions{byResult: []submit.Submission{{ID: "sub-1", BuildID: "bld-gone"}}}
+	api := appSubAPI(fs)
+	api.Builder = &fakeBuilder{}
+	w := do(api.ExternalHandler(), "GET", "/api/v1/me/submissions", "", nil)
+	if w.Code != http.StatusOK {
+		t.Fatalf("missing build row: code = %d, want 200 (%s)", w.Code, w.Body.String())
+	}
+	if strings.Contains(w.Body.String(), "build_status") {
+		t.Errorf("missing build row: body carries build fields: %s", w.Body.String())
+	}
+
+	// Store fault: the failure is reported, not hidden behind a 200.
+	fs = &fakeSubmissions{byResult: []submit.Submission{{ID: "sub-1", BuildID: "bld-1"}}}
+	api = appSubAPI(fs)
+	api.Builder = &fakeBuilder{getErr: errors.New("db down")}
+	w = do(api.ExternalHandler(), "GET", "/api/v1/me/submissions", "", nil)
+	if w.Code != http.StatusInternalServerError {
+		t.Fatalf("store fault: code = %d, want 500 (%s)", w.Code, w.Body.String())
+	}
+}
+
 // Every /submissions route is admin-tier: a plain user is rejected before the
 // handler runs.
 func TestSubmissionAdminRoutesAreAdminOnly(t *testing.T) {
@@ -265,6 +427,8 @@ func TestSubmissionAdminRoutesAreAdminOnly(t *testing.T) {
 		{"GET", "/api/v1/submissions", ""},
 		{"POST", "/api/v1/submissions/sub-1/approve", ""},
 		{"POST", "/api/v1/submissions/sub-1/reject", `{"reason":"no"}`},
+		{"DELETE", "/api/v1/submissions/sub-1", ""},
+		{"GET", "/api/v1/submissions/sub-1/context", ""},
 	}
 	for _, c := range cases {
 		api := adminSubAPI(&fakeSubmissions{})
@@ -280,9 +444,12 @@ func TestSubmissionAdminRoutesAreAdminOnly(t *testing.T) {
 func TestListSubmissionsAdmin(t *testing.T) {
 	fs := &fakeSubmissions{listed: []submit.Submission{
 		{ID: "sub-1", SubmittedBy: "user-7", Status: submit.StatusPendingReview},
-		{ID: "sub-2", SubmittedBy: "user-9", Status: submit.StatusApproved},
+		{ID: "sub-2", SubmittedBy: "user-9", Status: submit.StatusApproved, BuildID: "bld-2"},
 	}}
 	api := adminSubAPI(fs)
+	api.Builder = &fakeBuilder{getBuilds: map[string]*build.Build{
+		"bld-2": {ID: "bld-2", Status: build.StatusSucceeded},
+	}}
 	w := do(api.ExternalHandler(), "GET", "/api/v1/submissions", "", nil)
 	if w.Code != http.StatusOK {
 		t.Fatalf("code = %d, want 200 (%s)", w.Code, w.Body.String())
@@ -294,6 +461,10 @@ func TestListSubmissionsAdmin(t *testing.T) {
 	if len(got["submissions"]) != 2 {
 		t.Fatalf("submissions = %d, want 2", len(got["submissions"]))
 	}
+	// The admin queue carries the same build outcome enrichment.
+	if !strings.Contains(w.Body.String(), `"build_status":"succeeded"`) {
+		t.Errorf("admin queue lacks the linked build outcome: %s", w.Body.String())
+	}
 }
 
 // The reviewer is the admin principal's email, never client input.
@@ -396,3 +567,198 @@ func TestSubmissionRoutesWithoutServiceAre503(t *testing.T) {
 		t.Fatalf("admin route: code = %d, want 503", w.Code)
 	}
 }
+
+// The internal context route is the build Pod's only read path to a submission's
+// blob: it streams the bytes verbatim, and its error mapping distinguishes a
+// missing blob (404) from an unwired transport (503).
+func TestInternalSubmissionContextRoute(t *testing.T) {
+	newAPI := func(s SubmissionService) *API {
+		api := newTestAPI(newFakeRepo(), newFakeCluster())
+		api.Submissions = s
+		return api
+	}
+
+	t.Run("streams the blob", func(t *testing.T) {
+		fs := &fakeSubmissions{openBody: "\x1f\x8b\x08\x00blob"}
+		w := do(newAPI(fs).InternalHandler(), "GET", "/api/v1/internal/submissions/sub-7/context", "", nil)
+		if w.Code != http.StatusOK {
+			t.Fatalf("code = %d body %s", w.Code, w.Body.String())
+		}
+		if w.Body.String() != fs.openBody {
+			t.Fatalf("body = %q, want the stored blob %q", w.Body.String(), fs.openBody)
+		}
+		if fs.openedID != "sub-7" {
+			t.Fatalf("opened id = %q, want the path id", fs.openedID)
+		}
+		if ct := w.Header().Get("Content-Type"); ct != "application/gzip" {
+			t.Fatalf("content-type = %q, want application/gzip", ct)
+		}
+	})
+
+	t.Run("missing blob is 404", func(t *testing.T) {
+		fs := &fakeSubmissions{openErr: fmt.Errorf("%w: gone", submit.ErrBlobNotFound)}
+		w := do(newAPI(fs).InternalHandler(), "GET", "/api/v1/internal/submissions/sub-7/context", "", nil)
+		if w.Code != http.StatusNotFound {
+			t.Fatalf("code = %d, want 404", w.Code)
+		}
+	})
+
+	t.Run("unwired transport is 503", func(t *testing.T) {
+		w := do(newAPI(nil).InternalHandler(), "GET", "/api/v1/internal/submissions/sub-7/context", "", nil)
+		if w.Code != http.StatusServiceUnavailable {
+			t.Fatalf("code = %d, want 503", w.Code)
+		}
+	})
+}
+
+// The admin context route is the reviewer's read path to the blob they are
+// approving: an admin streams the stored bytes with a download disposition,
+// and the error mapping matches the internal route (404 missing, 503 unwired).
+// The admin-only gate itself is pinned by TestSubmissionAdminRoutesAreAdminOnly.
+func TestAdminSubmissionContextRoute(t *testing.T) {
+	t.Run("streams the blob with a download disposition", func(t *testing.T) {
+		fs := &fakeSubmissions{openBody: "\x1f\x8b\x08\x00blob"}
+		w := do(adminSubAPI(fs).ExternalHandler(), "GET", "/api/v1/submissions/sub-7/context", "", nil)
+		if w.Code != http.StatusOK {
+			t.Fatalf("code = %d body %s", w.Code, w.Body.String())
+		}
+		if w.Body.String() != fs.openBody {
+			t.Fatalf("body = %q, want the stored blob %q", w.Body.String(), fs.openBody)
+		}
+		if fs.openedID != "sub-7" {
+			t.Fatalf("opened id = %q, want the path id", fs.openedID)
+		}
+		if ct := w.Header().Get("Content-Type"); ct != "application/gzip" {
+			t.Fatalf("content-type = %q, want application/gzip", ct)
+		}
+		if cd := w.Header().Get("Content-Disposition"); cd != `attachment; filename="context.tar.gz"` {
+			t.Fatalf("content-disposition = %q", cd)
+		}
+	})
+
+	t.Run("missing blob is 404", func(t *testing.T) {
+		fs := &fakeSubmissions{openErr: fmt.Errorf("%w: gone", submit.ErrBlobNotFound)}
+		w := do(adminSubAPI(fs).ExternalHandler(), "GET", "/api/v1/submissions/sub-7/context", "", nil)
+		if w.Code != http.StatusNotFound {
+			t.Fatalf("code = %d, want 404", w.Code)
+		}
+	})
+
+	t.Run("unwired transport is 503", func(t *testing.T) {
+		api := adminSubAPI(nil)
+		api.Submissions = nil
+		w := do(api.ExternalHandler(), "GET", "/api/v1/submissions/sub-7/context", "", nil)
+		if w.Code != http.StatusServiceUnavailable {
+			t.Fatalf("code = %d, want 503", w.Code)
+		}
+	})
+}
+
+// A spent per-user allowance is 403 submission_quota_exceeded on both the create
+// and the upload path — distinctly NOT the 400 a malformed request gets, and not
+// the 429 the cooldown answers with.
+func TestSubmissionQuotaIs403(t *testing.T) {
+	t.Run("create", func(t *testing.T) {
+		fs := &fakeSubmissions{createErr: fmt.Errorf("%w: 5 submissions are already awaiting review", submit.ErrQuotaExceeded)}
+		w := do(appSubAPI(fs).ExternalHandler(), "POST", "/api/v1/me/submissions", `{"display_name":"Pack"}`, nil)
+		if w.Code != http.StatusForbidden {
+			t.Fatalf("code = %d, want 403 (%s)", w.Code, w.Body.String())
+		}
+		if got := decodeErr(t, w); got != "submission_quota_exceeded" {
+			t.Errorf("error code = %q, want submission_quota_exceeded", got)
+		}
+	})
+	t.Run("upload", func(t *testing.T) {
+		fs := &fakeSubmissions{uploadErr: fmt.Errorf("%w: exceeds your remaining storage allowance", submit.ErrQuotaExceeded)}
+		w := do(appSubAPI(fs).ExternalHandler(), "POST", "/api/v1/me/submissions/sub-9/context", "\x1f\x8bdata", nil)
+		if w.Code != http.StatusForbidden {
+			t.Fatalf("code = %d, want 403 (%s)", w.Code, w.Body.String())
+		}
+		if got := decodeErr(t, w); got != "submission_quota_exceeded" {
+			t.Errorf("error code = %q, want submission_quota_exceeded", got)
+		}
+	})
+}
+
+// The per-user create cooldown bounds review-queue growth: a second create in
+// the same window is 429 submission_cooldown and never reaches the service; the
+// window recovers afterwards.
+func TestCreateSubmissionRateLimited(t *testing.T) {
+	fs := &fakeSubmissions{}
+	api := appSubAPI(fs)
+	clock := time.Unix(1_700_000_000, 0)
+	api.Now = func() time.Time { return clock }
+	api.SubmitCreateCooldown = time.Minute
+	eh := api.ExternalHandler()
+
+	if w := do(eh, "POST", "/api/v1/me/submissions", `{"display_name":"First"}`, nil); w.Code != http.StatusCreated {
+		t.Fatalf("first create: code = %d, want 201 (%s)", w.Code, w.Body.String())
+	}
+	w := do(eh, "POST", "/api/v1/me/submissions", `{"display_name":"Second"}`, nil)
+	if w.Code != http.StatusTooManyRequests || decodeErr(t, w) != "submission_cooldown" {
+		t.Fatalf("immediate second create: code = %d body %s, want 429 submission_cooldown", w.Code, w.Body.String())
+	}
+	// The gate sits before the body handling: even a malformed request is
+	// refused while the window is closed, so it cannot be used to probe.
+	if w := do(eh, "POST", "/api/v1/me/submissions", `{`, nil); w.Code != http.StatusTooManyRequests {
+		t.Fatalf("malformed create during cooldown: code = %d, want 429", w.Code)
+	}
+	clock = clock.Add(time.Minute + time.Second)
+	if w := do(eh, "POST", "/api/v1/me/submissions", `{"display_name":"Third"}`, nil); w.Code != http.StatusCreated {
+		t.Fatalf("post-cooldown create: code = %d, want 201 (%s)", w.Code, w.Body.String())
+	}
+}
+
+// A failed create frees the window: only a row that was actually recorded burns
+// the cooldown, so a validation typo is not punished with a wait.
+func TestCreateSubmissionFailureDoesNotBurnCooldown(t *testing.T) {
+	fs := &fakeSubmissions{createErr: fmt.Errorf("%w: display name is required", submit.ErrInvalid)}
+	api := appSubAPI(fs)
+	api.Now = func() time.Time { return time.Unix(1_700_000_000, 0) }
+	api.SubmitCreateCooldown = time.Minute
+	eh := api.ExternalHandler()
+
+	if w := do(eh, "POST", "/api/v1/me/submissions", `{"display_name":""}`, nil); w.Code != http.StatusBadRequest {
+		t.Fatalf("failed create: code = %d, want 400", w.Code)
+	}
+	fs.createErr = nil
+	if w := do(eh, "POST", "/api/v1/me/submissions", `{"display_name":"Fixed"}`, nil); w.Code != http.StatusCreated {
+		t.Fatalf("retry at the same instant: code = %d, want 201 (%s)", w.Code, w.Body.String())
+	}
+}
+
+// The per-user upload cooldown bounds context streaming: a second upload in the
+// same window is 429 submission_cooldown, and a FAILED upload frees the window
+// for an immediate retry.
+func TestUploadSubmissionContextRateLimited(t *testing.T) {
+	fs := &fakeSubmissions{}
+	api := appSubAPI(fs)
+	clock := time.Unix(1_700_000_000, 0)
+	api.Now = func() time.Time { return clock }
+	api.SubmitUploadCooldown = time.Minute
+	eh := api.ExternalHandler()
+	body := "\x1f\x8b\x08\x00 the modpack bytes"
+
+	if w := do(eh, "POST", "/api/v1/me/submissions/sub-9/context", body, nil); w.Code != http.StatusOK {
+		t.Fatalf("first upload: code = %d, want 200 (%s)", w.Code, w.Body.String())
+	}
+	w := do(eh, "POST", "/api/v1/me/submissions/sub-9/context", body, nil)
+	if w.Code != http.StatusTooManyRequests || decodeErr(t, w) != "submission_cooldown" {
+		t.Fatalf("immediate second upload: code = %d body %s, want 429 submission_cooldown", w.Code, w.Body.String())
+	}
+
+	// A failed upload releases its reservation, so the user is not punished for
+	// a genuine failure (aborted transfer, spent quota) with a cooldown wait.
+	fs2 := &fakeSubmissions{uploadErr: submit.ErrUploadsUnavailable}
+	api2 := appSubAPI(fs2)
+	api2.Now = func() time.Time { return clock }
+	api2.SubmitUploadCooldown = time.Minute
+	eh2 := api2.ExternalHandler()
+	if w := do(eh2, "POST", "/api/v1/me/submissions/sub-9/context", body, nil); w.Code != http.StatusServiceUnavailable {
+		t.Fatalf("failed upload: code = %d, want 503", w.Code)
+	}
+	fs2.uploadErr = nil
+	if w := do(eh2, "POST", "/api/v1/me/submissions/sub-9/context", body, nil); w.Code != http.StatusOK {
+		t.Fatalf("retry at the same instant after failure: code = %d, want 200 (%s)", w.Code, w.Body.String())
+	}
+}

internal/backupjob/backup.go

@@ -87,9 +87,14 @@ type Config struct {
 	// CPULimit / MemLimit cap the backup container.
 	CPULimit string
 	MemLimit string
-	// RunAsUser / RunAsGroup / FSGroup are the Pod's runtime identity. FSGroup MUST
+	// RunAsUser / RunAsGroup / FSGroup are the Pod's runtime identity. They default
-	// match the operator StatefulSet's runtime group so the read-only world mount is
+	// to ROOT (0:0) for the same reason the operator's forwarding-init container
-	// readable by this Pod's uid.
+	// runs as root: the world volume is written by the game image's own UID (root
+	// for every Paper image we ship), and Paper saves files a non-root uid can
+	// never read — level.dat is written mode 0600 (tar walk: permission denied,
+	// verified live). DAC_OVERRIDE on the container covers images whose UID is
+	// neither root nor ours. Set 0/0/0 explicitly for root; FSGroup is omitted
+	// when zero.
 	RunAsUser  int64
 	RunAsGroup int64
 	FSGroup    int64
@@ -111,7 +116,6 @@ const (
 	defaultDeadline       = 30 * time.Minute
 	defaultCPULimit       = "1"
 	defaultMemLimit       = "1Gi"
-	defaultRunAsID        = int64(1000)
 	defaultTTL            = 10 * time.Minute
 )
 
@@ -145,15 +149,6 @@ func (c Config) withDefaults() Config {
 	if c.MemLimit == "" {
 		c.MemLimit = defaultMemLimit
 	}
-	if c.RunAsUser == 0 {
-		c.RunAsUser = defaultRunAsID
-	}
-	if c.RunAsGroup == 0 {
-		c.RunAsGroup = defaultRunAsID
-	}
-	if c.FSGroup == 0 {
-		c.FSGroup = defaultRunAsID
-	}
 	if c.TTLAfterFinished <= 0 {
 		c.TTLAfterFinished = defaultTTL
 	}

internal/backupjob/jobspec.go

@@ -148,7 +148,14 @@ func BackupJob(p JobParams) (*batchv1.Job, error) {
 			Privileged:               boolPtr(false),
 			AllowPrivilegeEscalation: boolPtr(false),
 			ReadOnlyRootFilesystem:   boolPtr(true),
-			Capabilities:             &corev1.Capabilities{Drop: []corev1.Capability{"ALL"}},
+			// DAC_OVERRIDE is granted on top of dropping ALL: the pod runs as root,
+			// but the world may have been written by a game image whose UID is
+			// neither root nor ours, and Paper's own files are mode 0600. It is the
+			// minimal extra power that makes the archive read every world shape.
+			Capabilities: &corev1.Capabilities{
+				Drop: []corev1.Capability{"ALL"},
+				Add:  []corev1.Capability{"DAC_OVERRIDE"},
+			},
 		},
 	}
 
@@ -175,13 +182,12 @@ func BackupJob(p JobParams) (*batchv1.Job, error) {
 					RestartPolicy:                corev1.RestartPolicyNever,
 					ServiceAccountName:           p.ServiceAccount,
 					AutomountServiceAccountToken: boolPtr(false),
-					SecurityContext: &corev1.PodSecurityContext{
+					// Root by default (see Config.RunAsUser): the world volume's
-						RunAsNonRoot: boolPtr(true),
+					// owner is the game image's UID, so only an owner-matching or
-						RunAsUser:    int64Ptr(p.RunAsUser),
+					// DAC-overriding uid can read it. FSGroup is omitted when unset
-						RunAsGroup:   int64Ptr(p.RunAsGroup),
+					// so a root pod never triggers a volume chgrp.
-						FSGroup:      int64Ptr(p.FSGroup),
+					SecurityContext: backupPodSecurityContext(p),
-					},
+					Containers:      []corev1.Container{container},
-					Containers: []corev1.Container{container},
 					Volumes: []corev1.Volume{
 						{
 							Name: worldVolume,
@@ -240,3 +246,20 @@ func resourceLimits(cpu, mem string) (corev1.ResourceList, error) {
 func boolPtr(b bool) *bool    { return &b }
 func int32Ptr(i int32) *int32 { return &i }
 func int64Ptr(i int64) *int64 { return &i }
+
+// backupPodSecurityContext pins the Pod identity. RunAsNonRoot is false because
+// the default identity is root: worlds are owned by the game image's UID (root
+// for the images we ship), and Paper writes mode-0600 files a non-root reader
+// cannot open. FSGroup stays unset unless configured — a root executor must not
+// needlessly chgrp the world volume.
+func backupPodSecurityContext(p JobParams) *corev1.PodSecurityContext {
+	sc := &corev1.PodSecurityContext{
+		RunAsNonRoot: boolPtr(false),
+		RunAsUser:    int64Ptr(p.RunAsUser),
+		RunAsGroup:   int64Ptr(p.RunAsGroup),
+	}
+	if p.FSGroup > 0 {
+		sc.FSGroup = int64Ptr(p.FSGroup)
+	}
+	return sc
+}

internal/backupjob/jobspec_test.go

@@ -23,9 +23,9 @@ func sampleJobParams() JobParams {
 		Deadline:         30 * time.Minute,
 		CPULimit:         "1",
 		MemLimit:         "1Gi",
-		RunAsUser:        1000,
+		RunAsUser:        0,
-		RunAsGroup:       1000,
+		RunAsGroup:       0,
-		FSGroup:          1000,
+		FSGroup:          0,
 		TTLAfterFinished: 10 * time.Minute,
 	}
 }
@@ -109,13 +109,30 @@ func TestBackupJobMountsTwoPVCsPlusConfigSecretOnly(t *testing.T) {
 	}
 }
 
-// The backup container is hardened exactly like the restore/build Job containers:
+// The backup container is hardened like the restore/build Job containers: no
-// no privilege, no escalation, read-only root fs, drop ALL capabilities.
+// privilege, no escalation, read-only root fs, ALL capabilities dropped — plus
+// DAC_OVERRIDE, because the Pod runs as root and the world may have been written
+// by a game image with a different UID (verified live: a uid-1000 executor cannot
+// read Paper's mode-0600 level.dat).
 func TestBackupJobContainerIsHardened(t *testing.T) {
 	job, err := BackupJob(sampleJobParams())
 	if err != nil {
 		t.Fatalf("BackupJob: %v", err)
 	}
+	pod := job.Spec.Template.Spec
+	if pod.SecurityContext == nil {
+		t.Fatal("pod SecurityContext is nil")
+	}
+	if pod.SecurityContext.RunAsNonRoot == nil || *pod.SecurityContext.RunAsNonRoot {
+		t.Error("pod must NOT require non-root: root is the owner-matching default for game-image worlds")
+	}
+	if pod.SecurityContext.RunAsUser == nil || *pod.SecurityContext.RunAsUser != 0 ||
+		pod.SecurityContext.RunAsGroup == nil || *pod.SecurityContext.RunAsGroup != 0 {
+		t.Errorf("pod must run as 0:0 by default, got %+v", pod.SecurityContext)
+	}
+	if pod.SecurityContext.FSGroup != nil {
+		t.Error("fsGroup must stay unset when zero (a root executor must not chgrp the world volume)")
+	}
 	sc := job.Spec.Template.Spec.Containers[0].SecurityContext
 	if sc == nil {
 		t.Fatal("container SecurityContext is nil")
@@ -132,6 +149,9 @@ func TestBackupJobContainerIsHardened(t *testing.T) {
 	if sc.Capabilities == nil || len(sc.Capabilities.Drop) != 1 || sc.Capabilities.Drop[0] != "ALL" {
 		t.Error("capabilities must drop ALL")
 	}
+	if len(sc.Capabilities.Add) != 1 || sc.Capabilities.Add[0] != "DAC_OVERRIDE" {
+		t.Errorf("capabilities must add exactly DAC_OVERRIDE, got %v", sc.Capabilities.Add)
+	}
 }
 
 // One-shot: a wedged archive must not loop, and a deadline caps it.

internal/backupjob/k8sjobs.go

@@ -24,10 +24,10 @@ func NewK8sJobs(c client.Client) *K8sJobs {
 	return &K8sJobs{c: c}
 }
 
-// CreateBackupJob renders and applies the backup Job. Its name is a deterministic
+// CreateBackupJob renders and applies the backup Job. Its name carries a fresh
-// function of the server (BackupJobName), so a concurrent backup of the same server
+// random suffix (see Backuper.Backup), so a repeated backup never collides with a
-// collides on Create; that collision is mapped to ErrAlreadyExists, which the
+// just-finished Job inside its TTL window; ErrAlreadyExists survives only as the
-// Backuper treats as success (idempotent enqueue).
+// defensive no-op for the astronomically unlikely suffix collision.
 func (k *K8sJobs) CreateBackupJob(ctx context.Context, p JobParams) error {
 	job, err := BackupJob(p)
 	if err != nil {

internal/build/build.go

@@ -215,6 +215,21 @@ type Config struct {
 	// RegistryURL is the internal registry the build pushes to and Trivy scans
 	// (spec §17). Image refs are validated to be under it.
 	RegistryURL string
+	// FelisImage is the platform image whose `fetch-context` entrypoint streams a
+	// submission's context from the internal face into the build Pod. Required
+	// only when a build's ContextRef is an http(s) URL (the submit lane's derived
+	// shape); an install that never builds user submissions can leave it empty.
+	FelisImage string
+	// TrivyDBRepository overrides where Trivy fetches its vulnerability DB
+	// (--db-repository). Empty keeps Trivy's upstream default, which the build
+	// egress lock denies — an install with builds must point this at an internal
+	// mirror (see config.RegistryConfig.TrivyDBRepository).
+	TrivyDBRepository string
+	// TrivyJavaDBRepository overrides where Trivy fetches its Java DB
+	// (--java-db-repository), downloaded lazily for images that contain Java
+	// artifacts — i.e. every real modpack. Same egress story as the
+	// vulnerability DB (see config.RegistryConfig.TrivyJavaDBRepository).
+	TrivyJavaDBRepository string
 	// KanikoImage / TrivyImage are the executor images.
 	KanikoImage string
 	TrivyImage  string
@@ -349,17 +364,20 @@ func (b *Builder) Submit(ctx context.Context, req Request) (*Build, error) {
 // jobParams projects a build + config onto the inputs jobspec.go renders.
 func (b *Builder) jobParams(bld *Build, cfg Config) JobParams {
 	return JobParams{
-		BuildID:        bld.ID,
+		BuildID:               bld.ID,
-		ImageRef:       bld.ImageRef,
+		ImageRef:              bld.ImageRef,
-		ContextRef:     bld.ContextRef,
+		ContextRef:            bld.ContextRef,
-		Namespace:      cfg.Namespace,
+		Namespace:             cfg.Namespace,
-		ServiceAccount: cfg.ServiceAccount,
+		ServiceAccount:        cfg.ServiceAccount,
-		RegistryURL:    cfg.RegistryURL,
+		RegistryURL:           cfg.RegistryURL,
-		KanikoImage:    cfg.KanikoImage,
+		FelisImage:            cfg.FelisImage,
-		TrivyImage:     cfg.TrivyImage,
+		TrivyDBRepository:     cfg.TrivyDBRepository,
-		Deadline:       cfg.Deadline,
+		TrivyJavaDBRepository: cfg.TrivyJavaDBRepository,
-		CPULimit:       cfg.CPULimit,
+		KanikoImage:           cfg.KanikoImage,
-		MemLimit:       cfg.MemLimit,
+		TrivyImage:            cfg.TrivyImage,
+		Deadline:              cfg.Deadline,
+		CPULimit:              cfg.CPULimit,
+		MemLimit:              cfg.MemLimit,
 	}
 }
 

internal/build/jobspec.go

@@ -2,8 +2,10 @@ package build
 
 import (
 	"fmt"
+	"strings"
 	"time"
 
+	"felis.lolicon.best/internal/naming"
 	batchv1 "k8s.io/api/batch/v1"
 	corev1 "k8s.io/api/core/v1"
 	networkingv1 "k8s.io/api/networking/v1"
@@ -32,8 +34,37 @@ const (
 const (
 	ContainerKaniko = "kaniko"
 	ContainerTrivy  = "trivy"
+	// ContainerFetch is the initContainer that pulls a submission's build context
+	// from the felis-api internal face and extracts it into the shared emptyDir.
+	// It exists only for an http(s) ContextRef (see BuildJob); a ref Kaniko can
+	// read natively (s3://) or a pre-mounted path renders no such container.
+	ContainerFetch = "context-fetch"
+
+	// contextVolume/contextMountPath carry a fetched build context: the fetch
+	// initContainer writes the extracted tree there, Kaniko reads it read-only.
+	contextVolume    = "context"
+	contextMountPath = "/context"
 )
 
+// contextSizeLimit bounds the extracted (attacker-controlled) context tree so a
+// tarball bomb wedges the build pod instead of the node's disk. The compressed
+// upload is capped at 1 GiB by the submit lane; 4 GiB leaves expansion room.
+var contextSizeLimit = resource.MustParse("4Gi")
+
+// buildJobTTL is how long a finished build Job survives before the Job
+// controller deletes it — and with it the Pod whose kaniko log is the admin
+// failure-triage surface (GET /api/v1/images/build/{id}/logs).
+//
+// Every other Job family the platform renders carries a TTL (fileedit 2m,
+// backup/restore 10m); the build lane deliberately keeps a much longer one
+// because the logs are the point. Without ANY TTL the Job and its completed
+// Pod accumulate one pair per build forever: they count against the node's
+// pod budget (110 on stock k3s), grow etcd, and eventually block new builds.
+// Sync already tolerates a vanished Job (JobUnknown → failed; terminal builds
+// are returned unchanged), so a week-old log falling off costs a 404, not a
+// status flip.
+const buildJobTTL = 7 * 24 * time.Hour
+
 // JobParams are the rendered inputs to a build Job. They are derived from a
 // Build + Config by the Builder; jobspec is a pure function of them so the
 // security-critical Job shape is unit-tested without a cluster.
@@ -44,11 +75,23 @@ type JobParams struct {
 	Namespace      string
 	ServiceAccount string
 	RegistryURL    string
-	KanikoImage    string
+	// FelisImage runs the context-fetch initContainer (the felis binary's
-	TrivyImage     string
+	// fetch-context entrypoint). Required when ContextRef is an http(s) URL.
-	Deadline       time.Duration
+	FelisImage string
-	CPULimit       string
+	// TrivyDBRepository overrides Trivy's vulnerability-DB source (the
-	MemLimit       string
+	// --db-repository flag). Empty keeps Trivy's own default; see
+	// build.Config.TrivyDBRepository for why an in-cluster install sets it.
+	TrivyDBRepository string
+	// TrivyJavaDBRepository overrides Trivy's Java-DB source (the
+	// --java-db-repository flag), fetched lazily when the image contains Java
+	// artifacts; empty keeps Trivy's own default, which the build egress lock
+	// denies — a jar-bearing image then fails the scan.
+	TrivyJavaDBRepository string
+	KanikoImage           string
+	TrivyImage            string
+	Deadline              time.Duration
+	CPULimit              string
+	MemLimit              string
 }
 
 // BuildJobName is the deterministic Job name for a build id.
@@ -90,44 +133,144 @@ func BuildJob(p JobParams) (*batchv1.Job, error) {
 		deadline = int64(defaultDeadline / time.Second)
 	}
 
-	// Hardened container security context shared by both build containers: no
+	// Hardened container security context baseline: no privilege, no privilege
-	// privilege, no privilege escalation, drop all capabilities. Kaniko needs a
+	// escalation, drop all capabilities. Kaniko needs a writable root filesystem
-	// writable root filesystem to unpack layers, so we do not force read-only
+	// to unpack layers, so we do not force read-only root here; it also needs a
-	// root here, but it gains no privilege.
+	// minimal capability subset added back (kanikoSec below), while fetch and
+	// trivy run with exactly this baseline.
 	sec := &corev1.SecurityContext{
 		Privileged:               boolPtr(false),
 		AllowPrivilegeEscalation: boolPtr(false),
 		Capabilities:             &corev1.Capabilities{Drop: []corev1.Capability{"ALL"}},
 	}
+	// Kaniko unpacks base-image layers as root, and the tar apply must chown/chmod
+	// files to the owners the layer recorded — impossible under drop-ALL (live:
+	// "failed to get filesystem from image: chown /etc/gshadow: operation not
+	// permitted" for any FROM <base image>; scratch builds masked this because
+	// COPY only ever creates files kaniko itself owns). Add back exactly the caps
+	// the unpack needs and nothing else: CHOWN/FOWNER for the ownership and mode
+	// restore, DAC_OVERRIDE to write entries whose bits would otherwise exclude
+	// even root once the capability-based exemption is gone.
+	kanikoSec := sec.DeepCopy()
+	kanikoSec.Capabilities = &corev1.Capabilities{
+		Drop: []corev1.Capability{"ALL"},
+		Add:  []corev1.Capability{"CHOWN", "DAC_OVERRIDE", "FOWNER"},
+	}
+
+	// The context Kaniko reads. An http(s) ref (the submit lane's derived ref: the
+	// API streams the blob on its internal face, because the build Pod can neither
+	// mount the control-plane uploads PVC across namespaces nor hold object-store
+	// credentials) is first fetched into a shared emptyDir; a ref Kaniko can read
+	// in place (s3://, or a path an installer pre-mounted) passes through untouched.
+	contextPath := p.ContextRef
+	initContainers := []corev1.Container{}
+	var kanikoMounts []corev1.VolumeMount
+	var podVolumes []corev1.Volume
+	if isHTTPContextRef(p.ContextRef) {
+		if p.FelisImage == "" {
+			return nil, fmt.Errorf("build: context ref %q needs FelisImage for the fetch initContainer", p.ContextRef)
+		}
+		contextPath = contextMountPath
+		// The fetch container runs as root while Kaniko keeps the image default
+		// (also root): Kaniko re-copies the Dockerfile out of the context and
+		// chowns/chmods it to the SOURCE file's owner, which fails for any other
+		// owner without CAP_CHOWN/CAP_FOWNER — capabilities this pod deliberately
+		// drops (the live drill hit exactly this: "copying dockerfile: chown
+		// /kaniko/Dockerfile: operation not permitted" with the distroless uid
+		// 65532). Extracting as root, the uid Kaniko itself runs as, keeps the
+		// context owned by the only user that can satisfy that copy. The pod is
+		// root by necessity regardless: Kaniko unpacks base-image layers into its
+		// own filesystem.
+		fetchSec := sec.DeepCopy()
+		fetchSec.RunAsUser = int64Ptr(0)
+		fetchSec.RunAsGroup = int64Ptr(0)
+		fetch := corev1.Container{
+			Name:  ContainerFetch,
+			Image: p.FelisImage,
+			Args: []string{
+				"fetch-context",
+				"--url=" + p.ContextRef,
+				"--out=" + contextMountPath,
+			},
+			// The internal face is service-token gated, and the token is read from a
+			// Secret the installer materializes in THIS namespace (secretKeyRef is
+			// namespace-local). It is mounted into this initContainer only: the Kaniko
+			// container executes the untrusted Dockerfile and must never hold it, and
+			// pod containers share neither environment nor PID namespace.
+			Env: []corev1.EnvVar{{
+				Name: "FELIS_SERVICE_TOKEN",
+				ValueFrom: &corev1.EnvVarSource{SecretKeyRef: &corev1.SecretKeySelector{
+					LocalObjectReference: corev1.LocalObjectReference{Name: naming.ServiceTokenSecretName},
+					Key:                  naming.ServiceTokenSecretKey,
+				}},
+			}},
+			VolumeMounts:    []corev1.VolumeMount{{Name: contextVolume, MountPath: contextMountPath}},
+			Resources:       corev1.ResourceRequirements{Limits: limits, Requests: buildRequests(limits)},
+			SecurityContext: fetchSec,
+		}
+		initContainers = append(initContainers, fetch)
+		kanikoMounts = []corev1.VolumeMount{{Name: contextVolume, MountPath: contextMountPath, ReadOnly: true}}
+		podVolumes = []corev1.Volume{{
+			Name: contextVolume,
+			VolumeSource: corev1.VolumeSource{EmptyDir: &corev1.EmptyDirVolumeSource{
+				// The extracted tree is attacker-controlled; bound it so a tarball
+				// bomb wedges THIS pod (admitted failure) instead of filling the
+				// node's disk. The compressed upload is capped at 1 GiB by the
+				// submit lane, and 4 GiB leaves room for a typical expansion.
+				SizeLimit: sizeLimitPtr(),
+			}},
+		}}
+	}
 
 	kaniko := corev1.Container{
 		Name:  ContainerKaniko,
 		Image: p.KanikoImage,
 		Args: []string{
 			"--dockerfile=Dockerfile",
-			"--context=" + p.ContextRef,
+			"--context=" + contextPath,
 			"--destination=" + p.ImageRef,
 			// The internal registry is in-cluster only and may serve plain HTTP;
-			// it is never a public ingress (spec §17).
+			// it is never a public ingress (spec §17). Both directions need the
+			// insecure flags: --insecure/--skip-tls-verify cover the PUSH, while
+			// the pull side needs its own pair — a Dockerfile's `FROM
+			// registry.felis.svc:5000/...` otherwise fails with "server gave
+			// HTTP response to HTTPS client", breaking every build based on a
+			// platform image (the canonical modpack shape).
 			"--insecure",
 			"--skip-tls-verify",
+			"--insecure-pull",
+			"--skip-tls-verify-pull",
 		},
-		Resources:       corev1.ResourceRequirements{Limits: limits, Requests: limits},
+		VolumeMounts:    kanikoMounts,
-		SecurityContext: sec,
+		Resources:       corev1.ResourceRequirements{Limits: limits, Requests: buildRequests(limits)},
+		SecurityContext: kanikoSec,
 	}
+	initContainers = append(initContainers, kaniko)
 
+	trivyArgs := []string{
+		"image",
+		"--exit-code", "1",
+		"--severity", "CRITICAL",
+		"--no-progress",
+		"--insecure",
+	}
+	// The DB source is configurable because the default (mirror.gcr.io/ghcr.io)
+	// is exactly what the build egress lock denies: an install that never mirrors
+	// the DB cannot complete a scan, and the gate fails closed on purpose. The
+	// supported shape is the internal registry (`--insecure` above already covers
+	// its plain HTTP).
+	if p.TrivyDBRepository != "" {
+		trivyArgs = append(trivyArgs, "--db-repository", p.TrivyDBRepository)
+	}
+	if p.TrivyJavaDBRepository != "" {
+		trivyArgs = append(trivyArgs, "--java-db-repository", p.TrivyJavaDBRepository)
+	}
+	trivyArgs = append(trivyArgs, p.ImageRef)
 	trivy := corev1.Container{
-		Name:  ContainerTrivy,
+		Name:            ContainerTrivy,
-		Image: p.TrivyImage,
+		Image:           p.TrivyImage,
-		Args: []string{
+		Args:            trivyArgs,
-			"image",
+		Resources:       corev1.ResourceRequirements{Limits: limits, Requests: buildRequests(limits)},
-			"--exit-code", "1",
-			"--severity", "CRITICAL",
-			"--no-progress",
-			"--insecure",
-			p.ImageRef,
-		},
-		Resources:       corev1.ResourceRequirements{Limits: limits, Requests: limits},
 		SecurityContext: sec,
 	}
 
@@ -141,14 +284,17 @@ func BuildJob(p JobParams) (*batchv1.Job, error) {
 			// A poisoned build must not loop — one shot, then a terminal verdict.
 			BackoffLimit:          int32Ptr(0),
 			ActiveDeadlineSeconds: int64Ptr(deadline),
+			// ...and a finished one must not linger forever (see buildJobTTL).
+			TTLSecondsAfterFinished: int32Ptr(int32(buildJobTTL / time.Second)),
 			Template: corev1.PodTemplateSpec{
 				ObjectMeta: metav1.ObjectMeta{Labels: buildLabels(p)},
 				Spec: corev1.PodSpec{
 					RestartPolicy:                corev1.RestartPolicyNever,
 					ServiceAccountName:           p.ServiceAccount,
 					AutomountServiceAccountToken: boolPtr(false),
-					InitContainers:               []corev1.Container{kaniko},
+					InitContainers:               initContainers,
 					Containers:                   []corev1.Container{trivy},
+					Volumes:                      podVolumes,
 				},
 			},
 		},
@@ -156,11 +302,24 @@ func BuildJob(p JobParams) (*batchv1.Job, error) {
 	return job, nil
 }
 
+// isHTTPContextRef reports whether ref is an http(s) URL — the shape the submit
+// lane derives when the API is the blob transport — i.e. a context only the
+// fetch initContainer can turn into a local path for Kaniko.
+func isHTTPContextRef(ref string) bool {
+	return strings.HasPrefix(ref, "http://") || strings.HasPrefix(ref, "https://")
+}
+
 // NetPolParams parameterises the build-namespace egress lock.
 type NetPolParams struct {
 	Namespace         string
 	RegistryNamespace string
 	RegistryPort      int32
+	// ControlNamespace and APIPort are where the felis-api internal face lives:
+	// the fetch initContainer's only egress besides DNS and the registry. Both
+	// defaults (felis, 8081) match platform.DefaultControlNamespace and the
+	// internal listener, so an unset Params is still the safe shape.
+	ControlNamespace string
+	APIPort          int32
 	// PackageSourceCIDRs is an optional, explicit allowlist of external package
 	// mirrors (spec §16: egress 仅 registry + 包源). Empty means the most
 	// locked-down default — no internet egress at all (默认拒外网).
@@ -177,10 +336,19 @@ func BuildNetworkPolicy(p NetPolParams) *networkingv1.NetworkPolicy {
 	if port == 0 {
 		port = 5000
 	}
+	controlNS := p.ControlNamespace
+	if controlNS == "" {
+		controlNS = "felis"
+	}
+	apiPort := p.APIPort
+	if apiPort == 0 {
+		apiPort = 8081
+	}
 	dnsUDP := corev1.ProtocolUDP
 	dnsTCP := corev1.ProtocolTCP
 	dns53 := intstr.FromInt32(53)
 	regPort := intstr.FromInt32(port)
+	ctxPort := intstr.FromInt32(apiPort)
 
 	egress := []networkingv1.NetworkPolicyEgressRule{
 		// DNS resolution: port-restricted to 53, so this is not an open-internet
@@ -203,6 +371,21 @@ func BuildNetworkPolicy(p NetPolParams) *networkingv1.NetworkPolicy {
 				{Protocol: &dnsTCP, Port: &regPort},
 			},
 		},
+		// felis-api's internal face, where the fetch initContainer streams the
+		// submission's build context from. Without this rule the build Pod could
+		// not read the context and every user build would fail in its first init
+		// step — the default-deny here is exactly why the transport had to be
+		// planned, not assumed.
+		{
+			To: []networkingv1.NetworkPolicyPeer{{
+				NamespaceSelector: &metav1.LabelSelector{
+					MatchLabels: map[string]string{"kubernetes.io/metadata.name": controlNS},
+				},
+			}},
+			Ports: []networkingv1.NetworkPolicyPort{
+				{Protocol: &dnsTCP, Port: &ctxPort},
+			},
+		},
 	}
 	// Explicit package-mirror CIDRs, when configured. No CIDR ⇒ no internet.
 	for _, cidr := range p.PackageSourceCIDRs {
@@ -279,6 +462,36 @@ func resourceLimits(cpu, mem string) (corev1.ResourceList, error) {
 	}, nil
 }
 
+// buildRequests is the scheduler floor a build container asks for while its
+// configured limit stays the safety cap. Reserving the full cap as a request is
+// what once made a default install on the platform's starter node (4 vCPU /
+// 5.5 GiB) unable to schedule ANY build — caught by the live end-to-end drill, not
+// by any unit test. A build is best-effort batch work: it may be throttled or
+// evicted under contention, which fails the Job loudly, and the caps still stop a
+// runaway build from exhausting the node.
+func buildRequests(limits corev1.ResourceList) corev1.ResourceList {
+	req := corev1.ResourceList{}
+	for res, floor := range map[corev1.ResourceName]resource.Quantity{
+		corev1.ResourceCPU:    resource.MustParse("250m"),
+		corev1.ResourceMemory: resource.MustParse("512Mi"),
+	} {
+		limit, ok := limits[res]
+		if ok && limit.Cmp(floor) < 0 {
+			floor = limit // never ask for more than the cap
+		}
+		req[res] = floor
+	}
+	return req
+}
+
 func boolPtr(b bool) *bool    { return &b }
 func int32Ptr(i int32) *int32 { return &i }
+
+// sizeLimitPtr returns a copy of contextSizeLimit for a VolumeSource (the API
+// object only ever gets serialized, but a shared pointer across rendered Jobs
+// invites accidental aliasing).
+func sizeLimitPtr() *resource.Quantity {
+	q := contextSizeLimit
+	return &q
+}
 func int64Ptr(i int64) *int64 { return &i }