fix: preserve recovery helper traversal

deploy/README.md

@@ -254,7 +254,7 @@ The deploy and migration workflows share the `staging-mutation` Actions concurre
 
 Use Gitea Actions → **Create Production Recovery Point** from the exact current `main` release SHA before every production schema migration. The manual workflow requires `main`, `staging`, public staging health, and the successful release gate to identify the same SHA. It uses the pinned production SSH identity, shares the `production-mutation` lock, and runs `deploy/run-production-recovery.sh` with shell tracing disabled.
 
-The host script writes an AES-256-CBC/PBKDF2 encrypted custom-format PostgreSQL dump under `/opt/jyotisha-production/backups`, restores it into a uniquely named disposable database, validates non-sensitive row/table counts, removes only that disposable database, and writes a mode-`0600` verification manifest. Before acquiring the shared host lock it validates that `.state` and `backups` are real directories, then reuses the existing passwordless Docker boundary to run the already-loaded PostgreSQL image with no network, a read-only root filesystem, all capabilities dropped except `CHOWN`, and only those two directories bind-mounted. That helper restores only the directory mount points and an existing regular `mutation.lock` to the current `deploy` UID/GID; it never recursively changes backup files or removes or replaces a lock inode. The workflow retrieves only the encrypted dump and verification metadata, verifies the SHA-256 digest, and uploads them as a 30-day Gitea Actions artifact with compression disabled. The artifact-backed `recovery_reference`, `recovery_created_at`, and `restore_verified=true` output are the inputs for **Migrate Production Database**. Do not use the attestation if backup, restore, retrieval, digest validation, ownership normalization, or artifact upload fails.
+The host script writes an AES-256-CBC/PBKDF2 encrypted custom-format PostgreSQL dump under `/opt/jyotisha-production/backups`, restores it into a uniquely named disposable database, validates non-sensitive row/table counts, removes only that disposable database, and writes a mode-`0600` verification manifest. Before acquiring the shared host lock it validates that `.state` and `backups` are real directories, then reuses the existing passwordless Docker boundary to run the already-loaded PostgreSQL image with no network, a read-only root filesystem, all capabilities dropped except `CHOWN`, and only those two directories bind-mounted. That helper restores an existing regular `mutation.lock` before changing its mode-`0700` parent directory, then restores the two directory mount points to the current `deploy` UID/GID. The child-first order lets the least-privilege helper retain traversal without adding `DAC_OVERRIDE`; it never recursively changes backup files or removes or replaces a lock inode. The workflow retrieves only the encrypted dump and verification metadata, verifies the SHA-256 digest, and uploads them as a 30-day Gitea Actions artifact with compression disabled. The artifact-backed `recovery_reference`, `recovery_created_at`, and `restore_verified=true` output are the inputs for **Migrate Production Database**. Do not use the attestation if backup, restore, retrieval, digest validation, ownership normalization, or artifact upload fails.
 
 Never restore over `jyotisha`, delete the PostgreSQL volume, print `.env.production.database`, expose `PRODUCTION_BACKUP_ENCRYPTION_KEY`, or substitute a staging recovery artifact. The encrypted local archive is preserved for repair; the Actions artifact supplies the required off-host copy.
 

deploy/run-production-recovery.sh

@@ -35,7 +35,7 @@ done
 install -d -m 700 "$state_directory" "$backup_directory"
 deployment_uid="$(id -u)"
 deployment_gid="$(id -g)"
-ownership_targets=("$state_directory" "$backup_directory")
+ownership_targets=()
 if [ -e "$lock_file" ]; then
   [ -f "$lock_file" ] && [ ! -L "$lock_file" ] || {
     echo "production mutation lock is unsafe" >&2
@@ -43,6 +43,7 @@ if [ -e "$lock_file" ]; then
   }
   ownership_targets+=("$lock_file")
 fi
+ownership_targets+=("$state_directory" "$backup_directory")
 
 mapfile -t postgres_containers < <(
   sudo -n docker ps -q \

docs/BUG_HISTORY.md

@@ -3415,10 +3415,10 @@
 - 首次发现：2026-08-15
 - 最近更新：2026-08-15
 - 影响面：`Create Production Recovery Point`、生产 schema migration 前置恢复门禁，以及后续 production deploy。
-- 用户现象：`main` 与 `staging` 已同步且 release gate 成功，但生产恢复 workflow Run 1865 在创建备份前失败，日志为 `/opt/jyotisha-production/.state/mutation.lock: Permission denied`。首次修复后的 Run 1869 仍在 `pg_dump` 前失败，准确日志为 `sudo: a password is required`；迁移和部署因此持续阻断，生产运行版本未改变。
-- 根因：生产 bootstrap 遗留的 `.state` 路径或既有 lock 仍为非 `deploy` 所有。原恢复脚本只执行 `install -d -m 700`；对已经存在的目录该命令不会恢复所有权，随后由 `deploy` 打开共享锁即被内核拒绝。首次修复又错误假设主机已为 `deploy` 配置独立的免密 `chown`，但真实生产 sudo 边界只允许已审查的 Docker 命令，因此 `sudo -n chown` 立即失败。两次失败都发生在 `pg_dump` 前，没有生成可用恢复证明，也没有执行 schema migration。
-- 修复：恢复脚本先对 `.state`、`backups` 和既有 `mutation.lock` 做类型与非符号链接校验，取得当前运行 PostgreSQL 容器的不可变本地 image ID，再复用既有 `sudo -n docker` 边界启动一次性所有权修复容器：`--pull never`、无网络、只读根文件系统、`no-new-privileges`、删除全部 capability 后只保留 `CHOWN`，并且仅绑定 `.state` 与 `backups`。helper 只恢复两个目录 mount point 和既有普通 lock 到当前 `deploy` UID/GID；不递归改动历史备份、不删除或替换 lock inode，随后仍用同一个 `flock -n` fail-closed 获取共享 host lock。同步更新生产 runbook 和静态安全契约测试，明确不得再扩大主机 sudoers。
+- 用户现象：`main` 与 `staging` 已同步且 release gate 成功，但生产恢复 workflow Run 1865 在创建备份前失败，日志为 `/opt/jyotisha-production/.state/mutation.lock: Permission denied`。首次修复后的 Run 1869 仍在 `pg_dump` 前失败，准确日志为 `sudo: a password is required`。切换到受限 helper 后，Run 1873 又在同一备份前阶段报告 `chown: /opt/jyotisha-production/.state/mutation.lock: Permission denied`；迁移和部署因此持续阻断，生产运行版本未改变。
+- 根因：生产 bootstrap 遗留的 `.state` 路径或既有 lock 仍为非 `deploy` 所有。原恢复脚本只执行 `install -d -m 700`；对已经存在的目录该命令不会恢复所有权，随后由 `deploy` 打开共享锁即被内核拒绝。首次修复又错误假设主机已为 `deploy` 配置独立的免密 `chown`，但真实生产 sudo 边界只允许已审查的 Docker 命令，因此 `sudo -n chown` 立即失败。第二次修复的 helper 只保留 `CHOWN` capability，却按“父目录在前、lock 子文件在后”的顺序处理；当 `.state` 先被改成 `deploy:deploy 0700` 后，容器内 UID 0 已不再是目录 owner，且没有 `DAC_OVERRIDE`，所以无法继续遍历到 lock。三次失败都发生在 `pg_dump` 前，没有生成可用恢复证明，也没有执行 schema migration。
+- 修复：恢复脚本先对 `.state`、`backups` 和既有 `mutation.lock` 做类型与非符号链接校验，取得当前运行 PostgreSQL 容器的不可变本地 image ID，再复用既有 `sudo -n docker` 边界启动一次性所有权修复容器：`--pull never`、无网络、只读根文件系统、`no-new-privileges`、删除全部 capability 后只保留 `CHOWN`，并且仅绑定 `.state` 与 `backups`。helper 按 child-first 顺序先恢复既有普通 lock，再恢复两个目录 mount point 到当前 `deploy` UID/GID，因此不需要增加 `DAC_OVERRIDE`；不递归改动历史备份、不删除或替换 lock inode，随后仍用同一个 `flock -n` fail-closed 获取共享 host lock。同步更新生产 runbook 和静态安全契约测试，明确不得再扩大主机 sudoers，并锁定子文件必须先于 mode-`0700` 父目录处理。
 - 验证：本地 `bash -n`、YAML parse、聚焦 workflow contract 与 `git diff --check` 必须通过；远端必须重新完成 staging quality/deploy、exact-SHA release gate、真实 production dump + disposable restore + off-site artifact，再允许 migration/deploy。
 - 防复发：生产私有状态目录必须保持 `deploy:deploy 0700`，共享 lock 必须是普通非 symlink 文件且不可通过删除重建来“修复”；任何恢复流程失败都不得手填 `restore_verified=true` 或跳过恢复门禁。
-- 相关记录：生产迁移 runbook、Run 1865、Run 1869
+- 相关记录：生产迁移 runbook、Run 1865、Run 1869、Run 1873
 - 修复版本：待提交（精确 SHA 以重新发布后的远端分支与 production health 为准）

docs/operations/production-server-migration-2026-08.md

@@ -90,7 +90,7 @@ Use distinct production credentials for PostgreSQL roles, Better Auth, Resend, b
 
 ## Database migration engineering gate
 
-Before importing data, dispatch Gitea Actions → `Create Production Recovery Point` for the exact accepted release SHA. It validates the private production `.state` and `backups` paths, obtains the immutable image ID of the already-running PostgreSQL container, and reuses the existing passwordless Docker boundary to run that local image with `--pull never`, no network, a read-only root filesystem, `no-new-privileges`, and only `CHOWN` retained. With only `.state` and `backups` bind-mounted, it restores the two directory mount points and an existing regular shared lock to documented `deploy` ownership when bootstrap ownership has drifted. It then creates an AES-256-CBC/PBKDF2 encrypted custom-format dump on the production host, restores it into a disposable database, verifies non-sensitive table counts, removes only that disposable database, and uploads the encrypted dump plus verification manifest as a 30-day Gitea Actions artifact. It must preserve the existing lock inode, avoid recursively changing historical backups, and fail closed on symlinks, non-regular lock paths, or an unexpected image identity. The workflow prints the non-sensitive `recovery_reference` and exact UTC `recovery_created_at`; use those values only after the artifact upload succeeds.
+Before importing data, dispatch Gitea Actions → `Create Production Recovery Point` for the exact accepted release SHA. It validates the private production `.state` and `backups` paths, obtains the immutable image ID of the already-running PostgreSQL container, and reuses the existing passwordless Docker boundary to run that local image with `--pull never`, no network, a read-only root filesystem, `no-new-privileges`, and only `CHOWN` retained. With only `.state` and `backups` bind-mounted, it restores an existing regular shared lock before changing its mode-`0700` parent directory, then restores the two directory mount points to documented `deploy` ownership when bootstrap ownership has drifted. The child-first order avoids adding `DAC_OVERRIDE` while retaining access to the lock path. It then creates an AES-256-CBC/PBKDF2 encrypted custom-format dump on the production host, restores it into a disposable database, verifies non-sensitive table counts, removes only that disposable database, and uploads the encrypted dump plus verification manifest as a 30-day Gitea Actions artifact. It must preserve the existing lock inode, avoid recursively changing historical backups, and fail closed on symlinks, non-regular lock paths, or an unexpected image identity. The workflow prints the non-sensitive `recovery_reference` and exact UTC `recovery_created_at`; use those values only after the artifact upload succeeds.
 
 Then dispatch Gitea Actions → `Migrate Production Database` for the same exact accepted release SHA. The workflow requires `main == staging == deploy_sha`, the same successful staging backend gate, the same manual release gate, and the public staging `/api/health` identity for that SHA. It also requires the recovery workflow's non-sensitive reference, exact UTC creation time, and `restore_verified=true`; the recovery point must be no more than 24 hours old and must already have passed the restore verification. It verifies the current production revision, obtains the gate-attested immutable Web image, runs the schema checker, applies only pending application schema migrations, and requires the checker to converge afterward.
 

frontend/tests/staging-backend-workflows.test.ts

@@ -1091,6 +1091,7 @@ test("production recovery workflow creates a verified encrypted off-site artifac
   assert.match(runner, /sudo -n docker run --rm --pull never --network none --read-only --user 0:0/);
   assert.match(runner, /--cap-drop ALL --cap-add CHOWN --security-opt no-new-privileges/);
   assert.match(runner, /--entrypoint chown "\$ownership_image"/);
+  assert.match(runner, /ownership_targets=\(\)[\s\S]*ownership_targets\+=\("\$lock_file"\)[\s\S]*ownership_targets\+=\("\$state_directory" "\$backup_directory"\)/);
   assert.match(runner, /"\$deployment_uid:\$deployment_gid" "\$\{ownership_targets\[@\]\}"/);
   assert.match(runner, /production mutation lock is unsafe/);
   assert.doesNotMatch(runner, /rm -f[^\n]*mutation\.lock|chown -R/);