feat(sense): add reconciliation safety controls [T-012]

2026-08-07 23:00:03 +08:00
parent 677ed732f7
commit 12857fdf32
35 changed files with 2817 additions and 124 deletions
@@ -1,6 +1,6 @@
 # Sense M1/M2 接入骨架

-本目录是 YoVision Sense 的 M1/M2 接入骨架。数据库保存期望态，ONVIF 和 MediaMTX 通过端口隔离；M1 默认使用 SQLite，T-009～T-011 增加 PostgreSQL 双 schema、Area 准入、本地审计 Outbox 和 Control API v1。默认关闭真实 ONVIF 与公共业务路由；T-006 的真实样机结论仅覆盖已批准的精确海康基线，不能据此宣称多品牌兼容。
+本目录是 YoVision Sense 的 M1/M2 接入骨架。数据库保存期望态，ONVIF 和 MediaMTX 通过端口隔离；M1 默认使用 SQLite，T-009～T-012 增加 PostgreSQL 双 schema、Area 准入、本地审计 Outbox、Control API v1、多实例调和 fencing 和孤儿受控处置。默认关闭真实 ONVIF 与公共业务路由；T-006 的真实样机结论仅覆盖已批准的精确海康基线，不能据此宣称多品牌兼容。

 ## 常用命令

@@ -14,7 +14,7 @@ go build -o bin/sense-api.exe ./cmd/sense-api
 go run ./cmd/sense-api
 ```

-Unix 将构建产物改为 `bin/sense-api`。服务默认监听 `127.0.0.1:8080`，SQLite 默认写入 `Sense/data/sense.db`，MediaMTX 控制 API 默认是 `http://127.0.0.1:9997`。默认运行只暴露 `/healthz` 与 `/readyz`；只有显式选择 PostgreSQL 并完成安全配置后才注册 7 个 `/api/v1` Control API 路由。
+Unix 将构建产物改为 `bin/sense-api`。服务默认监听 `127.0.0.1:8080`，SQLite 默认写入 `Sense/data/sense.db`，MediaMTX 控制 API 默认是 `http://127.0.0.1:9997`。默认运行暴露 `/healthz`、`/readyz` 和不含租户/设备标签的 `/metrics`；只有显式选择 PostgreSQL 并完成安全配置后才注册 7 个 `/api/v1` Control API 路由。

 常用环境变量：

@@ -26,7 +26,13 @@ Unix 将构建产物改为 `bin/sense-api`。服务默认监听 `127.0.0.1:8080`
 | `SENSE_DB_DSN` | `file:data/sense.db` | 所选 driver 的私有 DSN；不得写入日志、文档或仓库，PostgreSQL 密码优先由 `PGPASSFILE`/环境密钥提供 |
 | `SENSE_MEDIAMTX_URL` | `http://127.0.0.1:9997` | MediaMTX 控制 API；不得包含 userinfo |
 | `SENSE_RECONCILE_INTERVAL` | `5s` | 对账周期 |
+| `SENSE_RECONCILE_LEASE_DURATION` | `30s` | PostgreSQL due-row 租期；最大 5 分钟 |
+| `SENSE_RECONCILE_OPERATION_TIMEOUT` | `20s` | 单项 ONVIF/MediaMTX deadline；必须严格短于租期 |
 | `SENSE_PROBE_INTERVAL` | `10s` | path 探活周期 |
+| `SENSE_INSTANCE_ID` | 随进程随机生成 | 最多 64 位低基数字符串；多实例部署建议显式注入唯一实例 ID |
+| `SENSE_METRICS_ENABLED` | `true` | 是否注册低基数 Prometheus 文本 `/metrics` |
+| `SENSE_ORPHAN_SCAN_ENABLED` | PostgreSQL 为 `true`，SQLite 为 `false` | 周期执行只读 MediaMTX Path 差异报告；从不自动删除 |
+| `SENSE_ORPHAN_SCAN_INTERVAL` | `1m` | 孤儿只读扫描周期，最短 10 秒 |
 | `SENSE_ONVIF_MODE` | `disabled` | `disabled` 或 `standard`；默认不访问真实摄像头 |
 | `SENSE_ONVIF_RTSP_REWRITE_HOST` | 空 | NAT 或故障代理场景下重写 ONVIF 返回的 RTSP 主机 |
 | `SENSE_ONVIF_RTSP_REWRITE_PORT` | `0` | 非零时重写 ONVIF 返回的 RTSP 端口 |
@@ -50,9 +56,9 @@ SENSE_CREDENTIAL_<KEY>_RTSP_PASSWORD

 MediaMTX `v1.19.3` 应作为独立二进制启动并只在可信网络开放 API。获取与 SHA-256 校验值见 `docs/03-tech-stack.md`。生成客户端使用固定版本工具和 vendored 官方 OpenAPI；`internal/mtx/generated/client.gen.go` 不可手改。

-## T-009～T-011 PostgreSQL 17.10
+## T-009～T-012 PostgreSQL 17.10

-初始化与增量 SQL 位于 `deploy/postgres/`，由高权限部署步骤按文件名前缀执行；Sense 进程不会自动创建角色、schema 或 Bell 对象。`bell_app` 拥有 Site/Area、配额、`capture_policy` 及两个版本化视图，`sense_app` 只能读取两个视图，不能读取或写入 Bell 源表。T-011 的 v4 schema 另有资源版本、24 小时幂等收据和 batch operation 表；收据只保存 scope/body 摘要与脱敏响应快照。应用登录角色和密码由部署环境创建，不进入仓库。
+初始化与增量 SQL 位于 `deploy/postgres/`，由高权限部署步骤按文件名前缀执行；Sense 进程不会自动创建角色、schema 或 Bell 对象。`bell_app` 拥有 Site/Area、配额、`capture_policy` 及两个版本化视图，`sense_app` 只能读取两个视图，不能读取或写入 Bell 源表。T-011 的 v4 schema 增加资源版本、24 小时幂等收据和 batch operation；T-012 的 v5 schema 增加数据库时钟租约、Path 历史归属及脱敏孤儿报告/处置结果。表中不保存 MediaMTX source URI。应用登录角色和密码由部署环境创建，不进入仓库。

 PostgreSQL 新建设备必须携带匹配 tenant/Site 的 `area_id`。具有 `video_capture` 能力的设备在创建、移动 Area 和从 disabled 切到 enabled 时执行 Area 准入；`non_imaging_only` 拒绝成像设备但允许非成像设备。投影缺失、非法或版本回退只拒绝新变更，不关闭已有流。创建、配置修改和期望态受理都与对应脱敏 Outbox 事实同事务；停用后调和器只删除该设备的精确 MediaMTX path 并收敛为 offline，不枚举未知 path。Bell relay 尚未实现。

@@ -72,7 +78,18 @@ $env:SENSE_DB_DSN = '由部署环境私下设置'
 go run ./cmd/sense-api
 ```

-PostgreSQL 启动会检查 Sense migration 版本及当前角色对两个 Bell 投影视图的最小权限；权限过宽、视图不可读或 schema 未安装时 readiness 初始化失败。默认 SQLite 路径和 `cmd/sense-lab` 保持不变，但 SQLite 不实现生产 Area/Outbox 一致性，Control API feature flag 在 SQLite 下会直接拒绝启动。
+PostgreSQL 启动会检查 Sense v5 migration、当前角色对两个 Bell 投影视图和本地控制/对账表的最小权限；权限过宽、视图不可读或 schema 未安装时初始化失败。默认 SQLite 路径和 `cmd/sense-lab` 保持不变，但 SQLite 不实现生产 Area/Outbox、多实例租约或孤儿处置语义，Control API 与孤儿扫描在 SQLite 下不会启动。
+
+### 孤儿报告与受控处置
+
+`sense-api` 的 PostgreSQL 模式默认每分钟只读枚举一次 MediaMTX 配置 Path，并持久化汇总。`unowned` 表示没有 Sense 历史归属证据，永远不会由 Sense 删除；只有 `owned_stale` 可以进入处置候选。需要人工处置时从 `Sense/` 执行：
+
+```powershell
+go run ./cmd/sense-orphan -mode report
+go run ./cmd/sense-orphan -mode apply -scan-id scan_... -actor operator-id -confirm "DELETE scan_..."
+```
+
+`report` 输出不含 source URI 的 JSON 摘要。`apply` 会再次读取数据库和 MediaMTX，只处理仍属于原快照的 `owned_stale`，并要求快照不超过 15 分钟、候选不超过 128 且候选占当前全部配置 Path 不超过 10%；任何门禁失败均为零删除，没有 `force`。完整流程与恢复方法见 [`../docs/runbooks/sense-reconciliation.md`](../docs/runbooks/sense-reconciliation.md)。

 ### 开启 Control API

@@ -15,8 +15,10 @@ import (
 	"yovision/sense/internal/auth"
 	"yovision/sense/internal/config"
 	"yovision/sense/internal/controlapi"
+	"yovision/sense/internal/metrics"
 	"yovision/sense/internal/mtx"
 	"yovision/sense/internal/onvif"
+	"yovision/sense/internal/orphan"
 	"yovision/sense/internal/probe"
 	"yovision/sense/internal/reconcile"
 	"yovision/sense/internal/store"
@@ -37,6 +39,14 @@ func run(logger *slog.Logger) error {
 	if err != nil {
 		return fmt.Errorf("load configuration: %w", err)
 	}
+	instanceID := cfg.InstanceID
+	if instanceID == "" {
+		instanceID, err = metrics.GenerateInstanceID()
+		if err != nil {
+			return err
+		}
+	}
+	registry := metrics.New(instanceID, version)
 	ctx, stop := signal.NotifyContext(context.Background(), os.Interrupt, syscall.SIGTERM)
 	defer stop()

@@ -76,22 +86,36 @@ func run(logger *slog.Logger) error {
 		})
 	}
 	discovery := onvif.NewRouter(cameraAdapter, credentials)
-	reconciler := reconcile.New(repository, discovery, mediaClient)
+	reconciler := reconcile.NewWithOptions(repository, discovery, mediaClient, reconcile.Options{
+		InstanceID: instanceID, LeaseDuration: cfg.ReconcileLeaseDuration,
+		OperationTimeout: cfg.ReconcileOperationTimeout, Metrics: registry,
+	})
 	checker := probe.New(repository, mediaClient)
+	var orphanScanner *orphan.Manager
+	if cfg.OrphanScanEnabled {
+		orphanStore, ok := repository.(store.OrphanRepository)
+		if !ok {
+			return errors.New("selected repository does not support orphan scanning")
+		}
+		orphanScanner = orphan.New(orphanStore, mediaClient, instanceID, registry)
+	}
 	report := func(err error) {
 		// Domain and MediaMTX errors intentionally omit stream URIs and credentials.
 		logger.Warn("background convergence error", "error", err)
 	}
 	var background sync.WaitGroup
-	background.Add(2)
-	go func() {
-		defer background.Done()
-		reconciler.Run(ctx, cfg.ReconcileInterval, report)
-	}()
-	go func() {
-		defer background.Done()
-		checker.Run(ctx, cfg.ProbeInterval, report)
-	}()
+	startBackground := func(run func()) {
+		background.Add(1)
+		go func() {
+			defer background.Done()
+			run()
+		}()
+	}
+	startBackground(func() { reconciler.Run(ctx, cfg.ReconcileInterval, report) })
+	startBackground(func() { checker.Run(ctx, cfg.ProbeInterval, report) })
+	if orphanScanner != nil {
+		startBackground(func() { orphanScanner.Run(ctx, cfg.OrphanScanInterval, report) })
+	}

 	mux := http.NewServeMux()
 	mux.HandleFunc("GET /healthz", func(writer http.ResponseWriter, _ *http.Request) {
@@ -104,6 +128,9 @@ func run(logger *slog.Logger) error {
 		writer.WriteHeader(http.StatusOK)
 		_, _ = writer.Write([]byte(`{"status":"ready"}`))
 	})
+	if cfg.MetricsEnabled {
+		mux.Handle("GET /metrics", registry.Handler())
+	}
 	if cfg.ControlAPIEnabled {
 		mux.Handle("/api/v1/", controlHandler)
 	}
@@ -118,6 +145,7 @@ func run(logger *slog.Logger) error {
 	serverErrors := make(chan error, 1)
 	go func() {
 		logger.Info("Sense listening", "address", cfg.HTTPAddress, "version", version,
+			"instance_id", instanceID,
 			"control_api_enabled", cfg.ControlAPIEnabled)
 		serverErrors <- server.ListenAndServe()
 	}()
@@ -0,0 +1,94 @@
+// sense-orphan is a local, PostgreSQL-backed operational command. Report is
+// read-only; apply requires a fresh scan ID, actor and exact confirmation.
+package main
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"flag"
+	"fmt"
+	"net/http"
+	"os"
+	"os/signal"
+	"strings"
+	"syscall"
+
+	"yovision/sense/internal/config"
+	"yovision/sense/internal/metrics"
+	"yovision/sense/internal/mtx"
+	"yovision/sense/internal/orphan"
+	"yovision/sense/internal/store"
+)
+
+func main() {
+	if err := run(); err != nil {
+		_, _ = fmt.Fprintln(os.Stderr, "sense-orphan failed:", err)
+		os.Exit(1)
+	}
+}
+
+func run() error {
+	mode := flag.String("mode", "report", "report or apply")
+	scanID := flag.String("scan-id", "", "fresh scan ID required by apply")
+	actor := flag.String("actor", "", "operator identifier required by apply")
+	confirm := flag.String("confirm", "", "exact confirmation: DELETE <scan-id>")
+	flag.Parse()
+
+	cfg, err := config.Load()
+	if err != nil {
+		return err
+	}
+	if !strings.EqualFold(strings.TrimSpace(cfg.DatabaseDriver), store.DriverPostgres) {
+		return errors.New("sense-orphan requires SENSE_DB_DRIVER=postgres")
+	}
+	ctx, stop := signal.NotifyContext(context.Background(), os.Interrupt, syscall.SIGTERM)
+	defer stop()
+	repository, err := store.OpenPostgres(ctx, cfg.DatabaseDSN)
+	if err != nil {
+		return err
+	}
+	defer repository.Close()
+	media, err := mtx.NewClient(cfg.MediaMTXURL, &http.Client{Timeout: cfg.ReconcileOperationTimeout})
+	if err != nil {
+		return err
+	}
+	instanceID := cfg.InstanceID
+	if instanceID == "" {
+		instanceID, err = metrics.GenerateInstanceID()
+		if err != nil {
+			return err
+		}
+	}
+	manager := orphan.New(repository, media, instanceID, metrics.New(instanceID, "operator"))
+	encoder := json.NewEncoder(os.Stdout)
+	switch *mode {
+	case "report":
+		scan, err := manager.Report(ctx)
+		if err != nil {
+			return err
+		}
+		return encoder.Encode(struct {
+			ScanID        string `json:"scan_id"`
+			Observed      int    `json:"observed"`
+			OwnedStale    int    `json:"owned_stale"`
+			Unowned       int    `json:"unowned"`
+			SafetyAllowed bool   `json:"safety_allowed"`
+			SafetyReason  string `json:"safety_reason"`
+			ExpiresAt     string `json:"expires_at"`
+		}{
+			ScanID: scan.ID, Observed: scan.ObservedCount,
+			OwnedStale: scan.OwnedStaleCount, Unowned: scan.UnownedCount,
+			SafetyAllowed: scan.SafetyAllowed, SafetyReason: scan.SafetyReason,
+			ExpiresAt: scan.ExpiresAt.UTC().Format("2006-01-02T15:04:05Z07:00"),
+		})
+	case "apply":
+		result, err := manager.Apply(ctx, *scanID, *actor, *confirm)
+		if encodeErr := encoder.Encode(result); encodeErr != nil {
+			return encodeErr
+		}
+		return err
+	default:
+		return errors.New("mode must be report or apply")
+	}
+}
@@ -7,39 +7,52 @@ import (
 	"net/url"
 	"os"
 	"path/filepath"
+	"regexp"
 	"strconv"
 	"strings"
 	"time"
 )

 const (
-	defaultHTTPAddress     = "127.0.0.1:8080"
-	defaultDatabaseDriver  = "sqlite"
-	defaultDatabaseDSN     = "file:data/sense.db"
-	defaultMediaMTXURL     = "http://127.0.0.1:9997"
-	defaultReconcilePeriod = 5 * time.Second
-	defaultProbePeriod     = 10 * time.Second
-	defaultONVIFMode       = "disabled"
-	defaultControlAuthMode = "static-sha256"
+	defaultHTTPAddress      = "127.0.0.1:8080"
+	defaultDatabaseDriver   = "sqlite"
+	postgresDatabaseDriver  = "postgres"
+	defaultDatabaseDSN      = "file:data/sense.db"
+	defaultMediaMTXURL      = "http://127.0.0.1:9997"
+	defaultReconcilePeriod  = 5 * time.Second
+	defaultReconcileLease   = 30 * time.Second
+	defaultOperationTimeout = 20 * time.Second
+	defaultProbePeriod      = 10 * time.Second
+	defaultOrphanScanPeriod = time.Minute
+	defaultONVIFMode        = "disabled"
+	defaultControlAuthMode  = "static-sha256"
 )

+var instanceIDPattern = regexp.MustCompile(`^[A-Za-z0-9][A-Za-z0-9._-]{0,63}$`)
+
 type Config struct {
-	HTTPAddress              string
-	AllowNonLoopback         bool
-	DatabaseDriver           string
-	DatabaseDSN              string
-	MediaMTXURL              string
-	ReconcileInterval        time.Duration
-	ProbeInterval            time.Duration
-	ONVIFMode                string
-	RTSPRewriteHost          string
-	RTSPRewritePort          int
-	RTSPStripQuery           bool
-	ControlAPIEnabled        bool
-	ControlAuthMode          string
-	ControlAuthFile          string
-	ControlCursorKeyFile     string
-	ControlAllowInsecureHTTP bool
+	HTTPAddress               string
+	AllowNonLoopback          bool
+	DatabaseDriver            string
+	DatabaseDSN               string
+	MediaMTXURL               string
+	ReconcileInterval         time.Duration
+	ReconcileLeaseDuration    time.Duration
+	ReconcileOperationTimeout time.Duration
+	ProbeInterval             time.Duration
+	InstanceID                string
+	MetricsEnabled            bool
+	OrphanScanEnabled         bool
+	OrphanScanInterval        time.Duration
+	ONVIFMode                 string
+	RTSPRewriteHost           string
+	RTSPRewritePort           int
+	RTSPStripQuery            bool
+	ControlAPIEnabled         bool
+	ControlAuthMode           string
+	ControlAuthFile           string
+	ControlCursorKeyFile      string
+	ControlAllowInsecureHTTP  bool
 }

 func Load() (Config, error) {
@@ -51,6 +64,14 @@ func Load() (Config, error) {
 	if err != nil {
 		return Config{}, err
 	}
+	reconcileLease, err := durationEnv("SENSE_RECONCILE_LEASE_DURATION", defaultReconcileLease)
+	if err != nil {
+		return Config{}, err
+	}
+	operationTimeout, err := durationEnv("SENSE_RECONCILE_OPERATION_TIMEOUT", defaultOperationTimeout)
+	if err != nil {
+		return Config{}, err
+	}
 	probePeriod, err := durationEnv("SENSE_PROBE_INTERVAL", defaultProbePeriod)
 	if err != nil {
 		return Config{}, err
@@ -71,24 +92,44 @@ func Load() (Config, error) {
 	if err != nil {
 		return Config{}, err
 	}
+	metricsEnabled, err := boolEnv("SENSE_METRICS_ENABLED", true)
+	if err != nil {
+		return Config{}, err
+	}
+	databaseDriver := stringEnv("SENSE_DB_DRIVER", defaultDatabaseDriver)
+	orphanDefault := strings.EqualFold(strings.TrimSpace(databaseDriver), postgresDatabaseDriver)
+	orphanEnabled, err := boolEnv("SENSE_ORPHAN_SCAN_ENABLED", orphanDefault)
+	if err != nil {
+		return Config{}, err
+	}
+	orphanPeriod, err := durationEnv("SENSE_ORPHAN_SCAN_INTERVAL", defaultOrphanScanPeriod)
+	if err != nil {
+		return Config{}, err
+	}

 	cfg := Config{
-		HTTPAddress:              stringEnv("SENSE_HTTP_ADDR", defaultHTTPAddress),
-		AllowNonLoopback:         allow,
-		DatabaseDriver:           stringEnv("SENSE_DB_DRIVER", defaultDatabaseDriver),
-		DatabaseDSN:              stringEnv("SENSE_DB_DSN", defaultDatabaseDSN),
-		MediaMTXURL:              stringEnv("SENSE_MEDIAMTX_URL", defaultMediaMTXURL),
-		ReconcileInterval:        reconcilePeriod,
-		ProbeInterval:            probePeriod,
-		ONVIFMode:                stringEnv("SENSE_ONVIF_MODE", defaultONVIFMode),
-		RTSPRewriteHost:          stringEnv("SENSE_ONVIF_RTSP_REWRITE_HOST", ""),
-		RTSPRewritePort:          rewritePort,
-		RTSPStripQuery:           stripQuery,
-		ControlAPIEnabled:        controlEnabled,
-		ControlAuthMode:          stringEnv("SENSE_CONTROL_AUTH_MODE", defaultControlAuthMode),
-		ControlAuthFile:          stringEnv("SENSE_CONTROL_AUTH_FILE", ""),
-		ControlCursorKeyFile:     stringEnv("SENSE_CONTROL_CURSOR_KEY_FILE", ""),
-		ControlAllowInsecureHTTP: controlAllowInsecure,
+		HTTPAddress:               stringEnv("SENSE_HTTP_ADDR", defaultHTTPAddress),
+		AllowNonLoopback:          allow,
+		DatabaseDriver:            databaseDriver,
+		DatabaseDSN:               stringEnv("SENSE_DB_DSN", defaultDatabaseDSN),
+		MediaMTXURL:               stringEnv("SENSE_MEDIAMTX_URL", defaultMediaMTXURL),
+		ReconcileInterval:         reconcilePeriod,
+		ReconcileLeaseDuration:    reconcileLease,
+		ReconcileOperationTimeout: operationTimeout,
+		ProbeInterval:             probePeriod,
+		InstanceID:                stringEnv("SENSE_INSTANCE_ID", ""),
+		MetricsEnabled:            metricsEnabled,
+		OrphanScanEnabled:         orphanEnabled,
+		OrphanScanInterval:        orphanPeriod,
+		ONVIFMode:                 stringEnv("SENSE_ONVIF_MODE", defaultONVIFMode),
+		RTSPRewriteHost:           stringEnv("SENSE_ONVIF_RTSP_REWRITE_HOST", ""),
+		RTSPRewritePort:           rewritePort,
+		RTSPStripQuery:            stripQuery,
+		ControlAPIEnabled:         controlEnabled,
+		ControlAuthMode:           stringEnv("SENSE_CONTROL_AUTH_MODE", defaultControlAuthMode),
+		ControlAuthFile:           stringEnv("SENSE_CONTROL_AUTH_FILE", ""),
+		ControlCursorKeyFile:      stringEnv("SENSE_CONTROL_CURSOR_KEY_FILE", ""),
+		ControlAllowInsecureHTTP:  controlAllowInsecure,
 	}
 	if err := cfg.Validate(); err != nil {
 		return Config{}, err
@@ -129,6 +170,35 @@ func (c Config) Validate() error {
 	if c.ReconcileInterval <= 0 || c.ProbeInterval <= 0 {
 		return fmt.Errorf("loop intervals must be positive")
 	}
+	leaseDuration := c.ReconcileLeaseDuration
+	if leaseDuration == 0 {
+		leaseDuration = defaultReconcileLease
+	}
+	operationTimeout := c.ReconcileOperationTimeout
+	if operationTimeout == 0 {
+		operationTimeout = defaultOperationTimeout
+	}
+	if leaseDuration <= 0 || leaseDuration > 5*time.Minute {
+		return fmt.Errorf("SENSE_RECONCILE_LEASE_DURATION must be positive and at most 5m")
+	}
+	if operationTimeout <= 0 || operationTimeout >= leaseDuration {
+		return fmt.Errorf("SENSE_RECONCILE_OPERATION_TIMEOUT must be positive and shorter than the lease")
+	}
+	if c.InstanceID != "" && !instanceIDPattern.MatchString(c.InstanceID) {
+		return fmt.Errorf("invalid SENSE_INSTANCE_ID")
+	}
+	if c.OrphanScanEnabled {
+		if databaseDriver != postgresDatabaseDriver {
+			return fmt.Errorf("orphan scanning requires SENSE_DB_DRIVER=postgres")
+		}
+		orphanInterval := c.OrphanScanInterval
+		if orphanInterval == 0 {
+			orphanInterval = defaultOrphanScanPeriod
+		}
+		if orphanInterval < 10*time.Second {
+			return fmt.Errorf("SENSE_ORPHAN_SCAN_INTERVAL must be at least 10s")
+		}
+	}
 	if c.ONVIFMode != "" && c.ONVIFMode != "disabled" && c.ONVIFMode != "standard" {
 		return fmt.Errorf("SENSE_ONVIF_MODE must be disabled or standard")
 	}
@@ -3,6 +3,7 @@ package config
 import (
 	"path/filepath"
 	"testing"
+	"time"
 )

 func TestValidateRejectsNonLoopbackByDefault(t *testing.T) {
@@ -23,6 +24,35 @@ func TestValidateRejectsNonLoopbackByDefault(t *testing.T) {
 	}
 }

+func TestValidateReconcileLeaseAndOrphanProductionBoundary(t *testing.T) {
+	base := Config{
+		HTTPAddress: "127.0.0.1:8080", DatabaseDriver: "postgres",
+		DatabaseDSN: "postgres://sense-runtime@127.0.0.1/yovision?sslmode=disable",
+		MediaMTXURL: "http://127.0.0.1:9997", ReconcileInterval: time.Second,
+		ProbeInterval: time.Second, ReconcileLeaseDuration: 30 * time.Second,
+		ReconcileOperationTimeout: 20 * time.Second, OrphanScanEnabled: true,
+		OrphanScanInterval: time.Minute, InstanceID: "ins_edge-01",
+	}
+	if err := base.Validate(); err != nil {
+		t.Fatalf("valid multi-instance configuration failed: %v", err)
+	}
+	invalidTimeout := base
+	invalidTimeout.ReconcileOperationTimeout = invalidTimeout.ReconcileLeaseDuration
+	if err := invalidTimeout.Validate(); err == nil {
+		t.Fatal("operation timeout equal to the lease was accepted")
+	}
+	invalidInstance := base
+	invalidInstance.InstanceID = "tenant/site"
+	if err := invalidInstance.Validate(); err == nil {
+		t.Fatal("unbounded instance label was accepted")
+	}
+	sqlite := base
+	sqlite.DatabaseDriver, sqlite.DatabaseDSN = "sqlite", "file:test.db"
+	if err := sqlite.Validate(); err == nil {
+		t.Fatal("orphan scanner was accepted on SQLite")
+	}
+}
+
 func TestValidateRejectsCredentialsInMediaMTXURL(t *testing.T) {
 	t.Parallel()
 	cfg := Config{
@@ -0,0 +1,189 @@
+// Package metrics exposes a deliberately small, low-cardinality Prometheus
+// surface without adding a runtime dependency. Tenant, Site, device and Path
+// identifiers never become labels.
+package metrics
+
+import (
+	"crypto/rand"
+	"encoding/hex"
+	"errors"
+	"fmt"
+	"net/http"
+	"strconv"
+	"strings"
+	"sync/atomic"
+	"time"
+)
+
+func GenerateInstanceID() (string, error) {
+	value := make([]byte, 12)
+	if _, err := rand.Read(value); err != nil {
+		return "", errors.New("generate Sense instance identifier")
+	}
+	return "ins_" + hex.EncodeToString(value), nil
+}
+
+type Registry struct {
+	instanceID string
+	version    string
+
+	reconcileRunOK       atomic.Uint64
+	reconcileRunError    atomic.Uint64
+	reconcileItemOK      atomic.Uint64
+	reconcileItemError   atomic.Uint64
+	reconcileItemLost    atomic.Uint64
+	reconcileDurationNS  atomic.Uint64
+	reconcileTotal       atomic.Int64
+	reconcileUnconverged atomic.Int64
+
+	orphanScanOK       atomic.Uint64
+	orphanScanError    atomic.Uint64
+	orphanObserved     atomic.Int64
+	orphanOwnedStale   atomic.Int64
+	orphanUnowned      atomic.Int64
+	orphanBlockedRatio atomic.Uint64
+	orphanBlockedAge   atomic.Uint64
+	orphanBlockedScope atomic.Uint64
+	orphanDeleted      atomic.Uint64
+	orphanDeleteFailed atomic.Uint64
+}
+
+func New(instanceID, version string) *Registry {
+	return &Registry{instanceID: instanceID, version: version}
+}
+
+func (r *Registry) ObserveReconcileRun(err error, duration time.Duration) {
+	if err == nil {
+		r.reconcileRunOK.Add(1)
+	} else {
+		r.reconcileRunError.Add(1)
+	}
+	if duration < 0 {
+		duration = 0
+	}
+	r.reconcileDurationNS.Store(uint64(duration))
+}
+
+func (r *Registry) ObserveReconcileItem(result string) {
+	switch result {
+	case "success":
+		r.reconcileItemOK.Add(1)
+	case "lease_lost":
+		r.reconcileItemLost.Add(1)
+	default:
+		r.reconcileItemError.Add(1)
+	}
+}
+
+func (r *Registry) SetConvergence(total, unconverged int) {
+	r.reconcileTotal.Store(int64(total))
+	r.reconcileUnconverged.Store(int64(unconverged))
+}
+
+func (r *Registry) ObserveOrphanScan(observed, ownedStale, unowned int, err error) {
+	if err == nil {
+		r.orphanScanOK.Add(1)
+		r.orphanObserved.Store(int64(observed))
+		r.orphanOwnedStale.Store(int64(ownedStale))
+		r.orphanUnowned.Store(int64(unowned))
+	} else {
+		r.orphanScanError.Add(1)
+	}
+}
+
+func (r *Registry) ObserveOrphanCleanupBlocked(reason string) {
+	switch reason {
+	case "snapshot_expired":
+		r.orphanBlockedAge.Add(1)
+	case "ratio_exceeded":
+		r.orphanBlockedRatio.Add(1)
+	default:
+		r.orphanBlockedScope.Add(1)
+	}
+}
+
+func (r *Registry) ObserveOrphanCleanup(deleted, failed int) {
+	if deleted > 0 {
+		r.orphanDeleted.Add(uint64(deleted))
+	}
+	if failed > 0 {
+		r.orphanDeleteFailed.Add(uint64(failed))
+	}
+}
+
+func (r *Registry) Handler() http.Handler {
+	return http.HandlerFunc(func(writer http.ResponseWriter, _ *http.Request) {
+		writer.Header().Set("Content-Type", "text/plain; version=0.0.4; charset=utf-8")
+		writer.Header().Set("Cache-Control", "no-store")
+		_, _ = writer.Write([]byte(r.render()))
+	})
+}
+
+func (r *Registry) render() string {
+	var output strings.Builder
+	writeMetric(&output, "sense_build_info", "Sense process build and instance information.",
+		fmt.Sprintf("{instance_id=%s,version=%s}", quoteLabel(r.instanceID), quoteLabel(r.version)), 1)
+	writeMetric(&output, "sense_reconcile_runs_total", "Reconciliation runs by result.",
+		`{result="success"}`, r.reconcileRunOK.Load())
+	writeSample(&output, "sense_reconcile_runs_total", `{result="error"}`, r.reconcileRunError.Load())
+	writeMetric(&output, "sense_reconcile_items_total", "Reconciliation items by fixed result.",
+		`{result="success"}`, r.reconcileItemOK.Load())
+	writeSample(&output, "sense_reconcile_items_total", `{result="error"}`, r.reconcileItemError.Load())
+	writeSample(&output, "sense_reconcile_items_total", `{result="lease_lost"}`, r.reconcileItemLost.Load())
+	writeMetric(&output, "sense_reconcile_last_duration_seconds", "Duration of the last reconciliation run.",
+		"", float64(r.reconcileDurationNS.Load())/float64(time.Second))
+	writeMetric(&output, "sense_reconcile_devices", "Enabled video devices by convergence state.",
+		`{state="total"}`, r.reconcileTotal.Load())
+	writeSample(&output, "sense_reconcile_devices", `{state="unconverged"}`, r.reconcileUnconverged.Load())
+	writeMetric(&output, "sense_orphan_scan_runs_total", "MediaMTX orphan scans by result.",
+		`{result="success"}`, r.orphanScanOK.Load())
+	writeSample(&output, "sense_orphan_scan_runs_total", `{result="error"}`, r.orphanScanError.Load())
+	writeMetric(&output, "sense_orphan_paths", "Path counts from the last successful orphan scan.",
+		`{classification="observed"}`, r.orphanObserved.Load())
+	writeSample(&output, "sense_orphan_paths", `{classification="owned_stale"}`, r.orphanOwnedStale.Load())
+	writeSample(&output, "sense_orphan_paths", `{classification="unowned"}`, r.orphanUnowned.Load())
+	writeMetric(&output, "sense_orphan_cleanup_blocked_total", "Orphan cleanups blocked by a fixed safety reason.",
+		`{reason="ratio_exceeded"}`, r.orphanBlockedRatio.Load())
+	writeSample(&output, "sense_orphan_cleanup_blocked_total", `{reason="snapshot_expired"}`, r.orphanBlockedAge.Load())
+	writeSample(&output, "sense_orphan_cleanup_blocked_total", `{reason="scope_changed"}`, r.orphanBlockedScope.Load())
+	writeMetric(&output, "sense_orphan_cleanup_items_total", "Owned stale path cleanup results.",
+		`{result="deleted"}`, r.orphanDeleted.Load())
+	writeSample(&output, "sense_orphan_cleanup_items_total", `{result="failed"}`, r.orphanDeleteFailed.Load())
+	return output.String()
+}
+
+func writeMetric(builder *strings.Builder, name, help, labels string, value any) {
+	metricType := "gauge"
+	if strings.HasSuffix(name, "_total") {
+		metricType = "counter"
+	}
+	builder.WriteString("# HELP ")
+	builder.WriteString(name)
+	builder.WriteByte(' ')
+	builder.WriteString(help)
+	builder.WriteByte('\n')
+	builder.WriteString("# TYPE ")
+	builder.WriteString(name)
+	builder.WriteByte(' ')
+	builder.WriteString(metricType)
+	builder.WriteByte('\n')
+	writeSample(builder, name, labels, value)
+}
+
+func writeSample(builder *strings.Builder, name, labels string, value any) {
+	builder.WriteString(name)
+	builder.WriteString(labels)
+	builder.WriteByte(' ')
+	switch typed := value.(type) {
+	case float64:
+		builder.WriteString(strconv.FormatFloat(typed, 'f', 6, 64))
+	default:
+		builder.WriteString(fmt.Sprint(typed))
+	}
+	builder.WriteByte('\n')
+}
+
+func quoteLabel(value string) string {
+	replacer := strings.NewReplacer(`\`, `\\`, "\n", `\n`, `"`, `\"`)
+	return `"` + replacer.Replace(value) + `"`
+}
@@ -0,0 +1,84 @@
+package metrics
+
+import (
+	"net/http/httptest"
+	"strings"
+	"sync"
+	"testing"
+	"time"
+)
+
+func TestHandlerUsesOnlyFixedLowCardinalityLabels(t *testing.T) {
+	t.Parallel()
+	registry := New("ins_test-1", "v-test")
+	registry.ObserveReconcileRun(nil, 1500*time.Millisecond)
+	registry.ObserveReconcileItem("success")
+	registry.ObserveReconcileItem("lease_lost")
+	registry.SetConvergence(16, 2)
+	registry.ObserveOrphanScan(20, 1, 3, nil)
+	registry.ObserveOrphanCleanupBlocked("ratio_exceeded")
+	registry.ObserveOrphanCleanup(1, 1)
+
+	request := httptest.NewRequest("GET", "/metrics", nil)
+	response := httptest.NewRecorder()
+	registry.Handler().ServeHTTP(response, request)
+	body := response.Body.String()
+	for _, expected := range []string{
+		`sense_build_info{instance_id="ins_test-1",version="v-test"} 1`,
+		`sense_reconcile_items_total{result="lease_lost"} 1`,
+		`sense_reconcile_devices{state="unconverged"} 2`,
+		`sense_orphan_paths{classification="owned_stale"} 1`,
+		`sense_orphan_cleanup_blocked_total{reason="ratio_exceeded"} 1`,
+	} {
+		if !strings.Contains(body, expected) {
+			t.Fatalf("metrics output lacks %q:\n%s", expected, body)
+		}
+	}
+	for _, forbidden := range []string{
+		"tenant_id", "site_id", "device_id", "path_name", "rtsp://", "credential",
+	} {
+		if strings.Contains(body, forbidden) {
+			t.Fatalf("high-cardinality or secret-bearing label appeared: %q", forbidden)
+		}
+	}
+	if response.Header().Get("Cache-Control") != "no-store" {
+		t.Fatal("metrics responses must not be cached")
+	}
+}
+
+func TestRegistryIsSafeForConcurrentObservers(t *testing.T) {
+	t.Parallel()
+	registry := New("ins_concurrent", "dev")
+	var wait sync.WaitGroup
+	for index := 0; index < 32; index++ {
+		wait.Add(1)
+		go func() {
+			defer wait.Done()
+			for count := 0; count < 100; count++ {
+				registry.ObserveReconcileItem("success")
+				registry.SetConvergence(128, count%17)
+			}
+		}()
+	}
+	wait.Wait()
+	response := httptest.NewRecorder()
+	registry.Handler().ServeHTTP(response, httptest.NewRequest("GET", "/metrics", nil))
+	if !strings.Contains(response.Body.String(), `sense_reconcile_items_total{result="success"} 3200`) {
+		t.Fatalf("concurrent counter lost updates: %s", response.Body.String())
+	}
+}
+
+func TestGenerateInstanceIDShape(t *testing.T) {
+	t.Parallel()
+	first, err := GenerateInstanceID()
+	if err != nil {
+		t.Fatal(err)
+	}
+	second, err := GenerateInstanceID()
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(first) != 28 || !strings.HasPrefix(first, "ins_") || first == second {
+		t.Fatalf("unexpected instance identifiers: %q %q", first, second)
+	}
+}
@@ -5,6 +5,7 @@ import (
 	"errors"
 	"fmt"
 	"net/http"
+	"sort"
 	"strings"

 	mediamtxapi "yovision/sense/internal/mtx/generated"
@@ -31,9 +32,63 @@ type pathAPI interface {
 	ConfigPathsGetWithResponse(context.Context, string, ...mediamtxapi.RequestEditorFn) (*mediamtxapi.ConfigPathsGetResponse, error)
 	ConfigPathsPatchWithResponse(context.Context, string, mediamtxapi.ConfigPathsPatchJSONRequestBody, ...mediamtxapi.RequestEditorFn) (*mediamtxapi.ConfigPathsPatchResponse, error)
 	ConfigPathsDeleteWithResponse(context.Context, string, ...mediamtxapi.RequestEditorFn) (*mediamtxapi.ConfigPathsDeleteResponse, error)
+	ConfigPathsListWithResponse(context.Context, *mediamtxapi.ConfigPathsListParams, ...mediamtxapi.RequestEditorFn) (*mediamtxapi.ConfigPathsListResponse, error)
 	PathsGetWithResponse(context.Context, string, ...mediamtxapi.RequestEditorFn) (*mediamtxapi.PathsGetResponse, error)
 }

+// ListPathNames enumerates only configuration names. Sources are deliberately
+// discarded so inventory and orphan reports cannot expose stream URIs.
+func (c *Client) ListPathNames(ctx context.Context) ([]string, error) {
+	const (
+		itemsPerPage = 100
+		maxPages     = 1000
+	)
+	result := make(map[string]struct{})
+	seenPages := make(map[string]struct{})
+	for page := 0; page < maxPages; page++ {
+		pageValue, limitValue := page, itemsPerPage
+		response, err := c.api.ConfigPathsListWithResponse(ctx, &mediamtxapi.ConfigPathsListParams{
+			Page: &pageValue, ItemsPerPage: &limitValue,
+		})
+		if err != nil {
+			return nil, fmt.Errorf("MediaMTX list paths transport: %w", err)
+		}
+		if response.StatusCode() != http.StatusOK || response.JSON200 == nil ||
+			response.JSON200.PageCount == nil || response.JSON200.Items == nil {
+			return nil, &APIError{Operation: "list paths", StatusCode: response.StatusCode()}
+		}
+		pageCount := *response.JSON200.PageCount
+		if pageCount < 0 || pageCount > maxPages {
+			return nil, &APIError{Operation: "list paths pagination", StatusCode: response.StatusCode()}
+		}
+		pageNames := make([]string, 0, len(*response.JSON200.Items))
+		for _, item := range *response.JSON200.Items {
+			if item.Name == nil || strings.TrimSpace(*item.Name) == "" {
+				return nil, &APIError{Operation: "list paths response", StatusCode: response.StatusCode()}
+			}
+			pageNames = append(pageNames, *item.Name)
+			result[*item.Name] = struct{}{}
+		}
+		sort.Strings(pageNames)
+		signature := strings.Join(pageNames, "\x00")
+		if page > 0 && signature != "" {
+			if _, duplicate := seenPages[signature]; duplicate {
+				return nil, &APIError{Operation: "list paths repeated page", StatusCode: response.StatusCode()}
+			}
+		}
+		seenPages[signature] = struct{}{}
+		if int64(page+1) >= pageCount {
+			values := make([]string, 0, len(result))
+			for name := range result {
+				values = append(values, name)
+			}
+			sort.Strings(values)
+			return values, nil
+		}
+	}
+	return nil, &APIError{Operation: "list paths page limit", StatusCode: http.StatusOK}
+}
+
 type Client struct {
 	api pathAPI
 }
@@ -6,6 +6,8 @@ import (
 	"fmt"
 	"net/http"
 	"net/http/httptest"
+	"sort"
+	"strconv"
 	"strings"
 	"sync"
 	"testing"
@@ -21,6 +23,38 @@ func (f *fakeMediaMTX) ServeHTTP(writer http.ResponseWriter, request *http.Reque
 	f.mu.Lock()
 	defer f.mu.Unlock()
 	writer.Header().Set("Content-Type", "application/json")
+	if request.URL.Path == "/v3/config/paths/list" {
+		page, _ := strconv.Atoi(request.URL.Query().Get("page"))
+		limit, _ := strconv.Atoi(request.URL.Query().Get("itemsPerPage"))
+		if limit <= 0 {
+			limit = 100
+		}
+		names := make([]string, 0, len(f.paths))
+		for name := range f.paths {
+			names = append(names, name)
+		}
+		sort.Strings(names)
+		start := page * limit
+		if start > len(names) {
+			start = len(names)
+		}
+		end := start + limit
+		if end > len(names) {
+			end = len(names)
+		}
+		items := make([]map[string]any, 0, end-start)
+		for _, name := range names[start:end] {
+			items = append(items, map[string]any{"name": name})
+		}
+		pageCount := 0
+		if len(names) > 0 {
+			pageCount = (len(names) + limit - 1) / limit
+		}
+		_ = json.NewEncoder(writer).Encode(map[string]any{
+			"itemCount": len(names), "pageCount": pageCount, "items": items,
+		})
+		return
+	}
 	prefixes := map[string]string{
 		"/v3/config/paths/get/":    "get",
 		"/v3/config/paths/add/":    "add",
@@ -72,6 +106,54 @@ func (f *fakeMediaMTX) ServeHTTP(writer http.ResponseWriter, request *http.Reque
 	http.NotFound(writer, request)
 }

+func TestListPathNamesUsesPaginationAndDoesNotReturnSources(t *testing.T) {
+	t.Parallel()
+	paths := make(map[string]string)
+	for index := 0; index < 205; index++ {
+		paths[fmt.Sprintf("camera-%03d", index)] = fmt.Sprintf("rtsp://secret.invalid/%d", index)
+	}
+	fake := &fakeMediaMTX{paths: paths}
+	server := httptest.NewServer(fake)
+	defer server.Close()
+	client, err := NewClient(server.URL, server.Client())
+	if err != nil {
+		t.Fatal(err)
+	}
+	values, err := client.ListPathNames(context.Background())
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(values) != 205 || values[0] != "camera-000" || values[204] != "camera-204" {
+		t.Fatalf("unexpected path inventory: len=%d first=%q last=%q", len(values), values[0], values[len(values)-1])
+	}
+	for _, value := range values {
+		if strings.Contains(value, "rtsp") || strings.Contains(value, "secret") {
+			t.Fatalf("source leaked from path inventory: %q", value)
+		}
+	}
+}
+
+func TestListPathNamesRejectsRepeatedPage(t *testing.T) {
+	t.Parallel()
+	server := httptest.NewServer(http.HandlerFunc(func(writer http.ResponseWriter, request *http.Request) {
+		writer.Header().Set("Content-Type", "application/json")
+		if request.URL.Path != "/v3/config/paths/list" {
+			http.NotFound(writer, request)
+			return
+		}
+		_, _ = writer.Write([]byte(`{"itemCount":2,"pageCount":2,"items":[{"name":"same"}]}`))
+	}))
+	defer server.Close()
+	client, err := NewClient(server.URL, server.Client())
+	if err != nil {
+		t.Fatal(err)
+	}
+	if _, err := client.ListPathNames(context.Background()); err == nil ||
+		!strings.Contains(err.Error(), "repeated page") {
+		t.Fatalf("repeated pagination was accepted: %v", err)
+	}
+}
+
 func TestGeneratedClientCreateReadDeleteMapping(t *testing.T) {
 	t.Parallel()
 	fake := &fakeMediaMTX{paths: make(map[string]string)}
@@ -0,0 +1,400 @@
+// Package orphan reports MediaMTX configuration paths that do not match the
+// current Sense ledger. Only paths with durable Sense ownership evidence can
+// ever enter the controlled cleanup set.
+package orphan
+
+import (
+	"context"
+	"crypto/rand"
+	"errors"
+	"fmt"
+	"math/big"
+	"regexp"
+	"sort"
+	"strings"
+	"time"
+
+	"yovision/sense/internal/metrics"
+	"yovision/sense/internal/store"
+)
+
+const (
+	defaultLeaseDuration    = 30 * time.Second
+	defaultOperationTimeout = 20 * time.Second
+	reportTTL               = 15 * time.Minute
+	maxCleanupItems         = 128
+	crockford               = "0123456789ABCDEFGHJKMNPQRSTVWXYZ"
+)
+
+var (
+	ErrLeaseHeld       = errors.New("orphan operation lease held")
+	ErrConfirmation    = errors.New("orphan cleanup confirmation invalid")
+	ErrSnapshotExpired = errors.New("orphan scan snapshot expired")
+	ErrSafetyBlocked   = errors.New("orphan cleanup safety gate blocked")
+	operatorIDPattern  = regexp.MustCompile(`^[A-Za-z0-9][A-Za-z0-9._-]{0,63}$`)
+)
+
+type Runtime interface {
+	ListPathNames(context.Context) ([]string, error)
+	DeletePath(context.Context, string) error
+}
+
+type Manager struct {
+	repository       store.OrphanRepository
+	runtime          Runtime
+	instanceID       string
+	metrics          *metrics.Registry
+	now              func() time.Time
+	leaseDuration    time.Duration
+	operationTimeout time.Duration
+}
+
+type CleanupResult struct {
+	ScanID  string `json:"scan_id"`
+	Deleted int    `json:"deleted"`
+	Failed  int    `json:"failed"`
+}
+
+func New(
+	repository store.OrphanRepository,
+	runtime Runtime,
+	instanceID string,
+	registry *metrics.Registry,
+) *Manager {
+	return &Manager{
+		repository: repository, runtime: runtime, instanceID: instanceID, metrics: registry,
+		now: time.Now, leaseDuration: defaultLeaseDuration, operationTimeout: defaultOperationTimeout,
+	}
+}
+
+func (m *Manager) Report(ctx context.Context) (store.OrphanScan, error) {
+	now := m.now().UTC()
+	token, err := randomToken()
+	if err != nil {
+		return store.OrphanScan{}, err
+	}
+	acquired, err := m.repository.AcquireOperationalLease(
+		ctx, store.OperationalLeaseOrphanScan, m.instanceID, token, now, m.leaseDuration,
+	)
+	if err != nil {
+		m.observeScan(0, 0, 0, err)
+		return store.OrphanScan{}, err
+	}
+	if !acquired {
+		return store.OrphanScan{}, ErrLeaseHeld
+	}
+	saved := false
+	defer func() {
+		if saved {
+			return
+		}
+		releaseCtx, cancel := context.WithTimeout(context.Background(), time.Second)
+		defer cancel()
+		_ = m.repository.ReleaseOperationalLease(
+			releaseCtx, store.OperationalLeaseOrphanScan, m.instanceID, token, m.now().UTC(),
+		)
+	}()
+
+	operationCtx, cancel := context.WithTimeout(ctx, m.operationTimeout)
+	paths, err := m.runtime.ListPathNames(operationCtx)
+	cancel()
+	if err != nil {
+		m.observeScan(0, 0, 0, err)
+		return store.OrphanScan{}, fmt.Errorf("list MediaMTX paths: %w", err)
+	}
+	ownership, err := m.repository.ListMediaPathOwnership(ctx)
+	if err != nil {
+		m.observeScan(0, 0, 0, err)
+		return store.OrphanScan{}, err
+	}
+	findings, stale, unowned := classify(paths, ownership)
+	allowed, reason := safetyGate(stale, len(paths))
+	completedAt := m.now().UTC()
+	acquired, err = m.repository.AcquireOperationalLease(
+		ctx, store.OperationalLeaseOrphanScan, m.instanceID, token, completedAt, m.leaseDuration,
+	)
+	if err != nil || !acquired {
+		if err == nil {
+			err = store.ErrOperationalLeaseLost
+		}
+		m.observeScan(len(paths), stale, unowned, err)
+		return store.OrphanScan{}, err
+	}
+	id, err := newScanID(completedAt)
+	if err != nil {
+		m.observeScan(0, 0, 0, err)
+		return store.OrphanScan{}, err
+	}
+	scan := store.OrphanScan{
+		ID: id, InstanceID: m.instanceID, ObservedCount: len(paths),
+		OwnedStaleCount: stale, UnownedCount: unowned,
+		SafetyAllowed: allowed, SafetyReason: reason,
+		CompletedAt: completedAt, ExpiresAt: completedAt.Add(reportTTL), Findings: findings,
+	}
+	if err := m.repository.SaveOrphanScan(ctx, scan, m.instanceID, token); err != nil {
+		m.observeScan(len(paths), stale, unowned, err)
+		return store.OrphanScan{}, err
+	}
+	saved = true
+	m.observeScan(len(paths), stale, unowned, nil)
+	return scan, nil
+}
+
+func classify(
+	paths []string,
+	ownership []store.MediaPathOwnership,
+) ([]store.OrphanFinding, int, int) {
+	history := make(map[string]store.MediaPathOwnership, len(ownership))
+	for _, value := range ownership {
+		history[value.PathName] = value
+	}
+	unique := make(map[string]struct{}, len(paths))
+	for _, path := range paths {
+		unique[path] = struct{}{}
+	}
+	ordered := make([]string, 0, len(unique))
+	for path := range unique {
+		ordered = append(ordered, path)
+	}
+	sort.Strings(ordered)
+	findings := make([]store.OrphanFinding, 0)
+	stale, unowned := 0, 0
+	for _, path := range ordered {
+		value, known := history[path]
+		switch {
+		case known && value.CurrentClaim:
+			continue
+		case known:
+			stale++
+			findings = append(findings, store.OrphanFinding{
+				PathName: path, Classification: store.OrphanOwnedStale, DeviceID: value.DeviceID,
+			})
+		default:
+			unowned++
+			findings = append(findings, store.OrphanFinding{
+				PathName: path, Classification: store.OrphanUnowned,
+			})
+		}
+	}
+	return findings, stale, unowned
+}
+
+func safetyGate(candidates, observed int) (bool, string) {
+	switch {
+	case candidates <= 0:
+		return false, "no_candidates"
+	case observed <= 0:
+		return false, "empty_inventory"
+	case candidates > maxCleanupItems:
+		return false, "scope_exceeded"
+	case candidates*100 > observed*10:
+		return false, "ratio_exceeded"
+	default:
+		return true, "allowed"
+	}
+}
+
+func (m *Manager) Apply(
+	ctx context.Context,
+	scanID, actorID, confirmation string,
+) (CleanupResult, error) {
+	result := CleanupResult{ScanID: scanID}
+	if !operatorIDPattern.MatchString(actorID) || confirmation != "DELETE "+scanID {
+		m.blocked("scope_changed")
+		return result, ErrConfirmation
+	}
+	now := m.now().UTC()
+	token, err := randomToken()
+	if err != nil {
+		return result, err
+	}
+	acquired, err := m.repository.AcquireOperationalLease(
+		ctx, store.OperationalLeaseOrphanCleanup, m.instanceID, token, now, m.leaseDuration,
+	)
+	if err != nil {
+		return result, err
+	}
+	if !acquired {
+		return result, ErrLeaseHeld
+	}
+	defer func() {
+		releaseCtx, cancel := context.WithTimeout(context.Background(), time.Second)
+		defer cancel()
+		_ = m.repository.ReleaseOperationalLease(
+			releaseCtx, store.OperationalLeaseOrphanCleanup, m.instanceID, token, m.now().UTC(),
+		)
+	}()
+
+	scan, err := m.repository.GetOrphanScan(ctx, scanID)
+	if err != nil {
+		m.blocked("scope_changed")
+		return result, err
+	}
+	if !scan.ExpiresAt.After(now) {
+		m.blocked("snapshot_expired")
+		return result, ErrSnapshotExpired
+	}
+	if !scan.SafetyAllowed {
+		m.blocked(scan.SafetyReason)
+		return result, fmt.Errorf("%w: %s", ErrSafetyBlocked, scan.SafetyReason)
+	}
+	operationCtx, cancel := context.WithTimeout(ctx, m.operationTimeout)
+	paths, err := m.runtime.ListPathNames(operationCtx)
+	cancel()
+	if err != nil {
+		return result, fmt.Errorf("refresh MediaMTX path inventory: %w", err)
+	}
+	ownership, err := m.repository.ListMediaPathOwnership(ctx)
+	if err != nil {
+		return result, err
+	}
+	candidates := cleanupCandidates(scan, paths, ownership)
+	if len(candidates) == 0 {
+		return result, nil
+	}
+	allowed, reason := safetyGate(len(candidates), len(uniqueNames(paths)))
+	if !allowed {
+		m.blocked(reason)
+		return result, fmt.Errorf("%w: %s", ErrSafetyBlocked, reason)
+	}
+	var failures []error
+	for _, path := range candidates {
+		now = m.now().UTC()
+		acquired, err = m.repository.AcquireOperationalLease(
+			ctx, store.OperationalLeaseOrphanCleanup, m.instanceID, token, now, m.leaseDuration,
+		)
+		if err != nil || !acquired {
+			if err == nil {
+				err = ErrLeaseHeld
+			}
+			failures = append(failures, err)
+			break
+		}
+		itemCtx, itemCancel := context.WithTimeout(ctx, m.operationTimeout)
+		deleteErr := m.runtime.DeletePath(itemCtx, path)
+		itemCancel()
+		status, code := "deleted", ""
+		if deleteErr != nil {
+			status, code = "failed", "media_error"
+			result.Failed++
+			failures = append(failures, fmt.Errorf("delete owned stale path: %w", deleteErr))
+		} else {
+			result.Deleted++
+		}
+		if err := m.repository.RecordOrphanCleanup(
+			ctx, scanID, path, actorID, status, code, m.now().UTC(),
+		); err != nil {
+			failures = append(failures, err)
+			break
+		}
+	}
+	if m.metrics != nil {
+		m.metrics.ObserveOrphanCleanup(result.Deleted, result.Failed)
+	}
+	return result, errors.Join(failures...)
+}
+
+func cleanupCandidates(
+	scan store.OrphanScan,
+	paths []string,
+	ownership []store.MediaPathOwnership,
+) []string {
+	runtime := uniqueNames(paths)
+	history := make(map[string]store.MediaPathOwnership, len(ownership))
+	for _, value := range ownership {
+		history[value.PathName] = value
+	}
+	values := make([]string, 0)
+	for _, finding := range scan.Findings {
+		if finding.Classification != store.OrphanOwnedStale || finding.Deleted {
+			continue
+		}
+		_, present := runtime[finding.PathName]
+		owner, known := history[finding.PathName]
+		if present && known && !owner.CurrentClaim && owner.DeviceID == finding.DeviceID {
+			values = append(values, finding.PathName)
+		}
+	}
+	sort.Strings(values)
+	return values
+}
+
+func uniqueNames(values []string) map[string]struct{} {
+	result := make(map[string]struct{}, len(values))
+	for _, value := range values {
+		result[value] = struct{}{}
+	}
+	return result
+}
+
+func (m *Manager) Run(ctx context.Context, interval time.Duration, report func(error)) {
+	run := func() {
+		_, err := m.Report(ctx)
+		if err != nil && !errors.Is(err, ErrLeaseHeld) && ctx.Err() == nil && report != nil {
+			report(err)
+		}
+	}
+	run()
+	ticker := time.NewTicker(interval)
+	defer ticker.Stop()
+	for {
+		select {
+		case <-ctx.Done():
+			return
+		case <-ticker.C:
+			run()
+		}
+	}
+}
+
+func (m *Manager) observeScan(observed, stale, unowned int, err error) {
+	if m.metrics != nil {
+		m.metrics.ObserveOrphanScan(observed, stale, unowned, err)
+	}
+}
+
+func (m *Manager) blocked(reason string) {
+	if m.metrics != nil {
+		m.metrics.ObserveOrphanCleanupBlocked(reason)
+	}
+}
+
+func randomToken() (string, error) {
+	value := make([]byte, 16)
+	if _, err := rand.Read(value); err != nil {
+		return "", errors.New("generate fencing token")
+	}
+	const alphabet = "0123456789abcdefghijklmnopqrstuvwxyzABCDEFGHIJKLMNOPQRSTUVWXYZ"
+	result := make([]byte, 22)
+	number := new(big.Int).SetBytes(value)
+	base := big.NewInt(int64(len(alphabet)))
+	remainder := new(big.Int)
+	for index := len(result) - 1; index >= 0; index-- {
+		number.QuoRem(number, base, remainder)
+		result[index] = alphabet[remainder.Int64()]
+	}
+	return string(result), nil
+}
+
+func newScanID(now time.Time) (string, error) {
+	value := make([]byte, 16)
+	milliseconds := uint64(now.UTC().UnixMilli())
+	value[0], value[1], value[2] = byte(milliseconds>>40), byte(milliseconds>>32), byte(milliseconds>>24)
+	value[3], value[4], value[5] = byte(milliseconds>>16), byte(milliseconds>>8), byte(milliseconds)
+	if _, err := rand.Read(value[6:]); err != nil {
+		return "", errors.New("generate orphan scan identifier")
+	}
+	number := new(big.Int).SetBytes(value)
+	base := big.NewInt(32)
+	remainder := new(big.Int)
+	encoded := make([]byte, 26)
+	for index := len(encoded) - 1; index >= 0; index-- {
+		number.QuoRem(number, base, remainder)
+		encoded[index] = crockford[remainder.Int64()]
+	}
+	return "scan_" + string(encoded), nil
+}
+
+func Confirmation(scanID string) string {
+	return strings.Join([]string{"DELETE", scanID}, " ")
+}
@@ -0,0 +1,272 @@
+package orphan
+
+import (
+	"context"
+	"errors"
+	"sort"
+	"sync"
+	"testing"
+	"time"
+
+	"yovision/sense/internal/store"
+)
+
+type fakeLease struct {
+	owner, token string
+	expires      time.Time
+}
+
+type fakeRepository struct {
+	mu        sync.Mutex
+	leases    map[string]fakeLease
+	ownership []store.MediaPathOwnership
+	scans     map[string]store.OrphanScan
+	deleted   map[string]map[string]bool
+}
+
+func newFakeRepository() *fakeRepository {
+	return &fakeRepository{
+		leases: make(map[string]fakeLease), scans: make(map[string]store.OrphanScan),
+		deleted: make(map[string]map[string]bool),
+	}
+}
+
+func (f *fakeRepository) AcquireOperationalLease(
+	_ context.Context,
+	name, owner, token string,
+	now time.Time,
+	duration time.Duration,
+) (bool, error) {
+	f.mu.Lock()
+	defer f.mu.Unlock()
+	current, exists := f.leases[name]
+	if exists && current.expires.After(now) && (current.owner != owner || current.token != token) {
+		return false, nil
+	}
+	f.leases[name] = fakeLease{owner: owner, token: token, expires: now.Add(duration)}
+	return true, nil
+}
+
+func (f *fakeRepository) ReleaseOperationalLease(
+	_ context.Context,
+	name, owner, token string,
+	now time.Time,
+) error {
+	f.mu.Lock()
+	defer f.mu.Unlock()
+	current := f.leases[name]
+	if current.owner == owner && current.token == token {
+		current.expires = now
+		f.leases[name] = current
+	}
+	return nil
+}
+
+func (f *fakeRepository) ListMediaPathOwnership(context.Context) ([]store.MediaPathOwnership, error) {
+	f.mu.Lock()
+	defer f.mu.Unlock()
+	return append([]store.MediaPathOwnership(nil), f.ownership...), nil
+}
+
+func (f *fakeRepository) SaveOrphanScan(
+	_ context.Context,
+	scan store.OrphanScan,
+	owner, token string,
+) error {
+	f.mu.Lock()
+	defer f.mu.Unlock()
+	lease := f.leases[store.OperationalLeaseOrphanScan]
+	if lease.owner != owner || lease.token != token || !lease.expires.After(scan.CompletedAt) {
+		return store.ErrOperationalLeaseLost
+	}
+	f.scans[scan.ID] = cloneScan(scan)
+	lease.expires = scan.CompletedAt
+	f.leases[store.OperationalLeaseOrphanScan] = lease
+	return nil
+}
+
+func (f *fakeRepository) GetOrphanScan(_ context.Context, id string) (store.OrphanScan, error) {
+	f.mu.Lock()
+	defer f.mu.Unlock()
+	scan, exists := f.scans[id]
+	if !exists {
+		return store.OrphanScan{}, store.ErrOrphanScanNotFound
+	}
+	result := cloneScan(scan)
+	for index := range result.Findings {
+		result.Findings[index].Deleted = f.deleted[id][result.Findings[index].PathName]
+	}
+	return result, nil
+}
+
+func (f *fakeRepository) RecordOrphanCleanup(
+	_ context.Context,
+	scanID, pathName, _, status, _ string,
+	_ time.Time,
+) error {
+	f.mu.Lock()
+	defer f.mu.Unlock()
+	if f.deleted[scanID] == nil {
+		f.deleted[scanID] = make(map[string]bool)
+	}
+	if status == "deleted" {
+		f.deleted[scanID][pathName] = true
+	}
+	return nil
+}
+
+func cloneScan(value store.OrphanScan) store.OrphanScan {
+	value.Findings = append([]store.OrphanFinding(nil), value.Findings...)
+	return value
+}
+
+type fakeRuntime struct {
+	paths   map[string]bool
+	fail    map[string]bool
+	deleted []string
+}
+
+func (f *fakeRuntime) ListPathNames(context.Context) ([]string, error) {
+	values := make([]string, 0, len(f.paths))
+	for path, present := range f.paths {
+		if present {
+			values = append(values, path)
+		}
+	}
+	sort.Strings(values)
+	return values, nil
+}
+
+func (f *fakeRuntime) DeletePath(_ context.Context, path string) error {
+	if f.fail[path] {
+		return errors.New("redacted media failure")
+	}
+	delete(f.paths, path)
+	f.deleted = append(f.deleted, path)
+	return nil
+}
+
+func TestSafetyGateUsesExactTenPercentBoundary(t *testing.T) {
+	t.Parallel()
+	if allowed, reason := safetyGate(1, 10); !allowed || reason != "allowed" {
+		t.Fatalf("exact 10%% boundary was rejected: %v %s", allowed, reason)
+	}
+	if allowed, reason := safetyGate(1, 9); allowed || reason != "ratio_exceeded" {
+		t.Fatalf("more than 10%% was accepted: %v %s", allowed, reason)
+	}
+	if allowed, reason := safetyGate(129, 2000); allowed || reason != "scope_exceeded" {
+		t.Fatalf("129-item scope was accepted: %v %s", allowed, reason)
+	}
+}
+
+func TestReportClassifiesOwnershipAndApplyNeverDeletesUnowned(t *testing.T) {
+	repository := newFakeRepository()
+	repository.ownership = []store.MediaPathOwnership{
+		{PathName: "owned-stale", DeviceID: "old-device"},
+		{PathName: "owned-current", DeviceID: "live-device", CurrentClaim: true},
+	}
+	runtime := &fakeRuntime{paths: map[string]bool{
+		"owned-stale": true, "owned-current": true,
+		"unowned-1": true, "unowned-2": true, "unowned-3": true, "unowned-4": true,
+		"unowned-5": true, "unowned-6": true, "unowned-7": true, "unowned-8": true,
+	}, fail: make(map[string]bool)}
+	now := time.Date(2026, 8, 7, 0, 0, 0, 0, time.UTC)
+	manager := New(repository, runtime, "ins_test", nil)
+	manager.now = func() time.Time { return now }
+	scan, err := manager.Report(context.Background())
+	if err != nil {
+		t.Fatal(err)
+	}
+	if scan.ObservedCount != 10 || scan.OwnedStaleCount != 1 || scan.UnownedCount != 8 || !scan.SafetyAllowed {
+		t.Fatalf("unexpected report: %+v", scan)
+	}
+	result, err := manager.Apply(
+		context.Background(), scan.ID, "operator-1", Confirmation(scan.ID),
+	)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if result.Deleted != 1 || len(runtime.deleted) != 1 || runtime.deleted[0] != "owned-stale" {
+		t.Fatalf("cleanup escaped owned stale set: result=%+v deleted=%v", result, runtime.deleted)
+	}
+	for _, path := range []string{"owned-current", "unowned-1", "unowned-8"} {
+		if !runtime.paths[path] {
+			t.Fatalf("cleanup deleted protected path %q", path)
+		}
+	}
+}
+
+func TestApplyBlocksExpiredOrOversizedSnapshotBeforeDelete(t *testing.T) {
+	repository := newFakeRepository()
+	repository.ownership = []store.MediaPathOwnership{{PathName: "stale", DeviceID: "old"}}
+	runtime := &fakeRuntime{paths: map[string]bool{"stale": true}, fail: make(map[string]bool)}
+	now := time.Date(2026, 8, 7, 0, 0, 0, 0, time.UTC)
+	manager := New(repository, runtime, "ins_test", nil)
+	manager.now = func() time.Time { return now }
+	scan, err := manager.Report(context.Background())
+	if err != nil {
+		t.Fatal(err)
+	}
+	if scan.SafetyAllowed || scan.SafetyReason != "ratio_exceeded" {
+		t.Fatalf("single-path 100%% cleanup was not blocked: %+v", scan)
+	}
+	if _, err := manager.Apply(context.Background(), scan.ID, "operator", Confirmation(scan.ID)); !errors.Is(err, ErrSafetyBlocked) {
+		t.Fatalf("expected ratio gate, got %v", err)
+	}
+	if len(runtime.deleted) != 0 {
+		t.Fatal("ratio-blocked cleanup mutated MediaMTX")
+	}
+
+	for index := 0; index < 9; index++ {
+		runtime.paths[string(rune('a'+index))] = true
+	}
+	scan, err = manager.Report(context.Background())
+	if err != nil || !scan.SafetyAllowed {
+		t.Fatalf("expected a fresh 1/10 executable report: %+v %v", scan, err)
+	}
+	now = now.Add(reportTTL)
+	if _, err := manager.Apply(context.Background(), scan.ID, "operator", Confirmation(scan.ID)); !errors.Is(err, ErrSnapshotExpired) {
+		t.Fatalf("expected expired snapshot, got %v", err)
+	}
+	if len(runtime.deleted) != 0 {
+		t.Fatal("expired cleanup mutated MediaMTX")
+	}
+}
+
+func TestApplyIsRetryableAndCannotExpandBeyondSnapshot(t *testing.T) {
+	repository := newFakeRepository()
+	repository.ownership = []store.MediaPathOwnership{
+		{PathName: "stale-a", DeviceID: "old-a"},
+		{PathName: "stale-b", DeviceID: "old-b"},
+	}
+	runtime := &fakeRuntime{paths: make(map[string]bool), fail: map[string]bool{"stale-b": true}}
+	runtime.paths["stale-a"], runtime.paths["stale-b"] = true, true
+	for index := 0; index < 18; index++ {
+		runtime.paths[string(rune(0x100+index))] = true
+	}
+	now := time.Date(2026, 8, 7, 0, 0, 0, 0, time.UTC)
+	manager := New(repository, runtime, "ins_test", nil)
+	manager.now = func() time.Time { return now }
+	scan, err := manager.Report(context.Background())
+	if err != nil || !scan.SafetyAllowed {
+		t.Fatalf("expected 2/20 report: %+v %v", scan, err)
+	}
+	repository.ownership = append(repository.ownership,
+		store.MediaPathOwnership{PathName: "new-stale", DeviceID: "new-old"})
+	runtime.paths["new-stale"] = true
+	result, err := manager.Apply(context.Background(), scan.ID, "operator", Confirmation(scan.ID))
+	if err == nil || result.Deleted != 1 || result.Failed != 1 {
+		t.Fatalf("expected one partial failure: %+v %v", result, err)
+	}
+	if !runtime.paths["new-stale"] {
+		t.Fatal("cleanup expanded beyond the approved snapshot")
+	}
+	runtime.fail["stale-b"] = false
+	result, err = manager.Apply(context.Background(), scan.ID, "operator", Confirmation(scan.ID))
+	if err != nil || result.Deleted != 1 || result.Failed != 0 {
+		t.Fatalf("failed item was not retryable: %+v %v", result, err)
+	}
+	if !runtime.paths["new-stale"] {
+		t.Fatal("retry expanded beyond the approved snapshot")
+	}
+}
@@ -3,12 +3,15 @@ package reconcile

 import (
 	"context"
+	"crypto/rand"
+	"encoding/base64"
 	"errors"
 	"fmt"
 	"net/url"
 	"time"

 	"yovision/sense/internal/device"
+	"yovision/sense/internal/metrics"
 	"yovision/sense/internal/onvif"
 	"yovision/sense/internal/store"
 )
@@ -16,9 +19,11 @@ import (
 const defaultBatchSize = 128

 type Repository interface {
-	ListDueReconcile(ctx context.Context, now time.Time, limit int) ([]store.ReconcileCandidate, error)
-	MarkReconciled(ctx context.Context, id string, generation int64, now time.Time) error
-	MarkReconcileFailure(ctx context.Context, id string, failureCount int, nextAttempt time.Time, errorCode string, now time.Time) error
+	ClaimDueReconcile(context.Context, store.ReconcileClaim) ([]store.ReconcileCandidate, error)
+	RenewReconcileLease(context.Context, string, string, string, time.Time, time.Duration) (bool, error)
+	CompleteReconcile(context.Context, string, int64, string, string, time.Time) error
+	FailReconcile(context.Context, string, int, time.Time, string, string, string, time.Time) error
+	ConvergenceSnapshot(context.Context) (store.ConvergenceSnapshot, error)
 }

 type MediaPaths interface {
@@ -27,48 +32,134 @@ type MediaPaths interface {
 }

 type Reconciler struct {
-	repository  Repository
-	discovery   onvif.Adapter
-	media       MediaPaths
-	now         func() time.Time
-	baseBackoff time.Duration
-	maxBackoff  time.Duration
-	batchSize   int
+	repository       Repository
+	discovery        onvif.Adapter
+	media            MediaPaths
+	now              func() time.Time
+	baseBackoff      time.Duration
+	maxBackoff       time.Duration
+	batchSize        int
+	instanceID       string
+	leaseDuration    time.Duration
+	operationTimeout time.Duration
+	metrics          *metrics.Registry
+}
+
+type Options struct {
+	InstanceID       string
+	LeaseDuration    time.Duration
+	OperationTimeout time.Duration
+	Metrics          *metrics.Registry
 }

 func New(repository Repository, discovery onvif.Adapter, media MediaPaths) *Reconciler {
+	return NewWithOptions(repository, discovery, media, Options{})
+}
+
+func NewWithOptions(
+	repository Repository,
+	discovery onvif.Adapter,
+	media MediaPaths,
+	options Options,
+) *Reconciler {
+	if options.InstanceID == "" {
+		options.InstanceID = "single"
+	}
+	if options.LeaseDuration <= 0 {
+		options.LeaseDuration = 30 * time.Second
+	}
+	if options.OperationTimeout <= 0 {
+		options.OperationTimeout = 20 * time.Second
+	}
+	if options.OperationTimeout >= options.LeaseDuration {
+		options.OperationTimeout = options.LeaseDuration / 2
+	}
 	return &Reconciler{
 		repository: repository, discovery: discovery, media: media,
 		now: time.Now, baseBackoff: time.Second, maxBackoff: time.Minute, batchSize: defaultBatchSize,
+		instanceID: options.InstanceID, leaseDuration: options.LeaseDuration,
+		operationTimeout: options.OperationTimeout, metrics: options.Metrics,
 	}
 }

-func (r *Reconciler) RunOnce(ctx context.Context) error {
+func (r *Reconciler) RunOnce(ctx context.Context) (runErr error) {
+	started := time.Now()
+	defer func() {
+		if r.metrics != nil {
+			r.metrics.ObserveReconcileRun(runErr, time.Since(started))
+		}
+	}()
 	now := r.now().UTC()
-	candidates, err := r.repository.ListDueReconcile(ctx, now, r.batchSize)
+	token, err := newClaimToken()
 	if err != nil {
-		return fmt.Errorf("list reconciliation candidates: %w", err)
+		return err
+	}
+	candidates, err := r.repository.ClaimDueReconcile(ctx, store.ReconcileClaim{
+		Owner: r.instanceID, Token: token, Now: now,
+		LeaseDuration: r.leaseDuration, Limit: r.batchSize,
+	})
+	if err != nil {
+		return fmt.Errorf("claim reconciliation candidates: %w", err)
 	}
 	var runErrors []error
 	for _, candidate := range candidates {
 		if err := ctx.Err(); err != nil {
 			return err
 		}
-		if err := r.reconcileOne(ctx, candidate, now); err != nil {
+		itemNow := r.now().UTC()
+		renewed, err := r.repository.RenewReconcileLease(
+			ctx, candidate.Device.ID, r.instanceID, token, itemNow, r.leaseDuration,
+		)
+		if err != nil {
+			r.observeItem("error")
+			runErrors = append(runErrors, fmt.Errorf("renew reconcile device lease: %w", err))
+			continue
+		}
+		if !renewed {
+			r.observeItem("lease_lost")
+			continue
+		}
+		itemCtx, cancel := context.WithTimeout(ctx, r.operationTimeout)
+		err = r.reconcileOne(itemCtx, ctx, candidate, itemNow, r.instanceID, token)
+		cancel()
+		if errors.Is(err, store.ErrReconcileLeaseLost) {
+			r.observeItem("lease_lost")
+			continue
+		}
+		if err != nil {
+			r.observeItem("error")
 			runErrors = append(runErrors, fmt.Errorf("reconcile device %s: %w", candidate.Device.ID, err))
+		} else {
+			r.observeItem("success")
+		}
+	}
+	if r.metrics != nil {
+		snapshot, err := r.repository.ConvergenceSnapshot(ctx)
+		if err != nil {
+			runErrors = append(runErrors, fmt.Errorf("read convergence metrics: %w", err))
+		} else {
+			r.metrics.SetConvergence(snapshot.Total, snapshot.Unconverged)
 		}
 	}
 	return errors.Join(runErrors...)
 }

-func (r *Reconciler) reconcileOne(ctx context.Context, candidate store.ReconcileCandidate, now time.Time) error {
+func (r *Reconciler) reconcileOne(
+	ctx context.Context,
+	persistCtx context.Context,
+	candidate store.ReconcileCandidate,
+	now time.Time,
+	owner, token string,
+) error {
 	if candidate.Device.DesiredState == device.DesiredDisabled {
 		if err := r.media.DeletePath(ctx, candidate.Device.PathName); err == nil {
-			return r.repository.MarkReconciled(ctx, candidate.Device.ID, candidate.Device.Generation, now)
-		} else if ctx.Err() != nil {
-			return ctx.Err()
+			return r.repository.CompleteReconcile(
+				persistCtx, candidate.Device.ID, candidate.Device.Generation, owner, token, now,
+			)
+		} else if persistCtx.Err() != nil {
+			return persistCtx.Err()
 		} else {
-			return r.persistFailure(ctx, candidate, now, err)
+			return r.persistFailure(persistCtx, candidate, now, owner, token, err)
 		}
 	}
 	result, err := r.discovery.Probe(ctx, onvif.Target{
@@ -81,16 +172,22 @@ func (r *Reconciler) reconcileOne(ctx context.Context, candidate store.Reconcile
 		_, err = r.media.EnsurePath(ctx, candidate.Device.PathName, result.StreamURI)
 	}
 	if err == nil {
-		return r.repository.MarkReconciled(ctx, candidate.Device.ID, candidate.Device.Generation, now)
+		return r.repository.CompleteReconcile(
+			persistCtx, candidate.Device.ID, candidate.Device.Generation, owner, token, now,
+		)
 	}
-	if ctx.Err() != nil {
-		return ctx.Err()
+	if persistCtx.Err() != nil {
+		return persistCtx.Err()
 	}
-	return r.persistFailure(ctx, candidate, now, err)
+	return r.persistFailure(persistCtx, candidate, now, owner, token, err)
 }

 func (r *Reconciler) persistFailure(
-	ctx context.Context, candidate store.ReconcileCandidate, now time.Time, err error,
+	ctx context.Context,
+	candidate store.ReconcileCandidate,
+	now time.Time,
+	owner, token string,
+	err error,
 ) error {
 	failureCount := candidate.FailureCount + 1
 	nextAttempt := now.Add(r.backoff(failureCount))
@@ -99,14 +196,28 @@ func (r *Reconciler) persistFailure(
 	if !errors.As(err, &onvifError) {
 		errorCode = "media_error"
 	}
-	if markErr := r.repository.MarkReconcileFailure(
-		ctx, candidate.Device.ID, failureCount, nextAttempt, errorCode, now,
+	if markErr := r.repository.FailReconcile(
+		ctx, candidate.Device.ID, failureCount, nextAttempt, errorCode, owner, token, now,
 	); markErr != nil {
 		return errors.Join(err, fmt.Errorf("persist reconcile failure: %w", markErr))
 	}
 	return err
 }

+func (r *Reconciler) observeItem(result string) {
+	if r.metrics != nil {
+		r.metrics.ObserveReconcileItem(result)
+	}
+}
+
+func newClaimToken() (string, error) {
+	value := make([]byte, 18)
+	if _, err := rand.Read(value); err != nil {
+		return "", errors.New("generate reconcile claim token")
+	}
+	return base64.RawURLEncoding.EncodeToString(value), nil
+}
+
 func validateStreamURI(value string) error {
 	parsed, err := url.Parse(value)
 	if err != nil || parsed.Host == "" || (parsed.Scheme != "rtsp" && parsed.Scheme != "rtsps") {
@@ -162,6 +162,86 @@ func TestDisabledDeviceDeletesOnlyItsExactPath(t *testing.T) {
 	}
 }

+type lostLeaseRepository struct {
+	completed bool
+	failed    bool
+}
+
+func (r *lostLeaseRepository) ClaimDueReconcile(
+	context.Context,
+	store.ReconcileClaim,
+) ([]store.ReconcileCandidate, error) {
+	return []store.ReconcileCandidate{{Device: device.Device{
+		ID: "camera-lost", DesiredState: device.DesiredEnabled,
+		EndpointRef: "onvif://camera-lost", PathName: "camera-lost", Generation: 1,
+	}}}, nil
+}
+
+func (r *lostLeaseRepository) RenewReconcileLease(
+	context.Context, string, string, string, time.Time, time.Duration,
+) (bool, error) {
+	return false, nil
+}
+
+func (r *lostLeaseRepository) CompleteReconcile(
+	context.Context, string, int64, string, string, time.Time,
+) error {
+	r.completed = true
+	return nil
+}
+
+func (r *lostLeaseRepository) FailReconcile(
+	context.Context, string, int, time.Time, string, string, string, time.Time,
+) error {
+	r.failed = true
+	return nil
+}
+
+func (r *lostLeaseRepository) ConvergenceSnapshot(context.Context) (store.ConvergenceSnapshot, error) {
+	return store.ConvergenceSnapshot{}, nil
+}
+
+func TestLostLeaseSkipsAllExternalAndStoreMutations(t *testing.T) {
+	t.Parallel()
+	repository := &lostLeaseRepository{}
+	media := &recordingMedia{}
+	reconciler := New(repository, onvif.NewFake(nil), media)
+	if err := reconciler.RunOnce(context.Background()); err != nil {
+		t.Fatal(err)
+	}
+	if media.calls != 0 || repository.completed || repository.failed {
+		t.Fatalf("lost lease performed a side effect: media=%d completed=%v failed=%v",
+			media.calls, repository.completed, repository.failed)
+	}
+}
+
+func TestPerItemTimeoutPersistsRetryWhenParentIsAlive(t *testing.T) {
+	t.Parallel()
+	repository := openRepository(t, filepath.Join(t.TempDir(), "sense.db"))
+	createReconcileDevice(t, repository)
+	discovery := onvif.NewFake(map[string]onvif.FakeScenario{
+		"onvif://camera-1": {
+			DelayMillis: 100,
+			Result:      onvif.ProbeResult{StreamURI: "rtsp://media.invalid/camera-1"},
+		},
+	})
+	reconciler := NewWithOptions(repository, discovery, &recordingMedia{}, Options{
+		InstanceID: "ins_test", LeaseDuration: time.Second, OperationTimeout: 5 * time.Millisecond,
+	})
+	now := time.Date(2026, 8, 7, 0, 0, 0, 0, time.UTC)
+	reconciler.now = func() time.Time { return now }
+	if err := reconciler.RunOnce(context.Background()); err == nil {
+		t.Fatal("expected bounded operation timeout")
+	}
+	candidates, err := repository.ListDueReconcile(context.Background(), now.Add(time.Hour), 1)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(candidates) != 1 || candidates[0].FailureCount != 1 {
+		t.Fatalf("operation timeout did not persist retry state: %+v", candidates)
+	}
+}
+
 func openRepository(t *testing.T, path string) *store.SQLite {
 	t.Helper()
 	repository, err := store.OpenSQLite(context.Background(), "file:"+filepath.ToSlash(path))
@@ -514,6 +514,7 @@ func (s *Postgres) PatchControlDevice(
 		if reconcileChanged {
 			if _, err := tx.ExecContext(ctx, `UPDATE sense.reconcile_state SET
 				failure_count = 0, next_attempt_at = NULL, last_error_code = NULL,
+				lease_owner = NULL, lease_token = NULL, lease_until = NULL,
 				updated_at = $1 WHERE device_id = $2`, now, deviceID); err != nil {
 				return ControlMutationResult{}, errors.New("reset postgres Control API reconcile state")
 			}
@@ -631,6 +632,7 @@ func setControlDesiredStateTx(
 		}
 		if _, err := tx.ExecContext(ctx, `UPDATE sense.reconcile_state SET
 			failure_count = 0, next_attempt_at = NULL, last_error_code = NULL,
+			lease_owner = NULL, lease_token = NULL, lease_until = NULL,
 			updated_at = $1 WHERE device_id = $2`, now, deviceID); err != nil {
 			return ControlDevice{}, errors.New("reset postgres Control API desired-state reconciliation")
 		}
@@ -0,0 +1,58 @@
+package store
+
+import (
+	"context"
+	"errors"
+	"time"
+)
+
+const (
+	OperationalLeaseOrphanScan    = "mediamtx-orphan-scan"
+	OperationalLeaseOrphanCleanup = "mediamtx-orphan-cleanup"
+
+	OrphanOwnedStale = "owned_stale"
+	OrphanUnowned    = "unowned"
+)
+
+var ErrOrphanScanNotFound = errors.New("orphan scan not found")
+
+type MediaPathOwnership struct {
+	PathName     string
+	DeviceID     string
+	CurrentClaim bool
+}
+
+type OrphanFinding struct {
+	PathName       string
+	Classification string
+	DeviceID       string
+	Deleted        bool
+}
+
+type OrphanScan struct {
+	ID              string
+	InstanceID      string
+	ObservedCount   int
+	OwnedStaleCount int
+	UnownedCount    int
+	SafetyAllowed   bool
+	SafetyReason    string
+	CompletedAt     time.Time
+	ExpiresAt       time.Time
+	Findings        []OrphanFinding
+}
+
+type OrphanRepository interface {
+	AcquireOperationalLease(
+		context.Context, string, string, string, time.Time, time.Duration,
+	) (bool, error)
+	ReleaseOperationalLease(context.Context, string, string, string, time.Time) error
+	ListMediaPathOwnership(context.Context) ([]MediaPathOwnership, error)
+	SaveOrphanScan(context.Context, OrphanScan, string, string) error
+	GetOrphanScan(context.Context, string) (OrphanScan, error)
+	RecordOrphanCleanup(
+		context.Context, string, string, string, string, string, time.Time,
+	) error
+}
+
+var _ OrphanRepository = (*Postgres)(nil)
@@ -0,0 +1,268 @@
+package store
+
+import (
+	"context"
+	"database/sql"
+	"errors"
+	"fmt"
+	"strings"
+	"time"
+)
+
+const orphanReportRetention = 7 * 24 * time.Hour
+
+func (s *Postgres) AcquireOperationalLease(
+	ctx context.Context,
+	name, owner, token string,
+	_ time.Time,
+	duration time.Duration,
+) (bool, error) {
+	if strings.TrimSpace(name) == "" || strings.TrimSpace(owner) == "" ||
+		strings.TrimSpace(token) == "" || duration <= 0 {
+		return false, errors.New("invalid operational lease")
+	}
+	var acquired int
+	err := s.db.QueryRowContext(ctx, `INSERT INTO sense.operational_leases(
+		lease_name, owner_id, fencing_token, lease_until, updated_at
+	) VALUES (
+		$1, $2, $3,
+		clock_timestamp() + ($4 * interval '1 second'), clock_timestamp()
+	)
+	ON CONFLICT (lease_name) DO UPDATE SET
+		owner_id = EXCLUDED.owner_id,
+		fencing_token = EXCLUDED.fencing_token,
+		lease_until = EXCLUDED.lease_until,
+		updated_at = EXCLUDED.updated_at
+	WHERE sense.operational_leases.lease_until <= clock_timestamp()
+		OR (sense.operational_leases.owner_id = $2
+			AND sense.operational_leases.fencing_token = $3)
+	RETURNING 1`, name, owner, token, duration.Seconds()).Scan(&acquired)
+	if errors.Is(err, sql.ErrNoRows) {
+		return false, nil
+	}
+	if err != nil {
+		return false, errors.New("acquire postgres operational lease")
+	}
+	return acquired == 1, nil
+}
+
+func (s *Postgres) ReleaseOperationalLease(
+	ctx context.Context,
+	name, owner, token string,
+	_ time.Time,
+) error {
+	_, err := s.db.ExecContext(ctx, `UPDATE sense.operational_leases
+		SET lease_until = clock_timestamp(), updated_at = clock_timestamp()
+		WHERE lease_name = $1 AND owner_id = $2 AND fencing_token = $3`,
+		name, owner, token)
+	if err != nil {
+		return errors.New("release postgres operational lease")
+	}
+	return nil
+}
+
+func (s *Postgres) ListMediaPathOwnership(ctx context.Context) ([]MediaPathOwnership, error) {
+	rows, err := s.db.QueryContext(ctx, `SELECT
+		o.path_name, o.device_id,
+		EXISTS (
+			SELECT 1 FROM sense.devices d
+			WHERE d.id = o.device_id AND d.path_name = o.path_name
+				AND EXISTS (
+					SELECT 1 FROM sense.device_capabilities c
+					WHERE c.device_id = d.id AND c.capability = 'video_capture'
+				)
+		) AS current_claim
+		FROM sense.media_path_ownership o
+		ORDER BY o.path_name`)
+	if err != nil {
+		return nil, errors.New("list postgres MediaMTX path ownership")
+	}
+	defer rows.Close()
+	values := make([]MediaPathOwnership, 0)
+	for rows.Next() {
+		var value MediaPathOwnership
+		if err := rows.Scan(&value.PathName, &value.DeviceID, &value.CurrentClaim); err != nil {
+			return nil, errors.New("scan postgres MediaMTX path ownership")
+		}
+		values = append(values, value)
+	}
+	if err := rows.Err(); err != nil {
+		return nil, errors.New("iterate postgres MediaMTX path ownership")
+	}
+	return values, nil
+}
+
+func (s *Postgres) SaveOrphanScan(
+	ctx context.Context,
+	scan OrphanScan,
+	owner, token string,
+) error {
+	if err := validateOrphanScan(scan); err != nil {
+		return err
+	}
+	tx, err := s.db.BeginTx(ctx, nil)
+	if err != nil {
+		return errors.New("begin postgres orphan scan save")
+	}
+	defer tx.Rollback()
+	var lease int
+	err = tx.QueryRowContext(ctx, `SELECT 1 FROM sense.operational_leases
+		WHERE lease_name = $1 AND owner_id = $2 AND fencing_token = $3
+			AND lease_until > clock_timestamp()
+		FOR UPDATE`, OperationalLeaseOrphanScan, owner, token).Scan(&lease)
+	if errors.Is(err, sql.ErrNoRows) {
+		return ErrOperationalLeaseLost
+	}
+	if err != nil {
+		return errors.New("verify postgres orphan scan lease")
+	}
+	if _, err := tx.ExecContext(ctx, `INSERT INTO sense.orphan_scan_runs(
+		id, instance_id, observed_count, owned_stale_count, unowned_count,
+		safety_allowed, safety_reason, completed_at, expires_at
+	) VALUES ($1,$2,$3,$4,$5,$6,$7,$8,$9)`,
+		scan.ID, scan.InstanceID, scan.ObservedCount, scan.OwnedStaleCount, scan.UnownedCount,
+		scan.SafetyAllowed, scan.SafetyReason, scan.CompletedAt, scan.ExpiresAt,
+	); err != nil {
+		return errors.New("insert postgres orphan scan")
+	}
+	for _, finding := range scan.Findings {
+		var deviceID any
+		if finding.DeviceID != "" {
+			deviceID = finding.DeviceID
+		}
+		if _, err := tx.ExecContext(ctx, `INSERT INTO sense.orphan_scan_findings(
+			scan_id, path_name, classification, device_id
+		) VALUES ($1,$2,$3,$4)`, scan.ID, finding.PathName, finding.Classification, deviceID); err != nil {
+			return errors.New("insert postgres orphan scan finding")
+		}
+	}
+	if _, err := tx.ExecContext(ctx, `UPDATE sense.operational_leases
+		SET lease_until = clock_timestamp(), updated_at = clock_timestamp()
+		WHERE lease_name = $1 AND owner_id = $2 AND fencing_token = $3`,
+		OperationalLeaseOrphanScan, owner, token); err != nil {
+		return errors.New("release postgres orphan scan lease")
+	}
+	if _, err := tx.ExecContext(ctx, `DELETE FROM sense.orphan_scan_runs r
+		WHERE r.completed_at < $1
+			AND NOT EXISTS (
+				SELECT 1 FROM sense.orphan_cleanup_actions a WHERE a.scan_id = r.id
+			)`, scan.CompletedAt.Add(-orphanReportRetention)); err != nil {
+		return errors.New("expire postgres orphan scan reports")
+	}
+	if err := tx.Commit(); err != nil {
+		return errors.New("commit postgres orphan scan")
+	}
+	return nil
+}
+
+func validateOrphanScan(scan OrphanScan) error {
+	if strings.TrimSpace(scan.ID) == "" || strings.TrimSpace(scan.InstanceID) == "" ||
+		scan.ObservedCount < 0 || scan.OwnedStaleCount < 0 || scan.UnownedCount < 0 ||
+		scan.OwnedStaleCount+scan.UnownedCount > scan.ObservedCount ||
+		strings.TrimSpace(scan.SafetyReason) == "" || !scan.ExpiresAt.After(scan.CompletedAt) {
+		return errors.New("invalid orphan scan")
+	}
+	seen := make(map[string]struct{}, len(scan.Findings))
+	staleCount, unownedCount := 0, 0
+	for _, finding := range scan.Findings {
+		if strings.TrimSpace(finding.PathName) == "" {
+			return errors.New("invalid orphan finding path")
+		}
+		if _, duplicate := seen[finding.PathName]; duplicate {
+			return errors.New("duplicate orphan finding path")
+		}
+		seen[finding.PathName] = struct{}{}
+		switch finding.Classification {
+		case OrphanOwnedStale:
+			staleCount++
+			if strings.TrimSpace(finding.DeviceID) == "" {
+				return errors.New("owned stale finding lacks device")
+			}
+		case OrphanUnowned:
+			unownedCount++
+			if finding.DeviceID != "" {
+				return errors.New("unowned finding has device")
+			}
+		default:
+			return fmt.Errorf("invalid orphan finding classification %q", finding.Classification)
+		}
+	}
+	if staleCount != scan.OwnedStaleCount || unownedCount != scan.UnownedCount {
+		return errors.New("orphan scan counts do not match findings")
+	}
+	return nil
+}
+
+func (s *Postgres) GetOrphanScan(ctx context.Context, id string) (OrphanScan, error) {
+	var scan OrphanScan
+	err := s.db.QueryRowContext(ctx, `SELECT
+		id, instance_id, observed_count, owned_stale_count, unowned_count,
+		safety_allowed, safety_reason, completed_at, expires_at
+		FROM sense.orphan_scan_runs WHERE id = $1`, id).Scan(
+		&scan.ID, &scan.InstanceID, &scan.ObservedCount, &scan.OwnedStaleCount,
+		&scan.UnownedCount, &scan.SafetyAllowed, &scan.SafetyReason,
+		&scan.CompletedAt, &scan.ExpiresAt,
+	)
+	if errors.Is(err, sql.ErrNoRows) {
+		return OrphanScan{}, ErrOrphanScanNotFound
+	}
+	if err != nil {
+		return OrphanScan{}, errors.New("read postgres orphan scan")
+	}
+	rows, err := s.db.QueryContext(ctx, `SELECT
+		f.path_name, f.classification, COALESCE(f.device_id, ''),
+		COALESCE(a.status = 'deleted', false)
+		FROM sense.orphan_scan_findings f
+		LEFT JOIN sense.orphan_cleanup_actions a
+			ON a.scan_id = f.scan_id AND a.path_name = f.path_name
+		WHERE f.scan_id = $1 ORDER BY f.path_name`, id)
+	if err != nil {
+		return OrphanScan{}, errors.New("list postgres orphan scan findings")
+	}
+	defer rows.Close()
+	scan.Findings = make([]OrphanFinding, 0)
+	for rows.Next() {
+		var finding OrphanFinding
+		if err := rows.Scan(
+			&finding.PathName, &finding.Classification, &finding.DeviceID, &finding.Deleted,
+		); err != nil {
+			return OrphanScan{}, errors.New("scan postgres orphan finding")
+		}
+		scan.Findings = append(scan.Findings, finding)
+	}
+	if err := rows.Err(); err != nil {
+		return OrphanScan{}, errors.New("iterate postgres orphan findings")
+	}
+	return scan, nil
+}
+
+func (s *Postgres) RecordOrphanCleanup(
+	ctx context.Context,
+	scanID, pathName, actorID, status, errorCode string,
+	now time.Time,
+) error {
+	if status != "deleted" && status != "failed" {
+		return errors.New("invalid orphan cleanup status")
+	}
+	var storedError any
+	if status == "failed" {
+		if strings.TrimSpace(errorCode) == "" {
+			return errors.New("failed orphan cleanup requires an error code")
+		}
+		storedError = errorCode
+	}
+	_, err := s.db.ExecContext(ctx, `INSERT INTO sense.orphan_cleanup_actions(
+		scan_id, path_name, classification, actor_id, status, error_code, attempted_at
+	) VALUES ($1,$2,'owned_stale',$3,$4,$5,$6)
+	ON CONFLICT (scan_id, path_name) DO UPDATE SET
+		actor_id = EXCLUDED.actor_id,
+		status = EXCLUDED.status,
+		error_code = EXCLUDED.error_code,
+		attempted_at = EXCLUDED.attempted_at
+	WHERE sense.orphan_cleanup_actions.status <> 'deleted'`,
+		scanID, pathName, actorID, status, storedError, now)
+	if err != nil {
+		return errors.New("record postgres orphan cleanup result")
+	}
+	return nil
+}
@@ -53,8 +53,8 @@ func (s *Postgres) Close() error {
 func (s *Postgres) verifySchemaAndPrivileges(ctx context.Context) error {
 	var version sql.NullInt64
 	if err := s.db.QueryRowContext(ctx,
-		`SELECT MAX(version) FROM sense.schema_migrations`).Scan(&version); err != nil || !version.Valid || version.Int64 < 4 {
-		return errors.New("postgres sense schema migration v4 is required")
+		`SELECT MAX(version) FROM sense.schema_migrations`).Scan(&version); err != nil || !version.Valid || version.Int64 < 5 {
+		return errors.New("postgres sense schema migration v5 is required")
 	}
 	var canReadQuotaView, canWriteQuotaView, canReadSiteSource, canWriteSiteSource bool
 	var canReadAreaView, canWriteAreaView, canReadAreaSource, canWriteAreaSource bool
@@ -95,6 +95,28 @@ func (s *Postgres) verifySchemaAndPrivileges(ctx context.Context) error {
 		publicReceipts || publicOperations || publicOperationItems {
 		return errors.New("postgres role violates Control API state privilege boundary")
 	}
+	var canUseOwnership, canUseLeases, canUseScans, canUseFindings, canUseActions bool
+	var publicOwnership, publicLeases, publicScans, publicFindings, publicActions bool
+	if err := s.db.QueryRowContext(ctx, `SELECT
+		has_table_privilege(current_user, 'sense.media_path_ownership', 'SELECT,INSERT,UPDATE,DELETE'),
+		has_table_privilege(current_user, 'sense.operational_leases', 'SELECT,INSERT,UPDATE,DELETE'),
+		has_table_privilege(current_user, 'sense.orphan_scan_runs', 'SELECT,INSERT,UPDATE,DELETE'),
+		has_table_privilege(current_user, 'sense.orphan_scan_findings', 'SELECT,INSERT,UPDATE,DELETE'),
+		has_table_privilege(current_user, 'sense.orphan_cleanup_actions', 'SELECT,INSERT,UPDATE,DELETE'),
+		has_table_privilege('public', 'sense.media_path_ownership', 'SELECT,INSERT,UPDATE,DELETE'),
+		has_table_privilege('public', 'sense.operational_leases', 'SELECT,INSERT,UPDATE,DELETE'),
+		has_table_privilege('public', 'sense.orphan_scan_runs', 'SELECT,INSERT,UPDATE,DELETE'),
+		has_table_privilege('public', 'sense.orphan_scan_findings', 'SELECT,INSERT,UPDATE,DELETE'),
+		has_table_privilege('public', 'sense.orphan_cleanup_actions', 'SELECT,INSERT,UPDATE,DELETE')`).Scan(
+		&canUseOwnership, &canUseLeases, &canUseScans, &canUseFindings, &canUseActions,
+		&publicOwnership, &publicLeases, &publicScans, &publicFindings, &publicActions,
+	); err != nil {
+		return errors.New("verify postgres reconciliation safety privileges")
+	}
+	if !canUseOwnership || !canUseLeases || !canUseScans || !canUseFindings || !canUseActions ||
+		publicOwnership || publicLeases || publicScans || publicFindings || publicActions {
+		return errors.New("postgres role violates reconciliation safety privilege boundary")
+	}
 	return nil
 }

@@ -391,7 +413,8 @@ func (s *Postgres) SetDesiredState(ctx context.Context, id string, desired devic
 		return errors.New("update postgres desired state")
 	}
 	if _, err := tx.ExecContext(ctx, `UPDATE sense.reconcile_state SET
-		failure_count = 0, next_attempt_at = NULL, last_error_code = NULL, updated_at = $1
+		failure_count = 0, next_attempt_at = NULL, last_error_code = NULL,
+		lease_owner = NULL, lease_token = NULL, lease_until = NULL, updated_at = $1
 		WHERE device_id = $2`, now, id); err != nil {
 		return errors.New("reset postgres reconcile state")
 	}
@@ -486,6 +509,217 @@ func (s *Postgres) ListDueReconcile(ctx context.Context, now time.Time, limit in
 	return values, nil
 }

+func (s *Postgres) ClaimDueReconcile(
+	ctx context.Context,
+	claim ReconcileClaim,
+) ([]ReconcileCandidate, error) {
+	if claim.Limit <= 0 {
+		return nil, nil
+	}
+	if strings.TrimSpace(claim.Owner) == "" || strings.TrimSpace(claim.Token) == "" ||
+		claim.LeaseDuration <= 0 {
+		return nil, errors.New("invalid postgres reconcile claim")
+	}
+	tx, err := s.db.BeginTx(ctx, nil)
+	if err != nil {
+		return nil, errors.New("begin postgres reconcile claim")
+	}
+	defer tx.Rollback()
+	rows, err := tx.QueryContext(ctx, `SELECT `+postgresDeviceColumns+`, r.failure_count, r.next_attempt_at
+		FROM sense.devices d
+		JOIN sense.reconcile_state r ON r.device_id = d.id
+		WHERE EXISTS (SELECT 1 FROM sense.device_capabilities c
+				WHERE c.device_id = d.id AND c.capability = 'video_capture')
+			AND (r.observed_generation < d.generation
+				OR (d.desired_state = 'enabled' AND r.failure_count > 0))
+			AND (r.next_attempt_at IS NULL OR r.next_attempt_at <= clock_timestamp())
+			AND (r.lease_until IS NULL OR r.lease_until <= clock_timestamp())
+		ORDER BY d.updated_at, d.id
+		FOR UPDATE OF r SKIP LOCKED
+		LIMIT $1`, claim.Limit)
+	if err != nil {
+		return nil, errors.New("select postgres reconcile claims")
+	}
+	values := make([]ReconcileCandidate, 0)
+	for rows.Next() {
+		var candidate ReconcileCandidate
+		var areaID sql.NullString
+		var quotaVersion, areaVersion sql.NullInt64
+		var nextAttempt sql.NullTime
+		if err := rows.Scan(
+			&candidate.Device.ID, &candidate.Device.TenantID, &candidate.Device.SiteID,
+			&areaID, &candidate.Device.SerialNumber, &candidate.Device.Name, &candidate.Device.Modality,
+			&candidate.Device.DesiredState, &candidate.Device.ActualState,
+			&candidate.Device.EndpointRef, &candidate.Device.CredentialRef,
+			&candidate.Device.ProfileToken, &candidate.Device.PathName,
+			&candidate.Device.Generation, &candidate.Device.ResourceVersion,
+			&quotaVersion, &areaVersion,
+			&candidate.Device.CreatedAt, &candidate.Device.UpdatedAt,
+			&candidate.FailureCount, &nextAttempt,
+		); err != nil {
+			rows.Close()
+			return nil, errors.New("scan postgres reconcile claim")
+		}
+		candidate.Device.AreaID = areaID.String
+		candidate.Device.QuotaSourceVersion = quotaVersion.Int64
+		candidate.Device.AreaPolicySourceVersion = areaVersion.Int64
+		if nextAttempt.Valid {
+			value := nextAttempt.Time
+			candidate.NextAttempt = &value
+		}
+		values = append(values, candidate)
+	}
+	if err := rows.Close(); err != nil {
+		return nil, errors.New("close postgres reconcile claims")
+	}
+	if err := rows.Err(); err != nil {
+		return nil, errors.New("iterate postgres reconcile claims")
+	}
+	for _, candidate := range values {
+		if _, err := tx.ExecContext(ctx, `UPDATE sense.reconcile_state SET
+			lease_owner = $1, lease_token = $2,
+			lease_until = clock_timestamp() + ($3 * interval '1 second'),
+			updated_at = clock_timestamp()
+			WHERE device_id = $4`,
+			claim.Owner, claim.Token, claim.LeaseDuration.Seconds(), candidate.Device.ID,
+		); err != nil {
+			return nil, errors.New("persist postgres reconcile claim")
+		}
+	}
+	if err := tx.Commit(); err != nil {
+		return nil, errors.New("commit postgres reconcile claim")
+	}
+	for index := range values {
+		values[index].Device.Capabilities, err = s.capabilities(ctx, values[index].Device.ID)
+		if err != nil {
+			return nil, err
+		}
+	}
+	return values, nil
+}
+
+func (s *Postgres) RenewReconcileLease(
+	ctx context.Context,
+	id, owner, token string,
+	now time.Time,
+	duration time.Duration,
+) (bool, error) {
+	if strings.TrimSpace(owner) == "" || strings.TrimSpace(token) == "" || duration <= 0 {
+		return false, errors.New("invalid postgres reconcile lease renewal")
+	}
+	result, err := s.db.ExecContext(ctx, `UPDATE sense.reconcile_state SET
+		lease_until = clock_timestamp() + ($1 * interval '1 second'),
+		updated_at = clock_timestamp()
+		WHERE device_id = $2 AND lease_owner = $3 AND lease_token = $4
+			AND lease_until > clock_timestamp()`, duration.Seconds(), id, owner, token)
+	if err != nil {
+		return false, errors.New("renew postgres reconcile lease")
+	}
+	affected, err := result.RowsAffected()
+	if err != nil {
+		return false, errors.New("read postgres reconcile lease renewal")
+	}
+	return affected == 1, nil
+}
+
+func (s *Postgres) CompleteReconcile(
+	ctx context.Context,
+	id string,
+	generation int64,
+	owner, token string,
+	now time.Time,
+) error {
+	tx, err := s.db.BeginTx(ctx, nil)
+	if err != nil {
+		return errors.New("begin postgres fenced reconciliation completion")
+	}
+	defer tx.Rollback()
+	result, err := tx.ExecContext(ctx, `UPDATE sense.reconcile_state SET
+		failure_count = 0, next_attempt_at = NULL, last_error_code = NULL,
+		observed_generation = $1, lease_owner = NULL, lease_token = NULL,
+		lease_until = NULL, updated_at = $2
+		WHERE device_id = $3 AND lease_owner = $4 AND lease_token = $5
+			AND lease_until > clock_timestamp()`, generation, now, id, owner, token)
+	if err != nil {
+		return errors.New("complete postgres fenced reconciliation")
+	}
+	affected, err := result.RowsAffected()
+	if err != nil {
+		return errors.New("read postgres fenced reconciliation completion")
+	}
+	if affected != 1 {
+		return ErrReconcileLeaseLost
+	}
+	if _, err := tx.ExecContext(ctx, `INSERT INTO sense.media_path_ownership(
+		path_name, device_id, tenant_id, site_id, first_claimed_at, last_confirmed_at
+	)
+	SELECT path_name, id, tenant_id, site_id, $1, $1
+	FROM sense.devices
+	WHERE id = $2 AND desired_state = 'enabled' AND btrim(path_name) <> ''
+	ON CONFLICT (path_name) DO UPDATE SET
+		device_id = EXCLUDED.device_id,
+		tenant_id = EXCLUDED.tenant_id,
+		site_id = EXCLUDED.site_id,
+		last_confirmed_at = EXCLUDED.last_confirmed_at`, now, id); err != nil {
+		return errors.New("record postgres MediaMTX path ownership")
+	}
+	result, err = tx.ExecContext(ctx, `UPDATE sense.devices
+		SET actual_state = CASE WHEN desired_state = 'disabled' THEN 'offline' ELSE 'pending' END,
+		updated_at = $1 WHERE id = $2`, now, id)
+	if err != nil {
+		return errors.New("mark postgres fenced device state")
+	}
+	if affected, _ := result.RowsAffected(); affected != 1 {
+		return ErrNotFound
+	}
+	if err := tx.Commit(); err != nil {
+		return errors.New("commit postgres fenced reconciliation completion")
+	}
+	return nil
+}
+
+func (s *Postgres) FailReconcile(
+	ctx context.Context,
+	id string,
+	failureCount int,
+	nextAttempt time.Time,
+	errorCode, owner, token string,
+	now time.Time,
+) error {
+	tx, err := s.db.BeginTx(ctx, nil)
+	if err != nil {
+		return errors.New("begin postgres fenced reconciliation failure")
+	}
+	defer tx.Rollback()
+	result, err := tx.ExecContext(ctx, `UPDATE sense.reconcile_state SET
+		failure_count = $1, next_attempt_at = $2, last_error_code = $3,
+		lease_owner = NULL, lease_token = NULL, lease_until = NULL, updated_at = $4
+		WHERE device_id = $5 AND lease_owner = $6 AND lease_token = $7
+			AND lease_until > clock_timestamp()`, failureCount, nextAttempt, errorCode, now, id, owner, token)
+	if err != nil {
+		return errors.New("persist postgres fenced reconciliation failure")
+	}
+	affected, err := result.RowsAffected()
+	if err != nil {
+		return errors.New("read postgres fenced reconciliation failure")
+	}
+	if affected != 1 {
+		return ErrReconcileLeaseLost
+	}
+	result, err = tx.ExecContext(ctx, `UPDATE sense.devices
+		SET actual_state = 'failed', updated_at = $1 WHERE id = $2`, now, id)
+	if err != nil {
+		return errors.New("mark postgres fenced failed device")
+	}
+	if affected, _ := result.RowsAffected(); affected != 1 {
+		return ErrNotFound
+	}
+	if err := tx.Commit(); err != nil {
+		return errors.New("commit postgres fenced reconciliation failure")
+	}
+	return nil
+}
+
 func (s *Postgres) ListEnabledVideoDevices(ctx context.Context, limit int) ([]device.Device, error) {
 	if limit <= 0 {
 		return nil, nil
@@ -287,12 +287,12 @@ func TestPostgresOpenRejectsOverprivilegedRuntimeRole(t *testing.T) {
 	_, admin := openPostgresTestStore(t)
 	ctx := context.Background()
 	if _, err := admin.ExecContext(ctx,
-		`GRANT UPDATE ON bell.site_quota_v1 TO yovision_t011_sense`); err != nil {
+		`GRANT UPDATE ON bell.site_quota_v1 TO yovision_t012_sense`); err != nil {
 		t.Fatal(err)
 	}
 	defer func() {
 		_, _ = admin.ExecContext(context.Background(),
-			`REVOKE UPDATE ON bell.site_quota_v1 FROM yovision_t011_sense`)
+			`REVOKE UPDATE ON bell.site_quota_v1 FROM yovision_t012_sense`)
 	}()
 	value, err := OpenPostgres(ctx, os.Getenv(postgresTestDSNEnv))
 	if value != nil {
@@ -304,6 +304,27 @@ func TestPostgresOpenRejectsOverprivilegedRuntimeRole(t *testing.T) {
 	}
 }

+func TestPostgresOpenRejectsPublicReconciliationStatePrivilege(t *testing.T) {
+	_, admin := openPostgresTestStore(t)
+	ctx := context.Background()
+	if _, err := admin.ExecContext(ctx,
+		`GRANT SELECT ON sense.orphan_scan_runs TO PUBLIC`); err != nil {
+		t.Fatal(err)
+	}
+	defer func() {
+		_, _ = admin.ExecContext(context.Background(),
+			`REVOKE SELECT ON sense.orphan_scan_runs FROM PUBLIC`)
+	}()
+	value, err := OpenPostgres(ctx, os.Getenv(postgresTestDSNEnv))
+	if value != nil {
+		_ = value.Close()
+		t.Fatal("PUBLIC reconciliation state privilege was accepted")
+	}
+	if err == nil || !strings.Contains(err.Error(), "reconciliation safety privilege boundary") {
+		t.Fatalf("expected reconciliation privilege-boundary error, got %v", err)
+	}
+}
+
 func TestPostgresAreaPolicyAllowsNonImagingAndDeniesImagingCreate(t *testing.T) {
 	store, admin := openPostgresTestStore(t)
 	ctx := context.Background()
@@ -597,12 +618,12 @@ func TestPostgresOpenRejectsAreaSourcePrivilege(t *testing.T) {
 	_, admin := openPostgresTestStore(t)
 	ctx := context.Background()
 	if _, err := admin.ExecContext(ctx,
-		`GRANT SELECT ON bell.areas TO yovision_t011_sense`); err != nil {
+		`GRANT SELECT ON bell.areas TO yovision_t012_sense`); err != nil {
 		t.Fatal(err)
 	}
 	defer func() {
 		_, _ = admin.ExecContext(context.Background(),
-			`REVOKE SELECT ON bell.areas FROM yovision_t011_sense`)
+			`REVOKE SELECT ON bell.areas FROM yovision_t012_sense`)
 	}()
 	value, err := OpenPostgres(ctx, os.Getenv(postgresTestDSNEnv))
 	if value != nil {
@@ -969,6 +990,180 @@ func TestPostgresConcurrentControlBatchesUseStableDeviceLockOrder(t *testing.T)
 	}
 }

+func TestPostgresConcurrentReconcileClaimHasOneWinner(t *testing.T) {
+	first, admin := openPostgresTestStore(t)
+	insertBellSite(t, admin, "tenant", "site", 2)
+	if err := first.CreateDevice(context.Background(), videoDevice(1, "tenant", "site")); err != nil {
+		t.Fatal(err)
+	}
+	second, err := OpenPostgres(context.Background(), os.Getenv(postgresTestDSNEnv))
+	if err != nil {
+		t.Fatal(err)
+	}
+	t.Cleanup(func() { _ = second.Close() })
+	now := time.Date(2026, 8, 7, 0, 0, 0, 0, time.UTC)
+	start := make(chan struct{})
+	counts := make(chan int, 2)
+	errorsFound := make(chan error, 2)
+	var wait sync.WaitGroup
+	for index, repository := range []*Postgres{first, second} {
+		wait.Add(1)
+		go func(index int, repository *Postgres) {
+			defer wait.Done()
+			<-start
+			values, err := repository.ClaimDueReconcile(context.Background(), ReconcileClaim{
+				Owner: fmt.Sprintf("ins-%d", index), Token: fmt.Sprintf("token-%d", index),
+				Now: now, LeaseDuration: 30 * time.Second, Limit: 1,
+			})
+			if err != nil {
+				errorsFound <- err
+				return
+			}
+			counts <- len(values)
+		}(index, repository)
+	}
+	close(start)
+	wait.Wait()
+	close(counts)
+	close(errorsFound)
+	for err := range errorsFound {
+		t.Fatal(err)
+	}
+	total, winners := 0, 0
+	for count := range counts {
+		total += count
+		if count == 1 {
+			winners++
+		}
+	}
+	if total != 1 || winners != 1 {
+		t.Fatalf("due row was not exclusively claimed: total=%d winners=%d", total, winners)
+	}
+}
+
+func TestPostgresExpiredReconcileLeaseFencesOldWorkerAndRecordsOwnership(t *testing.T) {
+	postgres, admin := openPostgresTestStore(t)
+	insertBellSite(t, admin, "tenant", "site", 2)
+	value := videoDevice(1, "tenant", "site")
+	if err := postgres.CreateDevice(context.Background(), value); err != nil {
+		t.Fatal(err)
+	}
+	now := time.Date(2026, 8, 7, 0, 0, 0, 0, time.UTC)
+	first, err := postgres.ClaimDueReconcile(context.Background(), ReconcileClaim{
+		Owner: "ins-a", Token: "token-a", Now: now, LeaseDuration: 30 * time.Second, Limit: 1,
+	})
+	if err != nil || len(first) != 1 {
+		t.Fatalf("first claim failed: %+v %v", first, err)
+	}
+	early, err := postgres.ClaimDueReconcile(context.Background(), ReconcileClaim{
+		Owner: "ins-b", Token: "token-b", Now: now.Add(10 * time.Second),
+		LeaseDuration: 30 * time.Second, Limit: 1,
+	})
+	if err != nil || len(early) != 0 {
+		t.Fatalf("live lease was stolen: %+v %v", early, err)
+	}
+	if _, err := admin.Exec(`UPDATE sense.reconcile_state
+		SET lease_until = clock_timestamp() - interval '1 second'
+		WHERE device_id = $1`, value.ID); err != nil {
+		t.Fatal(err)
+	}
+	second, err := postgres.ClaimDueReconcile(context.Background(), ReconcileClaim{
+		Owner: "ins-b", Token: "token-b", Now: now.Add(31 * time.Second),
+		LeaseDuration: 30 * time.Second, Limit: 1,
+	})
+	if err != nil || len(second) != 1 {
+		t.Fatalf("expired lease was not recoverable: %+v %v", second, err)
+	}
+	if err := postgres.CompleteReconcile(
+		context.Background(), value.ID, value.Generation, "ins-a", "token-a", now.Add(32*time.Second),
+	); !errors.Is(err, ErrReconcileLeaseLost) {
+		t.Fatalf("old worker was not fenced: %v", err)
+	}
+	if err := postgres.CompleteReconcile(
+		context.Background(), value.ID, value.Generation, "ins-b", "token-b", now.Add(32*time.Second),
+	); err != nil {
+		t.Fatal(err)
+	}
+	var ownershipDevice string
+	if err := admin.QueryRow(`SELECT device_id FROM sense.media_path_ownership WHERE path_name = $1`,
+		value.PathName).Scan(&ownershipDevice); err != nil {
+		t.Fatal(err)
+	}
+	if ownershipDevice != value.ID {
+		t.Fatalf("wrong ownership was recorded: %q", ownershipDevice)
+	}
+	var leaseToken sql.NullString
+	if err := admin.QueryRow(`SELECT lease_token FROM sense.reconcile_state WHERE device_id = $1`,
+		value.ID).Scan(&leaseToken); err != nil {
+		t.Fatal(err)
+	}
+	if leaseToken.Valid {
+		t.Fatal("completion did not release the reconcile lease")
+	}
+}
+
+func TestPostgresOrphanReportLeaseAndCleanupAuditAreFencedAndIdempotent(t *testing.T) {
+	postgres, _ := openPostgresTestStore(t)
+	ctx := context.Background()
+	now := time.Date(2026, 8, 7, 0, 0, 0, 0, time.UTC)
+	acquired, err := postgres.AcquireOperationalLease(
+		ctx, OperationalLeaseOrphanScan, "ins-a", "scan-token-a", now, 30*time.Second,
+	)
+	if err != nil || !acquired {
+		t.Fatalf("scan lease failed: %v %v", acquired, err)
+	}
+	scan := OrphanScan{
+		ID: "scan_" + strings.Repeat("0", 26), InstanceID: "ins-a",
+		ObservedCount: 10, OwnedStaleCount: 1, UnownedCount: 1,
+		SafetyAllowed: true, SafetyReason: "allowed",
+		CompletedAt: now.Add(time.Second), ExpiresAt: now.Add(15 * time.Minute),
+		Findings: []OrphanFinding{
+			{PathName: "stale", Classification: OrphanOwnedStale, DeviceID: "old-device"},
+			{PathName: "unknown", Classification: OrphanUnowned},
+		},
+	}
+	if err := postgres.SaveOrphanScan(ctx, scan, "ins-a", "scan-token-a"); err != nil {
+		t.Fatal(err)
+	}
+	loaded, err := postgres.GetOrphanScan(ctx, scan.ID)
+	if err != nil || len(loaded.Findings) != 2 || !loaded.SafetyAllowed {
+		t.Fatalf("stored scan mismatch: %+v %v", loaded, err)
+	}
+	if err := postgres.RecordOrphanCleanup(
+		ctx, scan.ID, "stale", "operator", "deleted", "", now.Add(2*time.Second),
+	); err != nil {
+		t.Fatal(err)
+	}
+	if err := postgres.RecordOrphanCleanup(
+		ctx, scan.ID, "unknown", "operator", "deleted", "", now.Add(2*time.Second),
+	); err == nil {
+		t.Fatal("unowned path accepted a cleanup audit record")
+	}
+	if err := postgres.RecordOrphanCleanup(
+		ctx, scan.ID, "stale", "operator-2", "failed", "media_error", now.Add(3*time.Second),
+	); err != nil {
+		t.Fatal(err)
+	}
+	loaded, err = postgres.GetOrphanScan(ctx, scan.ID)
+	if err != nil || !loaded.Findings[0].Deleted {
+		t.Fatalf("successful cleanup was downgraded: %+v %v", loaded, err)
+	}
+
+	acquired, err = postgres.AcquireOperationalLease(
+		ctx, OperationalLeaseOrphanScan, "ins-b", "scan-token-b", now.Add(31*time.Second), 30*time.Second,
+	)
+	if err != nil || !acquired {
+		t.Fatalf("expired scan lease was not recoverable: %v %v", acquired, err)
+	}
+	staleScan := scan
+	staleScan.ID = "scan_" + strings.Repeat("1", 26)
+	staleScan.CompletedAt = now.Add(32 * time.Second)
+	staleScan.ExpiresAt = staleScan.CompletedAt.Add(15 * time.Minute)
+	if err := postgres.SaveOrphanScan(ctx, staleScan, "ins-a", "scan-token-a"); !errors.Is(err, ErrOperationalLeaseLost) {
+		t.Fatalf("stale scan worker was not fenced: %v", err)
+	}
+}
+
 func openPostgresTestStore(t *testing.T) (*Postgres, *sql.DB) {
 	t.Helper()
 	dsn := os.Getenv(postgresTestDSNEnv)
@@ -985,6 +1180,11 @@ func openPostgresTestStore(t *testing.T) (*Postgres, *sql.DB) {
 		t.Fatal("connect PostgreSQL test administrator")
 	}
 	if _, err := admin.ExecContext(context.Background(), `TRUNCATE
+		sense.orphan_cleanup_actions,
+		sense.orphan_scan_findings,
+		sense.orphan_scan_runs,
+		sense.operational_leases,
+		sense.media_path_ownership,
 		sense.control_idempotency_receipts,
 		sense.batch_operation_items,
 		sense.batch_operations,
@@ -21,6 +21,8 @@ var (
 	ErrAreaPolicyUnavailable      = errors.New("area policy unavailable")
 	ErrAreaPolicyInvalid          = errors.New("area policy invalid")
 	ErrAreaPolicyDenied           = errors.New("area policy denies imaging device")
+	ErrReconcileLeaseLost         = errors.New("reconcile lease lost")
+	ErrOperationalLeaseLost       = errors.New("operational lease lost")
 )

 // Repository is the storage boundary used by the Sense process. SQLite stays
@@ -31,6 +33,10 @@ type Repository interface {
 	SetDesiredState(context.Context, string, device.DesiredState) error
 	GetDevice(context.Context, string) (device.Device, error)
 	ListDueReconcile(context.Context, time.Time, int) ([]ReconcileCandidate, error)
+	ClaimDueReconcile(context.Context, ReconcileClaim) ([]ReconcileCandidate, error)
+	RenewReconcileLease(context.Context, string, string, string, time.Time, time.Duration) (bool, error)
+	CompleteReconcile(context.Context, string, int64, string, string, time.Time) error
+	FailReconcile(context.Context, string, int, time.Time, string, string, string, time.Time) error
 	ListEnabledVideoDevices(context.Context, int) ([]device.Device, error)
 	MarkReconciled(context.Context, string, int64, time.Time) error
 	MarkReconcileFailure(context.Context, string, int, time.Time, string, time.Time) error
@@ -39,6 +45,17 @@ type Repository interface {
 	ConvergenceSnapshot(context.Context) (ConvergenceSnapshot, error)
 }

+// ReconcileClaim identifies one short-lived batch claim. Token is unique per
+// run and fences a worker whose lease expired and was acquired by another
+// process. SQLite accepts the shape but remains explicitly single-process.
+type ReconcileClaim struct {
+	Owner         string
+	Token         string
+	Now           time.Time
+	LeaseDuration time.Duration
+	Limit         int
+}
+
 func OpenRepository(ctx context.Context, driver, dsn string) (Repository, error) {
 	switch strings.ToLower(strings.TrimSpace(driver)) {
 	case "", DriverSQLite:
@@ -430,6 +430,24 @@ func (s *SQLite) ListDueReconcile(ctx context.Context, now time.Time, limit int)
 	return candidates, nil
 }

+func (s *SQLite) ClaimDueReconcile(ctx context.Context, claim ReconcileClaim) ([]ReconcileCandidate, error) {
+	// SQLite is retained for one-process M1 development. It deliberately does
+	// not claim cross-process leases; PostgreSQL is the production M2 boundary.
+	return s.ListDueReconcile(ctx, claim.Now, claim.Limit)
+}
+
+func (s *SQLite) RenewReconcileLease(
+	ctx context.Context,
+	_, _, _ string,
+	_ time.Time,
+	_ time.Duration,
+) (bool, error) {
+	if err := ctx.Err(); err != nil {
+		return false, err
+	}
+	return true, nil
+}
+
 func (s *SQLite) ListEnabledVideoDevices(ctx context.Context, limit int) ([]device.Device, error) {
 	if limit <= 0 {
 		return nil, nil
@@ -499,6 +517,16 @@ func (s *SQLite) MarkReconciled(ctx context.Context, id string, generation int64
 	return nil
 }

+func (s *SQLite) CompleteReconcile(
+	ctx context.Context,
+	id string,
+	generation int64,
+	_, _ string,
+	now time.Time,
+) error {
+	return s.MarkReconciled(ctx, id, generation, now)
+}
+
 func (s *SQLite) MarkReconcileFailure(ctx context.Context, id string, failureCount int, nextAttempt time.Time, errorCode string, now time.Time) error {
 	tx, err := s.db.BeginTx(ctx, nil)
 	if err != nil {
@@ -526,6 +554,17 @@ func (s *SQLite) MarkReconcileFailure(ctx context.Context, id string, failureCou
 	return nil
 }

+func (s *SQLite) FailReconcile(
+	ctx context.Context,
+	id string,
+	failureCount int,
+	nextAttempt time.Time,
+	errorCode, _, _ string,
+	now time.Time,
+) error {
+	return s.MarkReconcileFailure(ctx, id, failureCount, nextAttempt, errorCode, now)
+}
+
 func (s *SQLite) UpdateActualState(ctx context.Context, id string, state device.ActualState, now time.Time) error {
 	if state != device.ActualPending && state != device.ActualOnline && state != device.ActualOffline && state != device.ActualFailed {
 		return fmt.Errorf("invalid actual state %q", state)
@@ -0,0 +1,147 @@
+-- Add fenced reconciliation claims and durable, secret-free MediaMTX path
+-- ownership/orphan reports. Runtime migrations are installed out of process.
+
+ALTER TABLE sense.reconcile_state
+    ADD COLUMN IF NOT EXISTS lease_owner text,
+    ADD COLUMN IF NOT EXISTS lease_token text,
+    ADD COLUMN IF NOT EXISTS lease_until timestamptz;
+
+DO $constraints$
+BEGIN
+    IF NOT EXISTS (
+        SELECT 1 FROM pg_constraint
+        WHERE conrelid = 'sense.reconcile_state'::regclass
+          AND conname = 'sense_reconcile_lease_all_or_none'
+    ) THEN
+        ALTER TABLE sense.reconcile_state
+            ADD CONSTRAINT sense_reconcile_lease_all_or_none CHECK (
+                (lease_owner IS NULL AND lease_token IS NULL AND lease_until IS NULL)
+                OR (btrim(lease_owner) <> '' AND btrim(lease_token) <> '' AND lease_until IS NOT NULL)
+            );
+    END IF;
+END
+$constraints$;
+
+CREATE INDEX IF NOT EXISTS sense_reconcile_lease_due_idx
+    ON sense.reconcile_state(lease_until, next_attempt_at);
+
+CREATE TABLE IF NOT EXISTS sense.media_path_ownership (
+    path_name text PRIMARY KEY,
+    device_id text NOT NULL,
+    tenant_id text NOT NULL,
+    site_id text NOT NULL,
+    first_claimed_at timestamptz NOT NULL,
+    last_confirmed_at timestamptz NOT NULL,
+    CONSTRAINT sense_media_path_ownership_not_blank CHECK (
+        btrim(path_name) <> '' AND btrim(device_id) <> ''
+        AND btrim(tenant_id) <> '' AND btrim(site_id) <> ''
+    )
+);
+ALTER TABLE sense.media_path_ownership OWNER TO sense_app;
+
+-- Keep the historical row independent from device lifecycle so a later device
+-- deletion cannot erase the evidence needed to classify an owned stale path.
+INSERT INTO sense.media_path_ownership(
+    path_name, device_id, tenant_id, site_id, first_claimed_at, last_confirmed_at
+)
+SELECT d.path_name, d.id, d.tenant_id, d.site_id, d.created_at, d.updated_at
+FROM sense.devices d
+WHERE btrim(d.path_name) <> ''
+  AND EXISTS (
+      SELECT 1 FROM sense.device_capabilities c
+      WHERE c.device_id = d.id AND c.capability = 'video_capture'
+  )
+ON CONFLICT (path_name) DO UPDATE SET
+    device_id = EXCLUDED.device_id,
+    tenant_id = EXCLUDED.tenant_id,
+    site_id = EXCLUDED.site_id,
+    last_confirmed_at = GREATEST(
+        sense.media_path_ownership.last_confirmed_at,
+        EXCLUDED.last_confirmed_at
+    );
+
+CREATE TABLE IF NOT EXISTS sense.operational_leases (
+    lease_name text PRIMARY KEY,
+    owner_id text NOT NULL,
+    fencing_token text NOT NULL,
+    lease_until timestamptz NOT NULL,
+    updated_at timestamptz NOT NULL,
+    CONSTRAINT sense_operational_lease_not_blank CHECK (
+        btrim(lease_name) <> '' AND btrim(owner_id) <> '' AND btrim(fencing_token) <> ''
+    )
+);
+ALTER TABLE sense.operational_leases OWNER TO sense_app;
+
+CREATE TABLE IF NOT EXISTS sense.orphan_scan_runs (
+    id text PRIMARY KEY,
+    instance_id text NOT NULL,
+    observed_count integer NOT NULL,
+    owned_stale_count integer NOT NULL,
+    unowned_count integer NOT NULL,
+    safety_allowed boolean NOT NULL,
+    safety_reason text NOT NULL,
+    completed_at timestamptz NOT NULL,
+    expires_at timestamptz NOT NULL,
+    CONSTRAINT sense_orphan_scan_id CHECK (id ~ '^scan_[0-9A-HJKMNP-TV-Z]{26}$'),
+    CONSTRAINT sense_orphan_scan_counts CHECK (
+        observed_count >= 0 AND owned_stale_count >= 0 AND unowned_count >= 0
+        AND owned_stale_count + unowned_count <= observed_count
+    ),
+    CONSTRAINT sense_orphan_scan_not_blank CHECK (
+        btrim(instance_id) <> '' AND btrim(safety_reason) <> ''
+    ),
+    CONSTRAINT sense_orphan_scan_expiry CHECK (expires_at > completed_at)
+);
+ALTER TABLE sense.orphan_scan_runs OWNER TO sense_app;
+
+CREATE TABLE IF NOT EXISTS sense.orphan_scan_findings (
+    scan_id text NOT NULL REFERENCES sense.orphan_scan_runs(id) ON DELETE CASCADE,
+    path_name text NOT NULL,
+    classification text NOT NULL,
+    device_id text,
+    PRIMARY KEY (scan_id, path_name),
+    UNIQUE (scan_id, path_name, classification),
+    CONSTRAINT sense_orphan_finding_classification CHECK (
+        classification IN ('owned_stale', 'unowned')
+    ),
+    CONSTRAINT sense_orphan_finding_not_blank CHECK (
+        btrim(path_name) <> '' AND (device_id IS NULL OR btrim(device_id) <> '')
+    ),
+    CONSTRAINT sense_orphan_finding_owner_shape CHECK (
+        (classification = 'owned_stale' AND device_id IS NOT NULL)
+        OR (classification = 'unowned' AND device_id IS NULL)
+    )
+);
+ALTER TABLE sense.orphan_scan_findings OWNER TO sense_app;
+
+CREATE TABLE IF NOT EXISTS sense.orphan_cleanup_actions (
+    scan_id text NOT NULL REFERENCES sense.orphan_scan_runs(id),
+    path_name text NOT NULL,
+    classification text NOT NULL DEFAULT 'owned_stale',
+    actor_id text NOT NULL,
+    status text NOT NULL,
+    error_code text,
+    attempted_at timestamptz NOT NULL,
+    PRIMARY KEY (scan_id, path_name),
+    FOREIGN KEY (scan_id, path_name, classification)
+        REFERENCES sense.orphan_scan_findings(scan_id, path_name, classification),
+    CONSTRAINT sense_orphan_cleanup_owned_only CHECK (classification = 'owned_stale'),
+    CONSTRAINT sense_orphan_cleanup_status CHECK (status IN ('deleted', 'failed')),
+    CONSTRAINT sense_orphan_cleanup_not_blank CHECK (
+        btrim(path_name) <> '' AND btrim(actor_id) <> ''
+        AND (error_code IS NULL OR btrim(error_code) <> '')
+    ),
+    CONSTRAINT sense_orphan_cleanup_error_shape CHECK (
+        (status = 'deleted' AND error_code IS NULL)
+        OR (status = 'failed' AND error_code IS NOT NULL)
+    )
+);
+ALTER TABLE sense.orphan_cleanup_actions OWNER TO sense_app;
+
+CREATE INDEX IF NOT EXISTS sense_orphan_scan_completed_idx
+    ON sense.orphan_scan_runs(completed_at);
+CREATE INDEX IF NOT EXISTS sense_orphan_cleanup_status_idx
+    ON sense.orphan_cleanup_actions(scan_id, status);
+
+INSERT INTO sense.schema_migrations(version) VALUES (5)
+ON CONFLICT (version) DO NOTHING;
@@ -0,0 +1,17 @@
+-- Extend the least-privilege boundary to reconciliation and orphan state.
+
+REVOKE ALL ON TABLE
+    sense.media_path_ownership,
+    sense.operational_leases,
+    sense.orphan_scan_runs,
+    sense.orphan_scan_findings,
+    sense.orphan_cleanup_actions
+FROM PUBLIC;
+
+GRANT SELECT, INSERT, UPDATE, DELETE ON TABLE
+    sense.media_path_ownership,
+    sense.operational_leases,
+    sense.orphan_scan_runs,
+    sense.orphan_scan_findings,
+    sense.orphan_cleanup_actions
+TO sense_app;
@@ -1,6 +1,6 @@
 # YoVision PostgreSQL 初始化

-本目录实现 T-009～T-011 的 PostgreSQL `17.10` schema。SQL 必须按文件名前缀顺序执行：`001`～`004` 创建 NOLOGIN 权限角色、Bell/Sense 初始对象和配额权限；`005`～`007` 增量增加 Area 与审计；`008`～`009` 增加 Control API 的资源版本、write-only profile token、24 小时幂等收据、batch operation 和对应最小权限。全部 SQL 可重放。应用登录角色及其密码由部署环境或密钥系统创建，并分别加入 `bell_app`、`sense_app`；仓库不保存登录凭据。
+本目录实现 T-009～T-012 的 PostgreSQL `17.10` schema。SQL 必须按文件名前缀顺序执行：`001`～`004` 创建 NOLOGIN 权限角色、Bell/Sense 初始对象和配额权限；`005`～`007` 增量增加 Area 与审计；`008`～`009` 增加 Control API 状态；`010`～`011` 增加调和 fencing、MediaMTX Path 历史归属、孤儿报告/受控处置结果和对应最小权限。全部 SQL 可重放。应用登录角色及其密码由部署环境或密钥系统创建，并分别加入 `bell_app`、`sense_app`；仓库不保存登录凭据。

 生产/共享实例必须由管理员先备份并在 YoVision 专用数据库中执行。Sense 进程不会用高权限自动建库或建角色。示例只使用私有环境变量，不把实际 DSN 写入脚本或日志：

@@ -22,8 +22,9 @@ Get-ChildItem deploy/postgres/[0-9][0-9][0-9]_*.sql |
 - `sense_app` 对 Bell 源表、Bell migration 表和 trigger function 没有权限；启动检查发现权限过宽时拒绝运行。
 - `sense.device_operation_outbox` 是本地持久化审计事实，不是 Bell 全局审计真相；relay 的传输、签名、确认和留存尚未实现。
 - `sense.control_idempotency_receipts` 不保存原始 Idempotency-Key，只保存 scope/request SHA-256 和脱敏响应快照；`batch_operations`/items 只保存逻辑 ID、状态和稳定错误，不保存连接秘密。
+- 调和与孤儿租约使用 PostgreSQL `clock_timestamp()` 和 fencing token；过期 worker 不能提交完成/失败或扫描报告。`media_path_ownership`、扫描和处置表不保存 endpoint、credential 或 source URI；数据库约束禁止为 `unowned` finding 写删除结果。
 - PUBLIC 对两个业务 schema 的表和函数没有权限。

 ## 恢复

-T-009～T-011 不迁移或删除现有 SQLite 数据，默认运行配置也仍为 SQLite。SQLite 不具备生产 Area/Outbox/Control API 一致性，业务路由不会在 SQLite 上注册。若 PostgreSQL 切换验证失败，先关闭 Control API feature flag；需要实验室回退时停止 PostgreSQL 模式进程并恢复 SQLite DSN，不得误报为生产语义等价。schema 不提供自动 down migration；需要移除时由管理员先备份，再只删除明确确认的 YoVision 专用数据库，禁止对共享实例做模糊或递归清理。
+T-009～T-012 不迁移或删除现有 SQLite 数据，默认运行配置也仍为 SQLite。SQLite 不具备生产 Area/Outbox/Control API、多实例租约或孤儿处置一致性，业务路由不会在 SQLite 上注册。若 PostgreSQL 切换验证失败，先关闭 Control API 和孤儿扫描；需要实验室回退时停止 PostgreSQL 模式进程并恢复 SQLite DSN，不得误报为生产语义等价。schema 不提供自动 down migration；需要移除时由管理员先备份，再只删除明确确认的 YoVision 专用数据库，禁止对共享实例做模糊或递归清理。
@@ -72,7 +72,7 @@ BEGIN
        RAISE EXCEPTION 'sense_app lacks access to its local audit Outbox';
    END IF;
    IF (SELECT max(version) FROM bell.schema_migrations) <> 2
-       OR (SELECT max(version) FROM sense.schema_migrations) <> 4 THEN
+       OR (SELECT max(version) FROM sense.schema_migrations) <> 5 THEN
        RAISE EXCEPTION 'schema migration version drift';
    END IF;

@@ -129,9 +129,9 @@ $assertions$;

 DO $control_api$
 BEGIN
-    IF NOT has_table_privilege('yovision_t011_sense', 'sense.control_idempotency_receipts', 'SELECT,INSERT,UPDATE,DELETE')
-       OR NOT has_table_privilege('yovision_t011_sense', 'sense.batch_operations', 'SELECT,INSERT,UPDATE,DELETE')
-       OR NOT has_table_privilege('yovision_t011_sense', 'sense.batch_operation_items', 'SELECT,INSERT,UPDATE,DELETE') THEN
+    IF NOT has_table_privilege('yovision_t012_sense', 'sense.control_idempotency_receipts', 'SELECT,INSERT,UPDATE,DELETE')
+       OR NOT has_table_privilege('yovision_t012_sense', 'sense.batch_operations', 'SELECT,INSERT,UPDATE,DELETE')
+       OR NOT has_table_privilege('yovision_t012_sense', 'sense.batch_operation_items', 'SELECT,INSERT,UPDATE,DELETE') THEN
        RAISE EXCEPTION 'Sense runtime lacks Control API table privileges';
    END IF;
    IF has_table_privilege('public', 'sense.control_idempotency_receipts', 'SELECT')
@@ -151,3 +151,31 @@ BEGIN
    END IF;
 END
 $control_api$;
+
+DO $reconcile_safety$
+DECLARE
+    target text;
+BEGIN
+    FOREACH target IN ARRAY ARRAY[
+        'sense.media_path_ownership',
+        'sense.operational_leases',
+        'sense.orphan_scan_runs',
+        'sense.orphan_scan_findings',
+        'sense.orphan_cleanup_actions'
+    ] LOOP
+        IF NOT has_table_privilege('yovision_t012_sense', target, 'SELECT,INSERT,UPDATE,DELETE') THEN
+            RAISE EXCEPTION 'Sense runtime lacks reconciliation safety privilege on %', target;
+        END IF;
+        IF has_table_privilege('public', target, 'SELECT,INSERT,UPDATE,DELETE') THEN
+            RAISE EXCEPTION 'Reconciliation safety state leaked to PUBLIC on %', target;
+        END IF;
+    END LOOP;
+    IF NOT EXISTS (
+        SELECT 1 FROM information_schema.columns
+        WHERE table_schema = 'sense' AND table_name = 'reconcile_state'
+          AND column_name = 'lease_token'
+    ) THEN
+        RAISE EXCEPTION 'reconcile fencing columns are missing';
+    END IF;
+END
+$reconcile_safety$;
@@ -40,13 +40,13 @@ MVP 以默认 16 路跑通一个场景的端到端闭环；架构、数据和 UI

 ## 当前阶段

-当前为 **M0 指定型号实机准入已完成，M1 Sense 五路混合源集成待执行**。后续本地开发统一使用已准入的一台海康样机，多路软件闭环使用独立合成 RTSP 源补足；真实多设备证据延后到客户/借用/租赁条件具备时执行。
+当前为 **M0 指定型号实机准入与 M1 Sense 五路混合源集成均已完成，M2 正在补齐生产接入边界**。后续本地开发统一使用已准入的一台海康样机，多路软件闭环使用独立合成 RTSP 源补足；真实多设备证据延后到客户/借用/租赁条件具备时执行。

 优先路径：

 1. M0：已用一台真实样机完成首期指定“型号 + 硬件版本 + 固件”ONVIF/RTSP 准入；结论不外推为多品牌兼容，真实断网恢复证据按负责人豁免留痕。
 2. M1：只在 `Sense/` 建立 MediaMTX 生产接入骨架，以 1 路准入实机 + 至少 4 路独立合成源完成五路自动建 path、探活和断线重建。
-3. M2：对账、多租户、隧道和至少一个站点的 16 路全流程。
+3. M2：Control API、多租户投影、调和 fencing、孤儿安全闸已完成；下一步是隧道和至少一个站点的 16 路全流程。
 4. M3：Brain + Bell 起步，默认 16 路端到端事件、预警、ack 与误报反馈。
 5. M4–M5：64/128 路分片、管理端和第二/第三场景包。

@@ -50,7 +50,9 @@ MediaMTX 官方 `v1.19.3` OpenAPI 固定保存于 `Sense/api/vendor/mediamtx-v1.

 M2 仍保留 SQLite 作为单机开发与 T-006 回归入口，但生产真相源切换目标是 PostgreSQL。运行时必须显式设置 `SENSE_DB_DRIVER=postgres` 和私有 `SENSE_DB_DSN` 才使用 PostgreSQL；仓库不提供带密码 DSN。初始 schema 由 `deploy/postgres/` 的顺序 migration 创建，不在 Sense 进程启动时使用高权限自动迁移。

-T-011 复用同一套冻结的 Go、PostgreSQL、pgx 与 `oapi-codegen`，未增加第三方依赖。Sense Control API v1 从仓库 OpenAPI 生成 `net/http` server glue；首版认证是可替换 port 下的外部静态 SHA-256 注册表，公共路由默认关闭且只在 PostgreSQL v4 schema 上启用。静态适配器是私有部署启动方案，不替代后续 Bell/JWT/OIDC 身份服务。
+T-011 复用同一套冻结的 Go、PostgreSQL、pgx 与 `oapi-codegen`，未增加第三方依赖。Sense Control API v1 从仓库 OpenAPI 生成 `net/http` server glue；首版认证是可替换 port 下的外部静态 SHA-256 注册表，公共路由默认关闭且只在 PostgreSQL 生产路径启用。T-011 首次落成 v4 schema，T-012 后进程统一要求 v5。静态适配器是私有部署启动方案，不替代后续 Bell/JWT/OIDC 身份服务。
+
+T-012 同样未增加第三方依赖：PostgreSQL v5 使用数据库时钟租约和 fencing token 协调多实例调和/孤儿扫描；MediaMTX Path 枚举继续使用同版生成客户端。`/metrics` 由 Go 标准库输出 Prometheus 文本格式，只包含固定结果枚举、构建/实例信息和汇总 gauge/counter，不使用 tenant、Site、device 或 Path 标签。孤儿处置是 PostgreSQL 专用本地命令，默认周期任务只报告、不自动删除。

 ## 2. 外部项目边界

@@ -46,7 +46,7 @@ Sense ── 视频流/触发信号 ──> Brain
 主流程：

 1. Bell 持有站点、Area、配额与 `capture_policy`；首期在同一 PostgreSQL 实例内发布 `bell.site_quota_v1` 和 `bell.area_policy_v1` 两个版本化只读视图。T-009/T-010 已实现 Bell 源表/视图、最小权限和 Sense PostgreSQL repository；Sense 按 Area→Site 的固定 advisory-lock 顺序执行策略与配额准入并记录所用版本。未来分库必须发布新版本契约，不能静默改变 v1 语义。
-2. Sense 维护设备期望态，通过 MediaMTX API 和对账器收敛实际态。
+2. Sense 维护设备期望态，通过 MediaMTX API 和对账器收敛实际态；PostgreSQL 多实例以数据库时钟短租约和 fencing token 领取 due row，过期 worker 不得提交结果。
 3. Brain 消费视频与触发信号，产生符合 v0.1 的事件。
 4. Bell 做 schema 与代码级断言，生成平台 ULID，保存不可变事件。
 5. 规则命中后创建独立 Alert，先落库再投递，等待 ack 并按策略升级。
@@ -82,8 +82,8 @@ Sense ── 视频流/触发信号 ──> Brain
 ## 7. 一致性与失败处理

 - PostgreSQL `sense` schema 是生产期望态真相源；SQLite 只保留为 M1 本地开发/回归路径。MediaMTX、推理 worker 和对象存储是可对账的实际态。
- 对账器水平触发、幂等、指数退避、限制并发；部分失败不做跨系统回滚，只持续收敛。
- 孤儿删除必须有 10% 安全闸和人工可观察指标。
+- 对账器水平触发、幂等、指数退避、限制并发；PostgreSQL 使用 `FOR UPDATE SKIP LOCKED`、每项续租和 fencing token，SQLite 只保留单进程开发语义。部分失败不做跨系统回滚，只持续收敛。
+- MediaMTX Path 扫描把“Sense 历史拥有但当前失配”和“从未归属 Sense”分开；未知归属永不自动删除。历史拥有项也只允许在 15 分钟二次快照、1～128 项和 `候选 × 100 <= 当前 Path 总数 × 10` 全部通过时由本地运维命令逐项处置，不提供绕过。
 - `bell.site_quota_v1` 行缺失、数值越界、版本回退或读取失败只阻止视频设备新增/启用，不中断已有流；降低配额导致超限时不自动停用，后续准入返回稳定错误并产生运维信号。多 Sense 实例使用 PostgreSQL transaction-scoped advisory lock 串行化同 tenant/site 的计数与写入，不能用进程内锁替代。
 - `bell.area_policy_v1` 缺失、非法、版本回退或读取失败时，PostgreSQL repository 拒绝相关新增/启用；`non_imaging_only` 允许非成像设备但拒绝具有 `video_capture` 的设备。已有设备保持原状态，策略冲突由 Bell 管理端显式迁移或取消。同库实时视图不以源记录年龄误判 freshness。
 - 设备创建和期望态受理在本地事务内同时写脱敏 `sense.device_operation_outbox`；Outbox 失败回滚业务写入，相同期望态不增加 generation 但仍审计。异步 relay 尚未实现。
@@ -107,14 +107,14 @@ Sense ── 视频流/触发信号 ──> Brain
 ## 9. 目录目标

 ```text
-Sense/cmd + Sense/internal/{device,onvif,mtx,reconcile,probe,trigger,tunnel,authcb,store}
+Sense/cmd + Sense/internal/{device,onvif,mtx,reconcile,orphan,metrics,probe,trigger,tunnel,auth,store}
 Brain/{pipeline,models,judge,emit,trigger,contracts}
 Bell/cmd + Bell/internal/{ingest,event,rule,alert,deliver,feedback,tenant,audit,store}
 Bell/{web,packs,contracts}
-deploy/postgres/{001_roles.sql,...,007_privileges_area_audit.sql,tests}
+deploy/postgres/{001_roles.sql,...,011_privileges_reconcile_safety.sql,tests}
 ```

-Sense 脚手架和 PostgreSQL `001`～`007` 已实现；Brain/Bell 应用目录仍是占位，真实服务必须由对应任务创建。
+Sense 脚手架和 PostgreSQL `001`～`011` 已实现；Brain/Bell 应用目录仍是占位，真实服务必须由对应任务创建。

 ## 10. 开发顺序

@@ -26,7 +26,7 @@
 - T-009：使用 PostgreSQL 17.10 建立 `sense`/`bell` schema、Bell 配额源表/只读视图、最小权限和 Sense PostgreSQL repository；SQLite 保留为 M1 本地路径。
 - T-010：建立 Bell Area/`capture_policy` 只读投影、Sense 成像准入与设备操作审计 Outbox 原子写入；relay 传输协议留给后续任务。
 - T-011：按 T-008 契约实现认证 tenant 上下文、7 个设备/operation handler、PostgreSQL 幂等收据、ETag/HMAC cursor、最多 128 项批量操作和停用精确 path 收敛；公共控制 API 默认关闭且只在 PostgreSQL 路径启用。
- 对账器幂等/退避/并发/10% 安全闸。
+- T-012：以 PostgreSQL 数据库时钟租约和 fencing token 防止多实例重复调和；增加可证明的 Path 历史归属、默认只读孤儿报告、15 分钟二次确认、10% 不可绕过安全闸和低基数 `/metrics`。
 - WireGuard 边缘隧道与断网恢复。
 - 16 路批量开通、停用和容量基准。

@@ -69,10 +69,10 @@ Sense 使用 MediaMTX 官方 OpenAPI 生成客户端并加薄封装。业务代
 | Port | 操作 | 数据所有者 / 失败语义 |
 | --- | --- | --- |
 | ONVIF adapter | `Probe(target)`、`SetSystemDateAndTime(target, time)` | 设备是外部来源；`target` 只含 endpoint ref 与不透明 credential ref。错误稳定映射为认证失败、超时、不可用、响应无效，不记录凭据或完整流地址 |
-| MediaMTX paths | `CreatePath`、`GetPath`、`EnsurePath`、`DeletePath`、`PathReady` | SQLite 设备台账持有期望态，MediaMTX 只持有运行配置；`EnsurePath` 相同 source 不写、不同 source patch、缺失时 add；当前调和器绝不枚举或删除孤儿 |
+| MediaMTX paths | `CreatePath`、`GetPath`、`EnsurePath`、`DeletePath`、`PathReady`、`ListPathNames` | 设备台账持有期望态，MediaMTX 只持有运行配置；分页枚举只返回名称、不返回 source。普通调和只操作精确台账 Path；独立孤儿流程只报告未知归属，受控处置只接受有历史归属的 stale Path |
 | Device repository | 设备、期望态、实际态、调和进度 | SQLite 是 M1 默认开发路径；PostgreSQL 是 M2 生产路径并只读 Site/Area 两个 Bell 视图。调和退避持久化；配额/Area 失败时拒绝相关新增或启用，不关闭已有流；高风险写入与脱敏 Outbox 同事务 |

-MediaMTX 薄封装调用同版官方 OpenAPI 的 `/v3/config/paths/get|add|patch|delete/{name}` 与 `/v3/paths/get/{name}`。生成源、版本和 SHA-256 见 `docs/03-tech-stack.md`；业务包不得直接 import 生成包。
+MediaMTX 薄封装调用同版官方 OpenAPI 的 `/v3/config/paths/get|add|patch|delete/{name}`、`/v3/config/paths/list` 与 `/v3/paths/get/{name}`。列表有最大页数和重复页保护，source URI 在薄封装内丢弃。生成源、版本和 SHA-256 见 `docs/03-tech-stack.md`；业务包不得直接 import 生成包。

 ### 5.2 设备台账语义

@@ -83,7 +83,9 @@ MediaMTX 薄封装调用同版官方 OpenAPI 的 `/v3/config/paths/get|add|patch

 ### 5.3 Sense 进程 HTTP 面

-`GET /healthz` 表示进程存活，`GET /readyz` 表示所选数据库已打开且 schema/权限前置检查完成；两者不要求认证，也不等价于摄像头、MediaMTX path 或里程碑健康。T-011 在 `SENSE_CONTROL_API_ENABLED=true`、PostgreSQL v4 schema 和外部安全文件全部有效时注册冻结的 `/api/v1` 路由；默认 SQLite 运行仍只暴露探针。业务路由不提供无 Site 边界的 `/api/v1/devices` 临时接口。
+`GET /healthz` 表示进程存活，`GET /readyz` 表示所选数据库已打开且 schema/权限前置检查完成；两者不要求认证，也不等价于摄像头、MediaMTX path 或里程碑健康。`GET /metrics` 默认输出无租户/设备/Path 标签的 Prometheus 汇总，可通过 `SENSE_METRICS_ENABLED=false` 关闭。T-011 在 `SENSE_CONTROL_API_ENABLED=true`、PostgreSQL v5 schema 和外部安全文件全部有效时注册冻结的 `/api/v1` 路由；默认 SQLite 不注册业务路由。业务路由不提供无 Site 边界的 `/api/v1/devices` 临时接口。
+
+孤儿报告/处置不是公共 HTTP API。它由 PostgreSQL 专用 `cmd/sense-orphan` 在受控运维主机执行，使用 15 分钟 scan ID、actor、精确确认文本和运行前二次快照；不改变 Sense Control API v1 的 7 个 endpoint。

 ## 6. 变更流程

@@ -4,8 +4,8 @@

 ## 当前阶段

- 阶段：M0 指定摄像头型号准入已完成；M1 的“一实机 + 四合成源”实验室软件闭环已通过；M2 已完成控制面契约、Site/Area 准入与本地审计事务基础，但五条独立真实上游和生产 SLA 尚未验收。
- 生产代码：Sense 已包含可构建进程、SQLite/PostgreSQL repository、Site/Area 准入、设备操作 Outbox、标准 ONVIF SOAP/WS-Security adapter、凭据引用、MediaMTX 生成客户端、对账、探活和 MediaMTX 重启重建；公共设备管理 API、认证、Outbox relay 和完整生产部署仍未实现。
+- 阶段：M0 指定摄像头型号准入已完成；M1 的“一实机 + 四合成源”实验室软件闭环已通过；M2 已完成 Control API、Site/Area 准入、本地审计事务、多实例调和 fencing 和孤儿受控处置基础，但 WireGuard、16 路容量出口、五条独立真实上游和生产 SLA 尚未验收。
+- 生产代码：Sense 已包含可构建进程、SQLite/PostgreSQL repository、Site/Area 准入、设备操作 Outbox、标准 ONVIF SOAP/WS-Security adapter、凭据引用、MediaMTX 生成客户端、Control API v1、对账/探活、数据库租约、孤儿只读扫描/受控命令和低基数指标；Bell 管理服务/JWT、Outbox relay、WireGuard 和完整生产部署仍未实现。
 - 默认容量：16 路；单站点本阶段上限 128 路，必须横向分片。

 ## 仓库现实
@@ -14,13 +14,14 @@
 - Sense 设备模型使用 `modality + capabilities`，SQLite 执行 v1 migration；视频配额默认 16、允许 1～128，17/128/129、新增/启用和“降低配额不关闭已有流”均有测试。
 - T-009 冻结 PostgreSQL `17.10` 和 `pgx/v5 v5.10.0`，实现 `bell`/`sense` schema、NOLOGIN 权限角色、Bell Site 版本 trigger、`bell.site_quota_v1` 和 Sense PostgreSQL repository；同站点并发准入用事务级 advisory lock，配额缺失/越界/版本回退时失败关闭且不改变已有流。
 - T-010 增量实现 `bell.areas`、`bell.area_policy_v1`、Area 版本观察和 `sense.device_operation_outbox`；`non_imaging_only` 拒绝成像设备创建/启用，失败不改变已有设备。设备创建/期望态受理与脱敏 Outbox 同事务，相同期望态不增加 generation 但仍审计。
- Windows 隔离测试使用 `D:\pgsql17\bin` 启动随机回环端口临时集群，`001`～`007` migration 可重放、权限断言和 18 个真实 PostgreSQL 测试通过后自动清理；现有 `D:\pgsql17\data` 和 5432 服务未被读取、停止或修改。
- MediaMTX 固定为独立二进制 `v1.19.3`，官方 OpenAPI 已按 SHA-256 vendoring，并由固定 `oapi-codegen v2.8.0` 生成客户端；手写薄封装有 create/read/delete、幂等 ensure 与探活假 HTTP 测试。
- T-003 对账进度与指数退避持久化，覆盖取消和 SQLite 重启恢复；T-006 增加真实 ONVIF adapter、RTSP router、实验室播种/状态工具、故障代理和五路自动验收。当前仍不枚举/删除孤儿。
+- Windows 隔离测试使用 `D:\pgsql17\bin` 启动随机回环端口临时集群，`001`～`011` migration 可重放、权限断言和 29 个真实 PostgreSQL 测试通过后自动清理；现有 `D:\pgsql17\data` 和 5432 服务未被读取、停止或修改。
+- MediaMTX 固定为独立二进制 `v1.19.3`，官方 OpenAPI 已按 SHA-256 vendoring，并由固定 `oapi-codegen v2.8.0` 生成客户端；手写薄封装有 create/read/delete、幂等 ensure、探活和只返回名称的受限分页枚举测试。
+- T-003 对账进度与指数退避持久化，覆盖取消和 SQLite 重启恢复；T-006 增加真实 ONVIF adapter、RTSP router、实验室播种/状态工具、故障代理和五路自动验收。T-012 的普通调和不枚举孤儿；独立 PostgreSQL 扫描默认只报告，未知归属永不删除。
 - T-006 正式使用 1 台准入实机和 4 个独立合成 publisher 连续观察 `1806.6 s` / 180 次采样，四类恢复均通过，最大与最终 `unconverged` 均为 0；详细证据见 `docs/research/sense-5-stream-integration.md`。
 - `docs/raw/01`～`08` 已记录需求、分析、方案、客户场景、事件比对和三系统职责。
 - `docs/raw/contracts/event-v0.1.schema.json` 已冻结，并有多份示例与语义说明。
- `docs/contracts/sense-control-v1.openapi.json` 的 7 个站点作用域/operation endpoint 已由 T-011 实现：外部静态 SHA-256 主体注册表、tenant/Site scope、HMAC cursor、ETag、PostgreSQL 24 小时幂等收据和最多 128 项 batch operation 均有代码与隔离集成测试。默认 SQLite 仍只暴露探针；Bell 管理服务、JWT/OIDC 和 Outbox relay 尚未实现。
+- `docs/contracts/sense-control-v1.openapi.json` 的 7 个站点作用域/operation endpoint 已由 T-011 实现：外部静态 SHA-256 主体注册表、tenant/Site scope、HMAC cursor、ETag、PostgreSQL 24 小时幂等收据和最多 128 项 batch operation 均有代码与隔离集成测试。默认 SQLite 只暴露运维探针与低基数 `/metrics`，不注册业务路由；Bell 管理服务、JWT/OIDC 和 Outbox relay 尚未实现。
+- T-012 把 PostgreSQL schema 提升到 v5：due row 用数据库时钟、`FOR UPDATE SKIP LOCKED`、逐项续租和 fencing token 协调；MediaMTX Path 历史归属、15 分钟孤儿快照、最多 10%/128 项安全闸、无 bypass 的本地处置命令及 `/metrics` 已实现。SQLite 明确保留单实例开发语义。
 - harness coding 文档、上下文清单、Gitea Issue/PR 模板和治理脚本已接入。
 - Gitea 已初始化 12 个协作标签；`status/waiting` 用于依赖或外部条件未满足的未领取任务，实时可领取状态必须从 Gitea 查询，不在本文复制。
 - T-002 已关闭架构影响型问题：首期 S2 民办寄宿学校、客户侧私有部署、ONVIF/RTSP、NVIDIA 主路径、自研 Bell、Web/H5 与客户侧证据存储等边界已批准。
@@ -55,7 +56,7 @@ python scripts/validate_harness_governance.py
 ./scripts/test_postgres.ps1 -PgRoot D:\pgsql17
 ```

-Sense 默认监听 `127.0.0.1:8080`，提供 `/healthz` 与 `/readyz` 运维探针；它们不代表摄像头或 M1 里程碑健康。MediaMTX 获取、校验和独立启动方法见 `Sense/README.md`。
+Sense 默认监听 `127.0.0.1:8080`，提供 `/healthz`、`/readyz` 运维探针和低基数 `/metrics`；探针不代表摄像头或 M1 里程碑健康。MediaMTX 获取、校验和独立启动方法见 `Sense/README.md`。

 ## 当前 blocker / 待确认

@@ -66,12 +67,12 @@ Sense 默认监听 `127.0.0.1:8080`，提供 `/healthz` 与 `/readyz` 运维探
 - 人脸方向已延后至 M5 的 S4 成人园区候选试点；必要性/PIP 影响评估、单独同意与替代方式、合法底库来源和删除流程未完成，阻塞人脸能力上线。
 - 短信/语音具体供应商未选；生产前必须选定两条独立投递路径并验证故障切换。
 - Python/Savant 的精确版本、目标硬件和 Bell 前端栈尚未冻结；Sense M1 的 Go、SQLite driver、MediaMTX、生成器及生成运行时版本已在 T-003 冻结，PostgreSQL/pgx 版本已在 T-009 冻结。
- 本机现有 PostgreSQL 5432 实例使用 SCRAM 且当前开发进程没有管理员密码；T-009～T-011 不绕过认证，自动验收使用隔离临时集群。向共享/生产实例安装 migration 前仍需管理员私下提供专用数据库、登录角色、外部 Control API 安全文件与备份方案。
+- 本机现有 PostgreSQL 5432 实例使用 SCRAM 且当前开发进程没有管理员密码；T-009～T-012 不绕过认证，自动验收使用隔离临时集群。向共享/生产实例安装 migration 前仍需管理员私下提供专用数据库、登录角色、外部 Control API 安全文件与备份方案。
 - 代码知识图谱在无业务代码阶段可能为空；工具不可用时使用 `rg` 处理文档与配置。

 ## 下一步

-下一项建议创建 T-012，补齐 M2 对账安全闸、孤儿检测（只报告/受控处置）与多实例可观测性，再进入 WireGuard 和 16 路批量开通基准；不要把 Control API 完成误报为 M2 容量出口。客户授权、借用或租赁条件具备后再执行 T-007 五条独立真实上游现场门禁。T-006 的合成结果不解除 T-007，也不形成容量或生产 SLA 承诺。
+下一项建议先冻结并实施 WireGuard 边缘隧道与断网恢复，再创建 16 路批量开通/停用和分维度容量基准任务；T-012 完成不等于 M2 容量出口。客户授权、借用或租赁条件具备后再执行 T-007 五条独立真实上游现场门禁。T-006 的合成结果不解除 T-007，也不形成容量或生产 SLA 承诺。

 ## 已知风险

@@ -0,0 +1,65 @@
+# Sense 对账与孤儿处置 Runbook
+
+本文适用于 PostgreSQL v5（migration `001`～`011`）与 MediaMTX `v1.19.3`。SQLite 仅用于单进程开发，不具备本 runbook 的多实例 fencing 或孤儿处置语义。
+
+## 观察入口
+
+- `GET /healthz`：进程存活。
+- `GET /readyz`：数据库 schema 和权限前置检查已完成。
+- `GET /metrics`：低基数进程指标；不包含 tenant、Site、device、Path 或 URI 标签。
+- 日志中的 `background convergence error`：只输出稳定、脱敏错误，不输出 stream URI 或凭据。
+
+重点指标：
+
+- `sense_reconcile_devices{state="unconverged"}`：未收敛的 enabled 视频设备数。
+- `sense_reconcile_items_total{result="lease_lost"}`：worker 在执行前或提交时失去 fencing token 的次数；失去租约的 worker 不得继续产生外部变更。
+- `sense_orphan_paths{classification="owned_stale|unowned"}`：最近一次成功扫描的两类差异。
+- `sense_orphan_cleanup_blocked_total`、`sense_orphan_cleanup_items_total`：安全闸阻断和人工处置结果。
+
+## 多实例调和
+
+每个实例应注入唯一、稳定且不含客户信息的 `SENSE_INSTANCE_ID`。PostgreSQL 使用数据库时钟、短事务、`FOR UPDATE SKIP LOCKED` 和随机 fencing token 领取 due row；批量等待中的每一项在调用 ONVIF/MediaMTX 前续租。默认租期 30 秒、单项 deadline 20 秒；operation timeout 必须严格短于租期。
+
+发现 `lease_lost` 时先检查实例 ID 是否重复、数据库时钟和外部调用延迟。不要通过延长到超过 5 分钟或关闭 fencing 规避问题；先定位超时，再在变更评审后同时调整租期和单项 deadline。
+
+## 生成孤儿报告
+
+在已安装 v5 migration、能访问同一 PostgreSQL 和目标 MediaMTX 的受控运维主机执行：
+
+```powershell
+$env:SENSE_DB_DRIVER = 'postgres'
+$env:SENSE_DB_DSN = '由部署环境私下设置'
+$env:SENSE_MEDIAMTX_URL = '受控 MediaMTX API 地址'
+go -C Sense run ./cmd/sense-orphan -mode report
+```
+
+命令只输出 `scan_id`、计数、安全闸结论和过期时间。分类含义：
+
+- `owned_stale`：Sense 有历史 Path 归属记录，但当前没有同一设备继续声明该 Path；可进入受控候选。
+- `unowned`：没有 Sense 归属证据；可能属于人工或其他系统，永远只报告。
+- 当前设备仍声明的 Path 不计入 finding。
+
+不要因为名称相似把 `unowned` 手工改成 owned。先调查其创建者和用途；需要接管时应走独立、可审计的迁移任务。
+
+## 执行受控处置
+
+仅当报告显示 `safety_allowed=true`，并在 15 分钟有效期内由授权运维人员执行：
+
+```powershell
+go -C Sense run ./cmd/sense-orphan -mode apply `
+  -scan-id 'scan_...' `
+  -actor 'operator-id' `
+  -confirm 'DELETE scan_...'
+```
+
+执行前程序会重新枚举 Path、重新读取归属并重新计算：候选必须仍是报告集合的子集、数量 1～128，且满足 `候选数 × 100 <= 当前 Path 总数 × 10`。过期、比例超限、归属变化、确认不匹配或租约冲突都会整批零删除；没有 `force`/bypass。删除逐项记入 PostgreSQL，成功项重复执行会跳过，失败项在快照仍有效且重新计算仍通过时可重试。
+
+## 故障与恢复
+
+- 扫描失败：不产生可执行快照；修复数据库/MediaMTX 连通性后重新 `report`。
+- `ratio_exceeded`：停止自动化，核对 MediaMTX 实例/分片是否选错、台账是否缺失或发生大面积配置漂移；不得拆小批次规避 10% 闸。
+- 部分删除失败：保留输出与数据库处置事实，修复 MediaMTX 后用同一 scan ID 重试；超过 15 分钟必须重新报告和审批。
+- 误删怀疑：立即停止 apply。数据库期望态仍是来源；当前设备声明的 Path 会由调和器重建。无当前设备声明的历史 Path 不自动恢复，应根据变更记录人工确认来源。
+- 回滚 Sense 二进制：先停用 `SENSE_ORPHAN_SCAN_ENABLED` 并停止所有新实例，再回滚；v5 表和已记录事实保留，不执行破坏性 down migration。
+
+本流程只处置 MediaMTX 配置 Path，不删除设备、录像、证据对象或 Bell 数据。
@@ -3,7 +3,7 @@ id: T-012
 title: 补齐 Sense 对账安全闸、孤儿受控处置与多实例可观测性
 phase: 2
 deps: [T-011]
-status: DOING
+status: DONE
 created: 2026-08-07
 issue: 43
 context_ref: 25357723a09218bae58626a9934309abcda34899
@@ -88,6 +88,16 @@ T-003 已实现幂等调和与持久化退避，T-011 已完成 Control API、

 ## 执行记录

+### 2026-08-07 完成多实例对账安全与孤儿受控处置
+
+- PostgreSQL schema 提升到 v5：`010`/`011` 为 due row 增加 owner/token/expiry fencing，新增不含 source URI 的 MediaMTX Path 历史归属、孤儿扫描租约、15 分钟报告、发现项和处置审计，并补齐 `sense_app` 最小权限与 PUBLIC 拒绝；启动会拒绝旧 schema 或越权运行角色。
+- 调和器改为 `FOR UPDATE SKIP LOCKED` 批量领取、逐项续租、短于租期的外部调用 deadline 和带 token 的完成/失败；数据库时钟判定租约，旧 worker 不能覆盖接管者结果。成功启用在同一完成事务中刷新 Path 归属；SQLite 明确保留单进程开发语义。
+- MediaMTX 薄客户端新增只返回排序去重 Path 名称的有界分页枚举，包含重复页保护且不读取/返回 source；独立孤儿扫描把差异分为 `owned_stale` 与永不删除的 `unowned`。周期任务默认只报告，人工 `sense-orphan apply` 必须使用 15 分钟内 scan ID、合法 actor 和精确确认文本，并在删除前重取库存/归属。
+- 处置安全闸固定为 1～128 项且不超过当前 Path 的 10%，使用整数交叉相乘，无 force/bypass；目标只能从原快照收窄。逐项结果可审计和幂等重试，数据库外键/约束也拒绝为 `unowned` 写入删除记录。
+- 新增无第三方运行时依赖的 `/metrics`：只使用构建/实例及固定结果枚举标签，覆盖调和 run/item/lease-lost/时长/未收敛和孤儿扫描/阻断/删除汇总；tenant、Site、device、Path、URI 与错误正文均不进入标签。
+- `./scripts/test_postgres.ps1 -PgRoot D:\pgsql17` 通过：隔离 PostgreSQL 17.10 临时集群连续重放 `001`～`011`，SQL 权限/约束断言及 29 个 `TestPostgres*` 全绿，覆盖两个 store 并发唯一领取、租约接管/旧 token fencing、归属记录、扫描租约、未知 Path 数据库拒删和处置幂等；随机端口实例已停止并清理，现有 `D:\pgsql17\data`/5432 未被读取、停止或修改。
+- `./init.ps1` 通过；Python 治理/契约测试共 52 项通过，T-012 定向静态契约 12 项通过；全部 Go 测试、生成漂移、`go vet`、`go build` 与 `git diff --check` 通过；`go test -race ./internal/metrics ./internal/mtx ./internal/orphan ./internal/reconcile ./internal/store` 通过。不需要摄像头/GPU，本结果不解除 T-007，也不形成 16/128 路容量承诺。
+
 ### 2026-08-07 领取任务

 - dispatcher `ila` 将 Issue #43 分配给 `codex`；`context_ref` 为 `25357723a09218bae58626a9934309abcda34899`，claim 为 `claims/T-012`，工作分支为 `agent/codex/T-012`。
@@ -40,7 +40,7 @@ $port = ([System.Net.IPEndPoint]$listener.LocalEndpoint).Port
 $listener.Stop()

 $systemTemp = [IO.Path]::GetFullPath([IO.Path]::GetTempPath())
-$clusterName = 'yovision-t011-pg-' + [guid]::NewGuid().ToString('N')
+$clusterName = 'yovision-t012-pg-' + [guid]::NewGuid().ToString('N')
 $dataDir = Join-Path $systemTemp $clusterName
 $logPath = Join-Path $systemTemp ($clusterName + '.log')
 $started = $false
@@ -69,9 +69,9 @@ try {
    $started = $true

    $adminRootDSN = "postgres://postgres@127.0.0.1:$port/postgres?sslmode=disable"
-    $databaseName = 'yovision_t011'
+    $databaseName = 'yovision_t012'
    $adminDatabaseDSN = "postgres://postgres@127.0.0.1:$port/${databaseName}?sslmode=disable"
-    $senseDSN = "postgres://yovision_t011_sense@127.0.0.1:$port/${databaseName}?sslmode=disable"
+    $senseDSN = "postgres://yovision_t012_sense@127.0.0.1:$port/${databaseName}?sslmode=disable"

    Invoke-Checked $psql '-X' '-v' 'ON_ERROR_STOP=1' '-d' $adminRootDSN '-f' (Join-Path $repoRoot 'deploy\postgres\001_roles.sql')
    Invoke-Checked $createdb '-h' '127.0.0.1' '-p' ([string]$port) '-U' 'postgres' $databaseName
@@ -85,12 +85,14 @@ try {
            '006_device_operation_outbox.sql',
            '007_privileges_area_audit.sql',
            '008_control_api.sql',
-            '009_privileges_control_api.sql'
+            '009_privileges_control_api.sql',
+            '010_reconcile_safety.sql',
+            '011_privileges_reconcile_safety.sql'
        )) {
            Invoke-Checked $psql '-X' '-v' 'ON_ERROR_STOP=1' '-d' $adminDatabaseDSN '-f' (Join-Path $repoRoot "deploy\postgres\$name")
        }
    }
-    Invoke-Checked $psql '-X' '-v' 'ON_ERROR_STOP=1' '-d' $adminRootDSN '-c' 'CREATE ROLE yovision_t011_sense LOGIN IN ROLE sense_app'
+    Invoke-Checked $psql '-X' '-v' 'ON_ERROR_STOP=1' '-d' $adminRootDSN '-c' 'CREATE ROLE yovision_t012_sense LOGIN IN ROLE sense_app'
    Invoke-Checked $psql '-X' '-v' 'ON_ERROR_STOP=1' '-d' $adminDatabaseDSN '-f' (Join-Path $repoRoot 'deploy\postgres\tests\assertions.sql')

    $env:YOVISION_TEST_POSTGRES_DSN = $senseDSN
@@ -105,7 +107,7 @@ finally {
    if (-not $started -or $stopped) {
        $resolvedData = [IO.Path]::GetFullPath($dataDir)
        if (-not $resolvedData.StartsWith($systemTemp, [StringComparison]::OrdinalIgnoreCase) -or
-            [IO.Path]::GetFileName($resolvedData) -notlike 'yovision-t011-pg-*') {
+            [IO.Path]::GetFileName($resolvedData) -notlike 'yovision-t012-pg-*') {
            throw "Refusing to clean unexpected temporary path."
        }
        if (Test-Path -LiteralPath $resolvedData) {
@@ -56,6 +56,8 @@ class PostgresContractTests(unittest.TestCase):
                "007_privileges_area_audit.sql",
                "008_control_api.sql",
                "009_privileges_control_api.sql",
+                "010_reconcile_safety.sql",
+                "011_privileges_reconcile_safety.sql",
            ],
            names,
        )
@@ -109,7 +111,7 @@ class PostgresContractTests(unittest.TestCase):
            "initdb.exe",
            "pg_ctl.exe",
            "127.0.0.1",
-            "yovision-t011-pg-",
+            "yovision-t012-pg-",
            "YOVISION_TEST_POSTGRES_DSN",
            "Get-NetTCPConnection",
        ):
@@ -0,0 +1,86 @@
+"""Static safety checks for the T-012 reconciliation/orphan boundary."""
+
+from __future__ import annotations
+
+import pathlib
+import re
+import unittest
+
+
+ROOT = pathlib.Path(__file__).resolve().parents[1]
+
+
+def text(relative: str) -> str:
+    return (ROOT / relative).read_text(encoding="utf-8")
+
+
+def normalized(value: str) -> str:
+    return re.sub(r"\s+", " ", value.lower())
+
+
+class SenseReconcileSafetyTests(unittest.TestCase):
+    def test_postgres_v5_has_fencing_ownership_and_private_reports(self) -> None:
+        schema = normalized(text("deploy/postgres/010_reconcile_safety.sql"))
+        privileges = normalized(
+            text("deploy/postgres/011_privileges_reconcile_safety.sql")
+        )
+        implementation = normalized(text("Sense/internal/store/postgres.go"))
+        for table in (
+            "media_path_ownership",
+            "operational_leases",
+            "orphan_scan_runs",
+            "orphan_scan_findings",
+            "orphan_cleanup_actions",
+        ):
+            self.assertIn(f"sense.{table}", schema)
+            self.assertIn(f"sense.{table}", privileges)
+        self.assertIn("insert into sense.schema_migrations(version) values (5)", schema)
+        self.assertIn("clock_timestamp()", implementation)
+        self.assertIn("for update of r skip locked", implementation)
+        self.assertIn("and lease_token = $5", implementation)
+        self.assertNotIn("endpoint_ref", schema)
+        self.assertNotIn("credential_ref", schema)
+        self.assertNotIn("source_uri", schema)
+
+    def test_cleanup_gate_has_no_force_or_unowned_delete_path(self) -> None:
+        source = text("Sense/internal/orphan/orphan.go")
+        normalized_source = normalized(source)
+        self.assertIn("candidates*100 > observed*10", normalized_source)
+        self.assertIn("candidates > maxcleanupitems", normalized_source)
+        self.assertIn('confirmation != "delete "+scanid', normalized_source)
+        self.assertIn(
+            "finding.classification != store.orphanownedstale", normalized_source
+        )
+        self.assertNotRegex(normalized_source, r"\bforce\b|bypass")
+        migration = normalized(text("deploy/postgres/010_reconcile_safety.sql"))
+        self.assertIn("classification = 'owned_stale'", migration)
+
+    def test_mediamtx_inventory_discards_sources_and_bounds_pagination(self) -> None:
+        source = text("Sense/internal/mtx/client.go")
+        self.assertIn("ListPathNames", source)
+        self.assertIn("maxPages", source)
+        self.assertIn("list paths repeated page", source)
+        inventory = source[source.index("func (c *Client) ListPathNames") :]
+        inventory = inventory[: inventory.index("func (c *Client) CreatePath")]
+        self.assertNotIn("item.Source", inventory)
+
+    def test_metrics_are_fixed_and_control_contract_is_untouched(self) -> None:
+        metrics = text("Sense/internal/metrics/metrics.go")
+        main = text("Sense/cmd/sense-api/main.go")
+        self.assertIn('mux.Handle("GET /metrics", registry.Handler())', main)
+        for forbidden in ("tenant_id", "site_id", "device_id", "path_name"):
+            self.assertNotIn(forbidden, metrics)
+        control = text("docs/contracts/sense-control-v1.openapi.json")
+        self.assertEqual(7, len(re.findall(r'"operationId"\s*:', control)))
+
+    def test_operator_command_requires_fresh_scan_confirmation(self) -> None:
+        command = text("Sense/cmd/sense-orphan/main.go")
+        manager = text("Sense/internal/orphan/orphan.go")
+        self.assertIn("exact confirmation: DELETE <scan-id>", command)
+        self.assertIn("scan.ExpiresAt.After(now)", manager)
+        self.assertIn("cleanupCandidates(scan, paths, ownership)", manager)
+        self.assertIn("RecordOrphanCleanup", manager)
+
+
+if __name__ == "__main__":
+    unittest.main()