feat: 接入 Instance 原生管理观测与删除保护

cmd/database.go

@@ -0,0 +1,26 @@
+package main
+
+import (
+	"git.ddupan.top/panxiao81/ayatori/internal/database/adapter/kubernetes"
+	"git.ddupan.top/panxiao81/ayatori/internal/database/adapter/postgresql"
+	"git.ddupan.top/panxiao81/ayatori/internal/database/application"
+	databasecontroller "git.ddupan.top/panxiao81/ayatori/internal/database/controller"
+	ctrl "sigs.k8s.io/controller-runtime"
+)
+
+func setupInstanceObservation(manager ctrl.Manager, namespace, rootCert string) (*application.InstanceService, error) {
+	credentials, err := kubernetes.NewSecretCredentials(manager.GetConfig(), namespace)
+	if err != nil {
+		return nil, err
+	}
+	service, err := application.NewInstanceService(credentials, postgresql.Connector{RootCert: rootCert})
+	if err != nil {
+		return nil, err
+	}
+	reconciler := &databasecontroller.InstanceReconciler{Observer: service, SecretNamespace: namespace}
+	if err := reconciler.SetupWithManager(manager); err != nil {
+		service.Close()
+		return nil, err
+	}
+	return service, nil
+}

cmd/main.go

@@ -22,6 +22,7 @@ import (
 
 	databasev1alpha1 "git.ddupan.top/panxiao81/ayatori/api/database/v1alpha1"
 	executionv1alpha1 "git.ddupan.top/panxiao81/ayatori/api/execution/v1alpha1"
+	"git.ddupan.top/panxiao81/ayatori/internal/database/application"
 	databasecontroller "git.ddupan.top/panxiao81/ayatori/internal/database/controller"
 	// +kubebuilder:scaffold:imports
 )
@@ -41,6 +42,10 @@ func init() {
 
 // nolint:gocyclo
 func main() {
+	var databaseNamespace, databaseRootCert string
+	flag.StringVar(&databaseNamespace, "database-secret-namespace", os.Getenv("POD_NAMESPACE"),
+		"固定管理 Secret namespace；为空时不启用 Instance 观测")
+	flag.StringVar(&databaseRootCert, "database-root-cert", "", "PostgreSQL 管理连接信任的公开 CA bundle 路径")
 	var metricsAddr string
 	var metricsCertPath, metricsCertName, metricsCertKey string
 	var webhookCertPath, webhookCertName, webhookCertKey string
@@ -145,7 +150,7 @@ func main() {
 		metricsServerOptions.KeyName = metricsCertKey
 	}
 
-	mgr, err := ctrl.NewManager(ctrl.GetConfigOrDie(), ctrl.Options{
+	managerOptions := ctrl.Options{
 		Scheme:                 scheme,
 		Metrics:                metricsServerOptions,
 		WebhookServer:          webhookServer,
@@ -163,13 +168,25 @@ func main() {
 		// if you are doing or is intended to do any operation such as perform cleanups
 		// after the manager stops then its usage might be unsafe.
 		// LeaderElectionReleaseOnCancel: true,
-	})
+	}
+	if databaseNamespace != "" {
+		managerOptions.Cache = databasecontroller.InstanceCacheOptions(databaseNamespace)
+	}
+	mgr, err := ctrl.NewManager(ctrl.GetConfigOrDie(), managerOptions)
 	if err != nil {
 		setupLog.Error(err, "Failed to start manager")
 		os.Exit(1)
 	}
 
 	// +kubebuilder:scaffold:builder
+	var instanceService *application.InstanceService
+	if databaseNamespace != "" {
+		instanceService, err = setupInstanceObservation(mgr, databaseNamespace, databaseRootCert)
+		if err != nil {
+			setupLog.Error(err, "Failed to set up Instance observation")
+			os.Exit(1)
+		}
+	}
 	if err := (&databasecontroller.BindingReconciler{}).SetupWithManager(context.Background(), mgr); err != nil {
 		setupLog.Error(err, "Failed to set up Database binding controller")
 		os.Exit(1)
@@ -185,7 +202,12 @@ func main() {
 	}
 
 	setupLog.Info("Starting manager")
-	if err := mgr.Start(ctrl.SetupSignalHandler()); err != nil {
+	err = mgr.Start(ctrl.SetupSignalHandler())
+	// worker 完全停止后才释放 pgxpool，避免与在途观察竞争。
+	if instanceService != nil {
+		instanceService.Close()
+	}
+	if err != nil {
 		setupLog.Error(err, "Failed to run manager")
 		os.Exit(1)
 	}

config/manager/manager.yaml

@@ -65,6 +65,11 @@ spec:
           - --health-probe-bind-address=:8081
         image: controller:latest
         name: manager
+        env:
+        - name: POD_NAMESPACE
+          valueFrom:
+            fieldRef:
+              fieldPath: metadata.namespace
         ports:
         - containerPort: 8081
           name: health

config/rbac/database_credentials_role.yaml

@@ -0,0 +1,23 @@
+apiVersion: rbac.authorization.k8s.io/v1
+kind: Role
+metadata:
+  name: database-management-credentials
+  namespace: system
+rules:
+- apiGroups: [""]
+  resources: [secrets]
+  verbs: [get, list, watch]
+---
+apiVersion: rbac.authorization.k8s.io/v1
+kind: RoleBinding
+metadata:
+  name: database-management-credentials
+  namespace: system
+roleRef:
+  apiGroup: rbac.authorization.k8s.io
+  kind: Role
+  name: database-management-credentials
+subjects:
+- kind: ServiceAccount
+  name: controller-manager
+  namespace: system

config/rbac/kustomization.yaml

@@ -7,6 +7,7 @@ resources:
 - service_account.yaml
 - role.yaml
 - role_binding.yaml
+- database_credentials_role.yaml
 - leader_election_role.yaml
 - leader_election_role_binding.yaml
 # The following RBAC configurations are used to protect

config/rbac/role.yaml

@@ -19,6 +19,7 @@ rules:
   - database.ayatori.ddupan.top
   resources:
   - postgresqldatabases/finalizers
+  - postgresqlinstances/finalizers
   - postgresqltenants/finalizers
   verbs:
   - update
@@ -26,6 +27,7 @@ rules:
   - database.ayatori.ddupan.top
   resources:
   - postgresqldatabases/status
+  - postgresqlinstances/status
   - postgresqltenants/status
   verbs:
   - get
@@ -35,13 +37,6 @@ rules:
   - database.ayatori.ddupan.top
   resources:
   - postgresqlinstances
-  verbs:
-  - get
-  - list
-  - watch
-- apiGroups:
-  - database.ayatori.ddupan.top
-  resources:
   - postgresqltenants
   verbs:
   - get

docs/database/README.md

@@ -1,7 +1,7 @@
 # Database 模块
 
-Database 是 Ayatori 首批实际产品领域之一。当前已包含 Instance 领域基础、管理凭据连接与
+Database 是 Ayatori 首批实际产品领域之一。当前已包含三资源 API、分层绑定与 Instance 原生
-metadata 观察切片，尚未完成 Database API/controller 和 Tenant 供应链路。
+管理能力观测；尚未完成 Database 供应/导入、Tenant 凭据交付与资源回收链路。
 
 ## 当前设计（2026-09-24）
 
@@ -12,7 +12,7 @@ Retain 后人工重新绑定与资源侧 Delete。撤销 PostgreSQL ownership re
 依据 [ADR-0009](../decisions/0009-database-resource-and-claim.md)，当前合同见
 [系统规格](specification.md)。下面的迁移来源与已存在代码不反向约束新设计。
 registry adapter、专属迁移/测试及 Instance 的 registry 判定现已撤除；Instance 根据完整管理
-能力观察直接判定 Ready。新增 Database 资源、绑定、导入、角色/凭据管理边界与回收链路尚未实现。wiki 同步位置见
+能力观察直接判定 Ready。Database 资源与绑定已接入，导入、角色/凭据供应及回收仍未完成。wiki 同步位置见
 `homelab-wiki/services/postgresql-tenant-operator.md`，跨仓库发布状态由 wiki 的同步记录维护。
 
 ## 来源基线
@@ -59,10 +59,10 @@ Secret metadata 和无关字段变化不重建连接。观测后再次读取 Sec
 不把旧连接的成功作为新凭据有效的证据；这不构成跨 Kubernetes/PostgreSQL 的原子事务。
 Instance UID、endpoint 或凭据引用变化也会释放旧连接；Forget/Close 只释放本地资源。
 
-当前通过 `ObserveMetadata` 读取服务器版本和可用扩展，`ObserveVersion` 只是其版本读取便捷入口，
+`ObserveMetadata` 读取服务器版本和可用扩展，`ObserveVersion` 是版本读取便捷入口；两者
-不能产生完整 CapabilityObservation 或 Ready。
+不会填充管理检查，不能产生 Ready。`ObserveManagement` 使用同一凭据/连接边界读取完整
-controller 接入、Secret watch、finalizer 与真实权限检查仍待后续切片；并发 CR 更新
+原生管理检查，返回绑定当前 target 的 `InstanceObservation`。controller 的资源呈现适配器
-必须由调用者通过 resourceVersion 校验。应用层沿用源实现的串行处理，本阶段未引入新的调度框架。
+使用 resourceVersion 拒绝过期写入。应用层沿用串行处理，不引入新的调度框架。
 
 运行 `make test-database-integration` 验证真实 API server + 一次性 PostgreSQL；fixture 不接受外部
 DSN，镜像固定摘要，使用随机本机回环端口并在退出时删除测试容器。覆盖缺失/错误凭据、RBAC、
@@ -83,14 +83,13 @@ SQL adapter 通过一条只读语句读取 `pg_catalog.current_setting('server_v
 
 `InstanceService` 保留原有 CredentialReader → Connector → Database 边界，复用同一个
 凭据读取、连接刷新和串行释放流程，不新增连接池封装或任意查询回调。每次重新查询 metadata，
-并在 Secret 有效值回读一致后生成不可变的 `MetadataObservation`，绑定本次 target（含当前
+并在 Secret 有效值回读一致后生成不可变的 `InstanceObservation`，绑定本次 target（含当前
 generation），不绑定建池时的旧 target。结果不包含凭据，扩展集合不与 driver 的可变 slice 共享。
 凭据中途变化、读取失败或查询失败时，返回零值观察并释放连接，不复用旧的扩展列表。
 
 调用方可将 `Target()` 与 `Extensions()` 交给 Instance 的 `ObserveExtensions`；应用调用链
 仍负责同轮次使用，不能持久化或跨轮缓存这份证据。metadata 读取不安装扩展、不初始化 registry、
-不设置 Ready，也不授予 Tenant 写权限。管理权限矩阵及 controller 的
+不设置 Ready，也不授予 Tenant 写权限。管理权限检查使用下面的独立入口。
-checkpoint/status/finalizer 链路仍是后续切片。
 
 真实 API server + PostgreSQL 测试验证未安装扩展可被观察、名称保持大小写、search_path 遮蔽
 不改变查询来源、低权限账号读取、权限撤回失败与恢复、Secret 中途变化丢弃扩展结果。
@@ -106,7 +105,51 @@ Instance 不再具有 InitializingRegistry 阶段、RegistryState、准备决策
 
 保留凭据读取与连接刷新、TLS、metadata/扩展观察及其真实后端测试。领域测试覆盖每项能力
 在初次验证和 Ready 重验时失败、依赖恢复、重启后重新取证、错误目标/阶段及删除保护。
-这不等于管理权限探测矩阵或三资源 controller 已实现。
+这不等于三资源供应、交付和回收已实现。
+
+## Instance 原生管理观测
+
+2026-09-25 维护者确认先使用原生非 superuser 方案，不引入 SECURITY DEFINER 接口。
+`InspectManagement` 通过同一条只读语句读取当前执行角色的 `CREATEROLE`、`CREATEDB`、
+superuser 属性、服务器可写状态、版本和扩展列表。不创建探针数据库/角色，不初始化 schema。
+只有非 superuser、具备两项原生属性且当前服务器/会话可写时，基础管理能力才通过。
+角色属性不可从继承成员关系推导；具体已有资源仍须检查 owner、membership 与授权范围。
+
+权限依据和真实测试对应：
+
+- role：当前角色具有 CREATEROLE，可创建普通登录角色；
+- database：当前角色具有 CREATEDB，且会话非只读、服务器不在 recovery；
+- grant：使用自己新建角色的管理权限，显式建立 SET membership，再以 owner 管理数据库 ACL；
+- extension：新建数据库 owner 可安装 trusted 扩展；可用列表不是安装授权，非 trusted 或
+  其他前提不满足的扩展仍可能失败，必须逐请求执行和回读。导入不继承此动态供应授权。
+
+参考 PostgreSQL 官方 [CREATE ROLE](https://www.postgresql.org/docs/18/sql-createrole.html)、
+[CREATE DATABASE](https://www.postgresql.org/docs/18/sql-createdatabase.html) 与
+[CREATE EXTENSION](https://www.postgresql.org/docs/18/sql-createextension.html)。这些检查是基础
+能力观察，不是未来操作必然成功的保证；权限、容量、连接数等仍可能在执行时变化。
+
+`InstanceReconciliation` 协调 finalizer、观察、领域判定和删除引用检查，controller 只连接
+事件、用例、状态呈现与重试。每轮重建无证据的领域对象；旧 Ready 不授权新一轮操作。
+失败清除当前版本结果并撤销 Ready；CR 在观察期间被修改则拒绝旧结果，下一轮重新读取。
+连接/权限变化由 Secret watch 和 30 秒重查驱动，单轮 IO 最长 15 秒；不持续写入相同状态。
+
+manager 通过 `--database-secret-namespace`（默认 `POD_NAMESPACE`）启用 Instance 观测；
+为空时不启用。本地运行需显式提供该参数。Deployment 使用 downward API 获取自身 namespace；
+Secret 的 get/list/watch 权限由该 namespace 的 Role 单独授予，不放入 ClusterRole。
+watch 使用 controller-runtime 的 metadata-only cache，读取有效凭据仍直连 API server。
+TLS 使用既有 endpoint 合同，公开 CA bundle 可由 `--database-root-cert` 指定；不会自动挂载
+生产证书或创建管理 Secret。manager worker 停止后统一关闭 pgxpool。
+
+Instance 删除首先释放本地连接并撤销 Ready。任何引用它的 Database（含 Released、删除中）
+或动态 Tenant 申请都会阻止 finalizer 解除；列表查询失败也等待。仅在引用全部解除后移除
+`database.ayatori.ddupan.top/instance-protection`，不删除 PostgreSQL、账号或凭据。
+引用查询与删除不是跨对象事务；后续供应仍必须拒绝已删除/删除中的 Instance。
+
+验收使用真实 PostgreSQL + API server：原生管理账号实际建库、owner 授权、trusted 扩展
+安装/回读，拒绝非 trusted 扩展；权限撤回/恢复、只读会话、superuser 拒绝和中途轮换。
+实际 manager 在生成的资源 RBAC 和 namespaced Secret Role 下验证缺失 Secret 后出现、
+轮换、删除、跨 namespace 拒绝和 watch。API 测试另覆盖写入版本冲突、幂等、新 reconciler
+恢复与引用删除保护。完整 DBaaS 仍需供应/导入、OpenBao/ESO、Retain/Delete 集成验收。
 
 ## 设计入口
 

docs/database/api-reference.md

@@ -53,7 +53,9 @@ Tenant 进入 `status.phase=Binding` 后由 CEL 固定申请目标；Database
 绑定顺序由 application service 协调，纯资格规则在领域层；Kubernetes adapter 负责快照
 映射、finalizer 和状态呈现。呈现前若资源版本已变化，返回冲突供下一轮重读，不覆盖其他修改。
 双向记录完成后 Tenant 为 Bound，Ready=False/BindingComplete，明确尚未供应或交付。
-生成的 manager RBAC 仅授予绑定所需资源读写，不包含 Secret 读取或后端凭据权限。
+生成的 manager ClusterRole 授予资源读写，不包含 Secret 读取；Instance 观测的管理 Secret
+权限由固定 namespace 的独立 Role 授予。Instance controller 已接入原生管理观察与引用删除
+保护，启用方式及 Ready 边界见 [模块说明](README.md#instance-原生管理观测)。
 
 当前有 Tenant/Database finalizer 保护，但**删除清理尚未实现**：Tenant 删除报告
 Ready=False/DeletionPending 并保留绑定与 finalizer，Database 的保护也不会被自动移除。

docs/database/domain-instance.md

@@ -1,8 +1,8 @@
 # Instance 领域对象规格
 
-日期：2026-09-24。资源模型修订依据
+日期：2026-09-25。资源模型修订依据
 [ADR-0009](../decisions/0009-database-resource-and-claim.md)，行为以
-[系统规格](specification.md) 为准。本页替代原 registry 准备与恢复合同；领域依赖已撤除，完整应用/controller 链路尚未接入。
+[系统规格](specification.md) 为准。本页替代原 registry 准备与恢复合同；领域依赖已撤除，Instance 应用/controller 观测链路已接入。
 
 ## 职责
 
@@ -85,4 +85,6 @@ finalizer 不阻止并发申请 CR 创建；新请求见 Instance 删除中/不
 
 Instance 领域代码、adapter 与测试的 registry 依赖已撤除。AssessManagement 根据完整观察
 直接完成验证；AssessReadiness 失败进入 Validating，依赖恢复后重新验证。领域测试覆盖各检查项
-在这两个入口的失败与恢复，但完整权限探测、Instance controller 和三资源生命周期尚未完成。
+在这两个入口的失败与恢复。原生权限检查、Instance controller、metadata-only Secret watch
+和引用删除保护已接入，验证矩阵见 [模块说明](README.md#instance-原生管理观测)；
+Database/Tenant 的供应、交付和回收仍未完成。

docs/database/operations.md

@@ -14,6 +14,13 @@ Database 的 `database.ayatori.ddupan.top/database-protection` 也尚无清理
 
 ## 日常检查
 
+Instance 观察已实现：先确认 manager 配置了 `--database-secret-namespace` 或 `POD_NAMESPACE`，
+再检查 Ready Reason、observedGeneration 与管理 Secret 名称/字段映射，切勿导出其 data。
+`InsufficientPrivileges` 表示当前原生方案要求的非 superuser、CREATEDB/CREATEROLE 不满足；
+`CredentialsChanged` 会丢弃中途轮换的结果并重验；`InstanceInUse` 消息定位阻塞删除的资源。
+Secret 事件立即入队，30 秒重查覆盖 PostgreSQL 权限等没有 Kubernetes 事件的外部变化。
+Instance 删除不要求 PostgreSQL 可达，但必须可读取所有 Database/Tenant 引用。
+
 先看 Instance、Database、Tenant 的 Ready Condition、绑定 UID、阶段与 observedGeneration，
 再核对 PostgreSQL catalog、OpenBao metadata、ExternalSecret 与 Secret 投射状态。
 具体 kubectl 资源名、finalizer 名称与人工确认字段在 API 实现后补齐，不提供猜测的 patch 命令。

docs/database/security.md

@@ -3,7 +3,7 @@
 | 项目 | 内容 |
 | --- | --- |
 | 状态 | Review |
-| 最后更新 | 2026-09-24 |
+| 最后更新 | 2026-09-25 |
 
 ## 保护目标
 
@@ -47,9 +47,11 @@ ESO 身份只读管理路径，租户 ESO 身份只读 tenant base path，二者
 使用管理凭据 Store。controller 对管理 Secret 的读取限于自身 namespace，Instance
 不能指定其他 namespace；controller 不创建或修改管理 Secret／ExternalSecret。
 
-PostgreSQL 管理 role 不应是 superuser。若平台选择 SECURITY DEFINER 函数承载创建或
+2026-09-25 维护者确认第一版使用原生非 superuser 管理 role，具有 CREATEDB/CREATEROLE，
-删除操作，函数必须固定 `search_path`、严格校验 identifier、拒绝任意 SQL，并仅向
+不引入 SECURITY DEFINER 接口。Instance 检查拒绝 superuser；具体已有资源的 owner 和
-controller role 授予 EXECUTE。controller 不调用 shell 或 `psql` 拼接用户输入。
+membership 仍需逐资源验证，不能把基础能力用于接管他人资源。扩展按实际权限安装，
+不因可用列表包含某个扩展就默认能安装它。controller 不调用 shell 或 `psql` 拼接用户输入。
+当前检查与真实权限矩阵见 [Instance 原生管理观测](README.md#instance-原生管理观测)。
 
 Kubernetes RBAC 应把 Instance 管理、Database 导入、Released 重新开放和回收限制给平台管理员。
 有权创建 Tenant 的申请者可显式申请未绑定且可用的 Database，不增加资源侧允许绑定名单

docs/database/specification.md

@@ -171,6 +171,9 @@ status 缺失不假定发生于正常重启。Instance 可重新探测能力；D
 
 ## 9. 权限、凭据与扩展
 
+2026-09-25 确认第一版管理账号使用原生非 superuser + CREATEDB/CREATEROLE 方案，
+不引入 SECURITY DEFINER 接口；权限检查与限制见 [安全合同](security.md#最小权限)。
+
 动态供应继续使用一个兼任 database owner 的 LOGIN role；应用角色不得具备 superuser、
 CREATEDB、CREATEROLE 或 replication 权限。撤销 PUBLIC CONNECT，再授予目标角色；
 不修改无关数据库和角色。identifier 匹配 `^[a-z][a-z0-9_]{0,62}$`，SQL 安全引用。

internal/database/adapter/kubernetes/instance_mapping.go

@@ -0,0 +1,45 @@
+package kubernetes
+
+import (
+	databasev1alpha1 "git.ddupan.top/panxiao81/ayatori/api/database/v1alpha1"
+	"git.ddupan.top/panxiao81/ayatori/internal/database/application"
+	"git.ddupan.top/panxiao81/ayatori/internal/database/domain/instance"
+)
+
+func instanceRecord(object *databasev1alpha1.PostgreSQLInstance) (*application.InstanceRecord, error) {
+	identity, err := instance.NewIdentity(string(object.UID), object.Name)
+	if err != nil {
+		return nil, err
+	}
+	revision, err := instance.NewRevision(object.Generation)
+	if err != nil {
+		return nil, err
+	}
+	spec := object.Spec
+	endpoint, err := instance.NewEndpoint(instance.EndpointValues{
+		Host: spec.Endpoint.Host, HostAddr: spec.Endpoint.HostAddr,
+		Port: int(spec.Endpoint.Port), ManagementDatabase: string(spec.Endpoint.Database),
+		TLSMode: instance.TLSMode(spec.Endpoint.SSLMode),
+	})
+	if err != nil {
+		return nil, err
+	}
+	credential, err := instance.NewCredentialReference(instance.CredentialReferenceValues{
+		Name:        string(spec.AdminCredentialRef.Name),
+		UsernameKey: spec.AdminCredentialRef.UsernameKey, PasswordKey: spec.AdminCredentialRef.PasswordKey,
+	})
+	if err != nil {
+		return nil, err
+	}
+	definition, err := instance.NewDefinition(endpoint, credential)
+	if err != nil {
+		return nil, err
+	}
+	target, err := instance.NewObservationTarget(identity, revision, definition)
+	if err != nil {
+		return nil, err
+	}
+	return &application.InstanceRecord{
+		Target: target, Revision: object.ResourceVersion, Deleting: !object.DeletionTimestamp.IsZero(),
+	}, nil
+}

internal/database/adapter/kubernetes/instance_resources.go

@@ -0,0 +1,112 @@
+package kubernetes
+
+import (
+	"context"
+	"errors"
+
+	databasev1alpha1 "git.ddupan.top/panxiao81/ayatori/api/database/v1alpha1"
+	"git.ddupan.top/panxiao81/ayatori/internal/database/application"
+	"git.ddupan.top/panxiao81/ayatori/internal/database/domain/instance"
+	"k8s.io/apimachinery/pkg/api/equality"
+	apierrors "k8s.io/apimachinery/pkg/api/errors"
+	"k8s.io/apimachinery/pkg/api/meta"
+	metav1 "k8s.io/apimachinery/pkg/apis/meta/v1"
+	"sigs.k8s.io/controller-runtime/pkg/client"
+	"sigs.k8s.io/controller-runtime/pkg/controller/controllerutil"
+)
+
+const InstanceFinalizer = "database.ayatori.ddupan.top/instance-protection"
+
+type InstanceResources struct {
+	Client client.Client
+	Reader client.Reader
+}
+
+func (r *InstanceResources) LoadInstance(ctx context.Context, name string) (*application.InstanceRecord, error) {
+	object := &databasev1alpha1.PostgreSQLInstance{}
+	if err := r.Reader.Get(ctx, client.ObjectKey{Name: name}, object); err != nil {
+		return nil, client.IgnoreNotFound(err)
+	}
+	return instanceRecord(object)
+}
+
+func (r *InstanceResources) ProtectInstance(ctx context.Context, record *application.InstanceRecord) (*application.InstanceRecord, error) {
+	object, err := r.instanceAtVersion(ctx, record)
+	if err != nil {
+		return nil, err
+	}
+	if controllerutil.AddFinalizer(object, InstanceFinalizer) {
+		if err := r.Client.Update(ctx, object); err != nil {
+			return nil, err
+		}
+	}
+	return instanceRecord(object)
+}
+
+func (r *InstanceResources) InstanceReferences(ctx context.Context, name string) (string, error) {
+	// 删除判断必须直读 API；包含 Released、删除中的 Database 和尚未绑定的申请。
+	// 不按旧 Instance UID 忽略引用，也不依赖 informer 索引的及时性。
+	databases := &databasev1alpha1.PostgreSQLDatabaseList{}
+	if err := r.Reader.List(ctx, databases); err != nil {
+		return "", err
+	}
+	for _, database := range databases.Items {
+		if string(database.Spec.InstanceRef.Name) == name {
+			return "Database/" + database.Name, nil
+		}
+	}
+	tenants := &databasev1alpha1.PostgreSQLTenantList{}
+	if err := r.Reader.List(ctx, tenants); err != nil {
+		return "", err
+	}
+	for _, tenant := range tenants.Items {
+		if tenant.Spec.Provision != nil && string(tenant.Spec.Provision.InstanceRef.Name) == name {
+			return "Tenant/" + tenant.Namespace + "/" + tenant.Name, nil
+		}
+	}
+	return "", nil
+}
+
+func (r *InstanceResources) PresentInstance(ctx context.Context, result application.InstanceResult) error {
+	if result.Record == nil {
+		return nil
+	}
+	object, err := r.instanceAtVersion(ctx, result.Record)
+	if err != nil {
+		return client.IgnoreNotFound(err)
+	}
+	previous := object.Status.DeepCopy()
+	object.Status.Phase = string(result.Snapshot.Phase)
+	object.Status.ObservedGeneration = object.Generation
+	object.Status.PostgreSQLVersion = result.Snapshot.ReportedVersion
+	ready := metav1.ConditionFalse
+	if result.Snapshot.Readiness == instance.Ready {
+		ready = metav1.ConditionTrue
+	}
+	meta.SetStatusCondition(&object.Status.Conditions, metav1.Condition{
+		Type: "Ready", Status: ready, ObservedGeneration: object.Generation,
+		Reason: result.Reason, Message: result.Message,
+	})
+	if !equality.Semantic.DeepEqual(*previous, object.Status) {
+		if err := r.Client.Status().Update(ctx, object); err != nil {
+			return err
+		}
+	}
+	if result.RemoveProtection && controllerutil.RemoveFinalizer(object, InstanceFinalizer) {
+		return r.Client.Update(ctx, object)
+	}
+	return nil
+}
+
+func (r *InstanceResources) instanceAtVersion(ctx context.Context, record *application.InstanceRecord) (*databasev1alpha1.PostgreSQLInstance, error) {
+	object := &databasev1alpha1.PostgreSQLInstance{}
+	name := record.Target.Identity().Name()
+	if err := r.Reader.Get(ctx, client.ObjectKey{Name: name}, object); err != nil {
+		return nil, err
+	}
+	if string(object.UID) != record.Target.Identity().UID() || object.ResourceVersion != record.Revision {
+		return nil, apierrors.NewConflict(databasev1alpha1.GroupVersion.WithResource("postgresqlinstances").GroupResource(),
+			name, errors.New("Instance 快照已过期，请重新观察"))
+	}
+	return object, nil
+}

internal/database/adapter/postgresql/fixture_integration_test.go

@@ -31,6 +31,7 @@ import (
 	corev1 "k8s.io/api/core/v1"
 	metav1 "k8s.io/apimachinery/pkg/apis/meta/v1"
 	"k8s.io/client-go/kubernetes"
+	"k8s.io/client-go/rest"
 	"sigs.k8s.io/controller-runtime/pkg/envtest"
 
 	secretadapter "git.ddupan.top/panxiao81/ayatori/internal/database/adapter/kubernetes"
@@ -40,15 +41,19 @@ import (
 )
 
 const (
-	fixtureHost         = "fixture.invalid"
+	fixtureAddress        = "127.0.0.1"
-	fixtureUser         = "postgres"
+	managementUsernameKey = "login"
-	fixtureExtension    = "plpgsql"
+	managementPasswordKey = "credential"
-	dockerExec          = "exec"
+	unrelatedNamespace    = "unrelated"
-	fixtureImage        = "postgres@sha256:18cfe3ef5e6815560c98237d6216d1e5119702fb0f3894c8785dd58b8bbe5d73"
+	fixtureHost           = "fixture.invalid"
-	fixturePassword     = "AYATORI-TEST-ONLY-initial-password"
+	fixtureUser           = "postgres"
-	rotatedPassword     = "AYATORI-TEST-ONLY-rotated-password"
+	fixtureExtension      = "plpgsql"
-	controllerNamespace = "database-controller"
+	dockerExec            = "exec"
-	secretName          = "management"
+	fixtureImage          = "postgres@sha256:18cfe3ef5e6815560c98237d6216d1e5119702fb0f3894c8785dd58b8bbe5d73"
+	fixturePassword       = "AYATORI-TEST-ONLY-initial-password"
+	rotatedPassword       = "AYATORI-TEST-ONLY-rotated-password"
+	controllerNamespace   = "database-controller"
+	secretName            = "management"
 )
 
 // fixture 不接受外部 DSN，只创建自己的临时容器并按确切 ID 清理。
@@ -79,7 +84,7 @@ func postgresFixture(t *testing.T, ctx context.Context) (string, int) {
 		t.Fatal("invalid fixture port")
 	}
 	// 初次 init 的临时服务器只监听 Unix socket，必须等最终 TCP listener。
-	for exec.CommandContext(ctx, "docker", dockerExec, id, "pg_isready", "-h", "127.0.0.1", "-U", fixtureUser).Run() != nil {
+	for exec.CommandContext(ctx, "docker", dockerExec, id, "pg_isready", "-h", fixtureAddress, "-U", fixtureUser).Run() != nil {
 		select {
 		case <-ctx.Done():
 			t.Fatal("fixture startup timed out")
@@ -154,7 +159,7 @@ func target(t *testing.T, port int, mode instance.TLSMode) instance.ObservationT
 	}
 	endpoint, err := instance.NewEndpoint(instance.EndpointValues{
 		Host:               fixtureHost,
-		HostAddr:           "127.0.0.1",
+		HostAddr:           fixtureAddress,
 		Port:               port,
 		ManagementDatabase: fixtureUser,
 		TLSMode:            mode,
@@ -164,8 +169,8 @@ func target(t *testing.T, port int, mode instance.TLSMode) instance.ObservationT
 	}
 	ref, err := instance.NewCredentialReference(instance.CredentialReferenceValues{
 		Name:        secretName,
-		UsernameKey: "login",
+		UsernameKey: managementUsernameKey,
-		PasswordKey: "credential",
+		PasswordKey: managementPasswordKey,
 	})
 	if err != nil {
 		t.Fatal(err)
@@ -196,6 +201,7 @@ func (r *gatedReader) Read(ctx context.Context, ref instance.CredentialReference
 
 // credentialFixture 为每个场景创建独立 API server、PostgreSQL 和应用服务。
 type credentialFixture struct {
+	config       *rest.Config
 	ctx          context.Context
 	client       *kubernetes.Clientset
 	reader       *secretadapter.SecretCredentials
@@ -227,7 +233,7 @@ func newCredentialFixture(t *testing.T) *credentialFixture {
 	if err != nil {
 		t.Fatal("cannot create test client")
 	}
-	for _, namespace := range []string{controllerNamespace, "unrelated"} {
+	for _, namespace := range []string{controllerNamespace, unrelatedNamespace} {
 		_, err := client.CoreV1().Namespaces().Create(
 			ctx,
 			&corev1.Namespace{Name: namespace},
@@ -260,6 +266,7 @@ func newCredentialFixture(t *testing.T) *credentialFixture {
 	t.Cleanup(service.Close)
 
 	return &credentialFixture{
+		config:       config,
 		ctx:          ctx,
 		client:       client,
 		reader:       reader,
@@ -277,8 +284,8 @@ func (f *credentialFixture) createSecret(t *testing.T, namespace string) {
 	secret := &corev1.Secret{
 		Name: secretName,
 		Data: map[string][]byte{
-			"login":      []byte(fixtureUser),
+			managementUsernameKey: []byte(fixtureUser),
-			"credential": []byte(fixturePassword),
+			managementPasswordKey: []byte(fixturePassword),
 		},
 	}
 	if _, err := f.client.CoreV1().Secrets(namespace).Create(f.ctx, secret, metav1.CreateOptions{}); err != nil {

internal/database/adapter/postgresql/instance_controller_integration_test.go

@@ -0,0 +1,210 @@
+//go:build integration
+
+package postgresql_test
+
+import (
+	"bytes"
+	"context"
+	"os"
+	"testing"
+	"time"
+
+	databasev1alpha1 "git.ddupan.top/panxiao81/ayatori/api/database/v1alpha1"
+	secretadapter "git.ddupan.top/panxiao81/ayatori/internal/database/adapter/kubernetes"
+	"git.ddupan.top/panxiao81/ayatori/internal/database/adapter/postgresql"
+	"git.ddupan.top/panxiao81/ayatori/internal/database/application"
+	databasecontroller "git.ddupan.top/panxiao81/ayatori/internal/database/controller"
+	corev1 "k8s.io/api/core/v1"
+	rbacv1 "k8s.io/api/rbac/v1"
+	apierrors "k8s.io/apimachinery/pkg/api/errors"
+	"k8s.io/apimachinery/pkg/api/meta"
+	metav1 "k8s.io/apimachinery/pkg/apis/meta/v1"
+	"k8s.io/apimachinery/pkg/runtime"
+	yamlutil "k8s.io/apimachinery/pkg/util/yaml"
+	"k8s.io/client-go/rest"
+	ctrl "sigs.k8s.io/controller-runtime"
+	"sigs.k8s.io/controller-runtime/pkg/client"
+	controllerconfig "sigs.k8s.io/controller-runtime/pkg/config"
+	"sigs.k8s.io/controller-runtime/pkg/envtest"
+	metricsserver "sigs.k8s.io/controller-runtime/pkg/metrics/server"
+	"sigs.k8s.io/yaml"
+)
+
+const watchRevisionAnnotation = "test.ayatori/observation"
+
+func TestInstanceControllerWithRealPostgreSQL(t *testing.T) {
+	f := newCredentialFixture(t)
+	if _, err := envtest.InstallCRDs(f.config, envtest.CRDInstallOptions{
+		Paths: []string{"../../../../config/crd/bases"}, ErrorIfPathMissing: true,
+	}); err != nil {
+		t.Fatal(err)
+	}
+	scheme := runtime.NewScheme()
+	for _, install := range []func(*runtime.Scheme) error{databasev1alpha1.AddToScheme, corev1.AddToScheme, rbacv1.AddToScheme} {
+		if err := install(scheme); err != nil {
+			t.Fatal(err)
+		}
+	}
+	apiClient, err := client.New(f.config, client.Options{Scheme: scheme})
+	if err != nil {
+		t.Fatal(err)
+	}
+	restricted := instanceControllerRBAC(t, f, apiClient)
+	credentials, err := secretadapter.NewSecretCredentials(restricted, controllerNamespace)
+	if err != nil {
+		t.Fatal(err)
+	}
+	service, err := application.NewInstanceService(credentials, postgresql.Connector{})
+	if err != nil {
+		t.Fatal(err)
+	}
+	// 同进程 -count 重复启动测试 manager；生产继续校验 controller 名称唯一。
+	skipRepeatedName := true
+	manager, err := ctrl.NewManager(restricted, ctrl.Options{
+		Scheme: scheme, Cache: databasecontroller.InstanceCacheOptions(controllerNamespace),
+		Metrics: metricsserver.Options{BindAddress: "0"}, HealthProbeBindAddress: "0",
+		Controller: controllerconfig.Controller{SkipNameValidation: &skipRepeatedName},
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	reconciler := &databasecontroller.InstanceReconciler{Observer: service, SecretNamespace: controllerNamespace}
+	if err := reconciler.SetupWithManager(manager); err != nil {
+		t.Fatal(err)
+	}
+	managerContext, stop := context.WithCancel(f.ctx)
+	done := make(chan error, 1)
+	go func() { done <- manager.Start(managerContext) }()
+	t.Cleanup(func() {
+		stop()
+		select {
+		case err := <-done:
+			if err != nil {
+				t.Error(err)
+			}
+		case <-time.After(20 * time.Second):
+			t.Error("Instance manager 未停止")
+		}
+		service.Close()
+	})
+	object := &databasev1alpha1.PostgreSQLInstance{}
+	object.Name = "native-instance"
+	object.Spec.Endpoint = databasev1alpha1.PostgreSQLEndpoint{
+		Host: fixtureHost, HostAddr: fixtureAddress, Port: int32(f.port), SSLMode: "disable",
+	}
+	object.Spec.AdminCredentialRef = databasev1alpha1.AdminCredentialReference{
+		Name: secretName, UsernameKey: managementUsernameKey, PasswordKey: managementPasswordKey,
+	}
+	if err := apiClient.Create(f.ctx, object); err != nil {
+		t.Fatal(err)
+	}
+	awaitInstanceReason(t, f, apiClient, object, "DependencyUnavailable")
+	// 30 秒轮询前必须收到 Secret 创建事件；实际 controller 使用 namespace Role + metadata watch。
+	useNativeManager(t, f)
+	awaitInstanceReason(t, f, apiClient, object, "ManagementReady")
+	before := f.backendIDs(t)
+	f.updateSecret(t, func(secret *corev1.Secret) {
+		secret.Annotations = map[string]string{watchRevisionAnnotation: "changed"}
+	})
+	// 用实际 API 事件触发重验，metadata 改动不应换池。
+	time.Sleep(200 * time.Millisecond)
+	if f.backendIDs(t) != before {
+		t.Fatal("无关 Secret metadata 修改重建了连接")
+	}
+	f.queryPostgres(t, "ALTER ROLE native_manager PASSWORD '"+rotatedPassword+"'")
+	f.updateSecret(t, func(secret *corev1.Secret) { secret.Data[managementPasswordKey] = []byte("invalid-test-password") })
+	awaitInstanceReason(t, f, apiClient, object, "AuthenticationFailed")
+	f.updateSecret(t, func(secret *corev1.Secret) { secret.Data[managementPasswordKey] = []byte(rotatedPassword) })
+	awaitInstanceReason(t, f, apiClient, object, "ManagementReady")
+	if f.backendIDs(t) == before {
+		t.Fatal("凭据轮换没有替换旧连接")
+	}
+	f.queryPostgres(t, "ALTER ROLE native_manager NOCREATEROLE")
+	f.updateSecret(t, func(secret *corev1.Secret) { secret.Annotations[watchRevisionAnnotation] = "recheck" })
+	awaitInstanceReason(t, f, apiClient, object, "InsufficientPrivileges")
+	f.queryPostgres(t, "ALTER ROLE native_manager CREATEROLE")
+	f.updateSecret(t, func(secret *corev1.Secret) { secret.Annotations[watchRevisionAnnotation] = "recovered" })
+	awaitInstanceReason(t, f, apiClient, object, "ManagementReady")
+	if err := f.client.CoreV1().Secrets(controllerNamespace).Delete(f.ctx, secretName, metav1.DeleteOptions{}); err != nil {
+		t.Fatal(err)
+	}
+	awaitInstanceReason(t, f, apiClient, object, "DependencyUnavailable")
+	if f.backendIDs(t) != "" {
+		t.Fatal("Secret 删除后旧连接未释放")
+	}
+}
+
+func awaitInstanceReason(t *testing.T, f *credentialFixture, apiClient client.Client,
+	object *databasev1alpha1.PostgreSQLInstance, reason string) {
+	t.Helper()
+	deadline := time.Now().Add(10 * time.Second)
+	for time.Now().Before(deadline) {
+		if err := apiClient.Get(f.ctx, client.ObjectKeyFromObject(object), object); err != nil {
+			t.Fatal(err)
+		}
+		condition := meta.FindStatusCondition(object.Status.Conditions, "Ready")
+		if condition != nil && condition.Reason == reason && condition.ObservedGeneration == object.Generation {
+			if (condition.Status == metav1.ConditionTrue) != (reason == "ManagementReady") {
+				t.Fatal("Ready 与检查结果不一致")
+			}
+			return
+		}
+		time.Sleep(50 * time.Millisecond)
+	}
+	t.Fatalf("watch 未及时推进到 %s", reason)
+}
+
+func instanceControllerRBAC(t *testing.T, f *credentialFixture, apiClient client.Client) *rest.Config {
+	t.Helper()
+	roleBytes, err := os.ReadFile("../../../../config/rbac/role.yaml")
+	if err != nil {
+		t.Fatal(err)
+	}
+	role := &rbacv1.ClusterRole{}
+	if err := yaml.Unmarshal(roleBytes, role); err != nil {
+		t.Fatal(err)
+	}
+	if err := apiClient.Create(f.ctx, role); err != nil {
+		t.Fatal(err)
+	}
+	user := "instance-controller-test"
+	binding := &rbacv1.ClusterRoleBinding{}
+	binding.Name = user
+	binding.RoleRef = rbacv1.RoleRef{APIGroup: rbacv1.GroupName, Kind: "ClusterRole", Name: role.Name}
+	binding.Subjects = []rbacv1.Subject{{Kind: "User", APIGroup: rbacv1.GroupName, Name: user}}
+	if err := apiClient.Create(f.ctx, binding); err != nil {
+		t.Fatal(err)
+	}
+	credentialBytes, err := os.ReadFile("../../../../config/rbac/database_credentials_role.yaml")
+	if err != nil {
+		t.Fatal(err)
+	}
+	namespaceRole := &rbacv1.Role{}
+	decoder := yamlutil.NewYAMLOrJSONDecoder(bytes.NewReader(credentialBytes), 4096)
+	if err := decoder.Decode(namespaceRole); err != nil {
+		t.Fatal(err)
+	}
+	namespaceRole.Namespace = controllerNamespace
+	if err := apiClient.Create(f.ctx, namespaceRole); err != nil {
+		t.Fatal(err)
+	}
+	namespaceBinding := &rbacv1.RoleBinding{}
+	namespaceBinding.Name, namespaceBinding.Namespace = user, controllerNamespace
+	namespaceBinding.RoleRef = rbacv1.RoleRef{APIGroup: rbacv1.GroupName, Kind: "Role", Name: namespaceRole.Name}
+	namespaceBinding.Subjects = binding.Subjects
+	if err := apiClient.Create(f.ctx, namespaceBinding); err != nil {
+		t.Fatal(err)
+	}
+	config := rest.CopyConfig(f.config)
+	config.Impersonate.UserName = user
+	restrictedClient, err := client.New(config, client.Options{Scheme: apiClient.Scheme()})
+	if err != nil {
+		t.Fatal(err)
+	}
+	secret := &corev1.Secret{}
+	err = restrictedClient.Get(f.ctx, client.ObjectKey{Namespace: unrelatedNamespace, Name: secretName}, secret)
+	if !apierrors.IsForbidden(err) {
+		t.Fatal("Instance controller 可以跨 namespace 读取 Secret")
+	}
+	return config
+}

internal/database/adapter/postgresql/management.go

@@ -0,0 +1,58 @@
+package postgresql
+
+import (
+	"context"
+
+	"git.ddupan.top/panxiao81/ayatori/internal/database/application"
+	"git.ddupan.top/panxiao81/ayatori/internal/database/domain/instance"
+)
+
+// 只读取当前执行角色的属性，不能从可继承的角色成员关系推导 CREATEDB/CREATEROLE。
+// 所有事实来自同一条语句；不创建探针数据库，不修改管理账号或持久 schema。
+const inspectManagementStatement = `
+SELECT
+    pg_catalog.current_setting('server_version'),
+    ARRAY(SELECT name::text FROM pg_catalog.pg_available_extensions ORDER BY name),
+    role.rolsuper,
+    role.rolcreaterole,
+    role.rolcreatedb,
+    pg_catalog.pg_is_in_recovery() OR
+        pg_catalog.current_setting('transaction_read_only')::boolean
+FROM pg_catalog.pg_roles AS role
+WHERE role.rolname = current_user`
+
+func (d *database) InspectManagement(ctx context.Context) (application.DatabaseMetadata, error) {
+	var metadata application.DatabaseMetadata
+	var superuser, createRole, createDatabase, readOnly bool
+	err := d.pool.QueryRow(ctx, inspectManagementStatement).Scan(
+		&metadata.Version, &metadata.AvailableExtensions,
+		&superuser, &createRole, &createDatabase, &readOnly,
+	)
+	if err != nil {
+		return application.DatabaseMetadata{}, safeError(err, application.ErrObservation)
+	}
+	checks := instance.ManagementChecks{
+		Connection: instance.CheckPassed,
+		Metadata:   instance.CheckPassed,
+		Roles:      nativePrivilege(createRole && !superuser),
+		Databases:  nativePrivilege(createDatabase && !superuser),
+		// CREATEROLE 可管理自己新建角色的 membership；供应时必须显式取得 SET 权限，
+		// 再以 owner 操作数据库 ACL。这里不授权操作任意导入角色或他人数据库。
+		Grants: nativePrivilege(createRole && createDatabase && !superuser),
+		// 新建数据库 owner 可安装 trusted 扩展。具体扩展仍需逐请求执行和回读，
+		// 非 trusted 扩展不能因出现在 available 列表就视为可安装。
+		Extensions: nativePrivilege(createRole && createDatabase && !superuser),
+	}
+	if readOnly {
+		checks.Databases = instance.CheckUnavailable
+	}
+	metadata.Management = checks
+	return metadata, nil
+}
+
+func nativePrivilege(allowed bool) instance.CheckResult {
+	if allowed {
+		return instance.CheckPassed
+	}
+	return instance.CheckInsufficientPrivileges
+}

internal/database/adapter/postgresql/management_integration_test.go

@@ -0,0 +1,152 @@
+//go:build integration
+
+package postgresql_test
+
+import (
+	"context"
+	"errors"
+	"testing"
+
+	"github.com/jackc/pgx/v5"
+	corev1 "k8s.io/api/core/v1"
+
+	"git.ddupan.top/panxiao81/ayatori/internal/database/application"
+	"git.ddupan.top/panxiao81/ayatori/internal/database/domain/instance"
+)
+
+const nativeManager = "native_manager"
+
+func useNativeManager(t *testing.T, f *credentialFixture) {
+	t.Helper()
+	f.queryPostgres(t, "CREATE ROLE native_manager LOGIN CREATEDB CREATEROLE PASSWORD '"+fixturePassword+"'")
+	f.createSecret(t, controllerNamespace)
+	f.updateSecret(t, func(secret *corev1.Secret) { secret.Data[managementUsernameKey] = []byte(nativeManager) })
+}
+
+func assessManagement(t *testing.T, f *credentialFixture) instance.Snapshot {
+	t.Helper()
+	observation, err := f.service.ObserveManagement(f.ctx, f.target)
+	if err != nil {
+		t.Fatal(err)
+	}
+	aggregate, err := instance.Reconstitute(f.target, instance.Snapshot{}, false)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if err := aggregate.BeginValidation(); err != nil {
+		t.Fatal(err)
+	}
+	capabilities, err := observation.Capabilities()
+	if err != nil {
+		t.Fatal(err)
+	}
+	if err := aggregate.AssessManagement(capabilities); err != nil {
+		t.Fatal(err)
+	}
+	return aggregate.Snapshot()
+}
+
+func TestNativeManagementPrivileges(t *testing.T) {
+	f := newCredentialFixture(t)
+	useNativeManager(t, f)
+	if snapshot := assessManagement(t, f); snapshot.Readiness != instance.Ready {
+		t.Fatal("原生非 superuser 管理账号未通过检查")
+	}
+	before := f.backendIDs(t)
+	for _, attribute := range []string{"NOCREATEROLE", "NOCREATEDB"} {
+		f.queryPostgres(t, "ALTER ROLE native_manager "+attribute)
+		if snapshot := assessManagement(t, f); snapshot.Failure != instance.InsufficientPrivileges {
+			t.Fatal("已有连接忽略了管理权限撤回")
+		}
+		f.queryPostgres(t, "ALTER ROLE native_manager CREATEROLE CREATEDB")
+		if snapshot := assessManagement(t, f); snapshot.Readiness != instance.Ready {
+			t.Fatal("管理权限恢复后无法重新就绪")
+		}
+	}
+	if f.backendIDs(t) != before {
+		t.Fatal("权限检查不应要求重建连接才生效")
+	}
+	f.queryPostgres(t, "ALTER ROLE native_manager SET default_transaction_read_only = on")
+	f.service.Forget(f.target.Identity().Name())
+	if snapshot := assessManagement(t, f); snapshot.Failure != instance.DependencyUnavailable {
+		t.Fatal("只读会话不应标记可供应")
+	}
+	f.queryPostgres(t, "ALTER ROLE native_manager RESET default_transaction_read_only")
+	f.service.Forget(f.target.Identity().Name())
+	if snapshot := assessManagement(t, f); snapshot.Readiness != instance.Ready {
+		t.Fatal("恢复可写会话后没有就绪")
+	}
+	f.updateSecret(t, func(secret *corev1.Secret) { secret.Data[managementUsernameKey] = []byte(fixtureUser) })
+	if snapshot := assessManagement(t, f); snapshot.Failure != instance.InsufficientPrivileges {
+		t.Fatal("不应以 superuser 绕过非特权账号合同")
+	}
+	reads := 0
+	f.gate.beforeRead = func() {
+		reads++
+		if reads == 2 {
+			f.updateSecret(t, func(secret *corev1.Secret) { secret.Data[managementPasswordKey] = []byte(rotatedPassword) })
+		}
+	}
+	observation, err := f.service.ObserveManagement(f.ctx, f.target)
+	if !errors.Is(err, application.ErrCredentialsChanged) || observation.Target().Validate() == nil {
+		t.Fatal("管理观察期间凭据轮换应丢弃全部能力结果")
+	}
+	if f.backendIDs(t) != "" {
+		t.Fatal("中途轮换后不应保留旧管理连接")
+	}
+}
+
+// 以实际非 superuser 会话验证能力矩阵的依据，不用超级用户执行 SQL 模拟管理账号。
+// 这些固定名称只存在于本测试独占容器，生产观察本身不会创建探针对象。
+func TestNativeManagementSupplyContract(t *testing.T) {
+	f := newCredentialFixture(t)
+	useNativeManager(t, f)
+	config, err := pgx.ParseConfig("")
+	if err != nil {
+		t.Fatal("无法装配隔离测试连接")
+	}
+	config.Host, config.Port = fixtureAddress, uint16(f.port)
+	config.Database, config.User, config.Password = fixtureUser, nativeManager, fixturePassword
+	config.TLSConfig, config.Fallbacks = nil, nil
+	connection, err := pgx.ConnectConfig(f.ctx, config)
+	if err != nil {
+		t.Fatal("非 superuser 测试连接失败")
+	}
+	t.Cleanup(func() { _ = connection.Close(context.Background()) })
+	execute := func(statement string) {
+		t.Helper()
+		if _, err := connection.Exec(f.ctx, statement); err != nil {
+			t.Fatalf("原生管理能力合同未满足，步骤 %q", statement)
+		}
+	}
+	execute("CREATE ROLE managed_owner LOGIN NOSUPERUSER NOCREATEDB NOCREATEROLE NOREPLICATION")
+	execute("GRANT managed_owner TO native_manager WITH SET TRUE")
+	execute("CREATE DATABASE managed_database OWNER managed_owner")
+	execute("SET ROLE managed_owner")
+	execute("REVOKE CONNECT ON DATABASE managed_database FROM PUBLIC")
+	execute("GRANT CONNECT ON DATABASE managed_database TO managed_owner")
+	execute("RESET ROLE")
+	config.Database = "managed_database"
+	tenantConnection, err := pgx.ConnectConfig(f.ctx, config)
+	if err != nil {
+		t.Fatal("管理账号无法访问其受管数据库")
+	}
+	defer func() { _ = tenantConnection.Close(context.Background()) }()
+	if _, err := tenantConnection.Exec(f.ctx, "SET ROLE managed_owner; CREATE EXTENSION hstore"); err != nil {
+		t.Fatal("owner 无法安装 trusted 扩展")
+	}
+	var installed bool
+	if err := tenantConnection.QueryRow(f.ctx, "SELECT EXISTS (SELECT FROM pg_catalog.pg_extension WHERE extname = 'hstore')").Scan(&installed); err != nil || !installed {
+		t.Fatal("扩展安装后实际回读失败")
+	}
+	if _, err := tenantConnection.Exec(f.ctx, "CREATE EXTENSION file_fdw"); err == nil {
+		t.Fatal("非 trusted 扩展不应被 Ready 隐式授权")
+	}
+	if err := tenantConnection.Close(f.ctx); err != nil {
+		t.Fatal("关闭目标数据库连接失败")
+	}
+	execute("SET ROLE managed_owner")
+	execute("DROP DATABASE managed_database")
+	execute("RESET ROLE")
+	execute("DROP ROLE managed_owner")
+}

internal/database/application/instance_reconciliation.go

@@ -0,0 +1,149 @@
+package application
+
+import (
+	"context"
+	"errors"
+
+	"git.ddupan.top/panxiao81/ayatori/internal/database/domain/instance"
+)
+
+const (
+	instanceDependencyUnavailable = "DependencyUnavailable"
+	instanceAuthenticationFailed  = "AuthenticationFailed"
+)
+
+// InstanceRecord 是 API 快照；Revision 仅用于持久化并发保护，不是领域版本。
+type InstanceRecord struct {
+	Target   instance.ObservationTarget
+	Revision string
+	Deleting bool
+}
+
+type InstanceResources interface {
+	LoadInstance(context.Context, string) (*InstanceRecord, error)
+	ProtectInstance(context.Context, *InstanceRecord) (*InstanceRecord, error)
+	// InstanceReferences 返回一个可定位的阻塞引用；空字符串表示没有引用。
+	InstanceReferences(context.Context, string) (string, error)
+}
+
+type InstanceObserver interface {
+	ObserveManagement(context.Context, instance.ObservationTarget) (InstanceObservation, error)
+	Forget(string)
+}
+
+type InstanceResult struct {
+	Record           *InstanceRecord
+	Snapshot         instance.Snapshot
+	Reason           string
+	Message          string
+	RemoveProtection bool
+}
+
+// InstanceReconciliation 协调 API 保护、实时观察和领域判断，不拼装 Kubernetes status。
+type InstanceReconciliation struct {
+	Resources InstanceResources
+	Observer  InstanceObserver
+}
+
+func (s *InstanceReconciliation) Reconcile(ctx context.Context, name string) (InstanceResult, error) {
+	record, err := s.Resources.LoadInstance(ctx, name)
+	if err != nil {
+		return InstanceResult{}, err
+	}
+	if record == nil {
+		s.Observer.Forget(name)
+		return InstanceResult{}, nil
+	}
+	if record.Deleting {
+		return s.deleting(ctx, record)
+	}
+	record, err = s.Resources.ProtectInstance(ctx, record)
+	if err != nil {
+		return InstanceResult{}, err
+	}
+	// 每轮从无证据的领域对象开始；持久化 Ready 和连接存活不能替代本轮检查。
+	aggregate, err := instance.Reconstitute(record.Target, instance.Snapshot{}, false)
+	if err != nil {
+		return InstanceResult{}, err
+	}
+	if err := aggregate.BeginValidation(); err != nil {
+		return InstanceResult{}, err
+	}
+	observation, observationErr := s.Observer.ObserveManagement(ctx, record.Target)
+	result := InstanceResult{Record: record}
+	if observationErr != nil {
+		result.Snapshot = aggregate.Snapshot()
+		result.Snapshot.Readiness = instance.NotReady
+		result.Snapshot.ObservedRevision = record.Target.Revision().Value()
+		result.Reason, result.Message = observationFailure(observationErr)
+		return result, nil
+	}
+	capabilities, err := observation.Capabilities()
+	if err != nil {
+		return InstanceResult{}, err
+	}
+	if err := aggregate.AssessManagement(capabilities); err != nil {
+		return InstanceResult{}, err
+	}
+	result.Snapshot = aggregate.Snapshot()
+	result.Snapshot.ReportedVersion = observation.Version()
+	result.Reason, result.Message = managementResult(result.Snapshot.Failure)
+	return result, nil
+}
+
+func (s *InstanceReconciliation) deleting(ctx context.Context, record *InstanceRecord) (InstanceResult, error) {
+	name := record.Target.Identity().Name()
+	s.Observer.Forget(name)
+	aggregate, err := instance.Reconstitute(record.Target, instance.Snapshot{}, true)
+	if err != nil {
+		return InstanceResult{}, err
+	}
+	if err := aggregate.BeginDeletion(); err != nil {
+		return InstanceResult{}, err
+	}
+	result := InstanceResult{Record: record, Snapshot: aggregate.Snapshot(), Reason: "Deleting"}
+	reference, err := s.Resources.InstanceReferences(ctx, name)
+	if err != nil {
+		result.Reason = instanceDependencyUnavailable
+		result.Message = "无法确认 Database/Tenant 引用已解除；保留 Instance 删除保护并重试"
+		return result, nil
+	}
+	if reference != "" {
+		result.Reason = "InstanceInUse"
+		result.Message = "仍被 " + reference + " 引用；先处理该资源，不会级联删除外部数据库"
+		return result, nil
+	}
+	result.Message = "引用已解除，仅移除登记保护；不删除 PostgreSQL 或凭据"
+	result.RemoveProtection = true
+	return result, nil
+}
+
+func observationFailure(err error) (string, string) {
+	switch {
+	case errors.Is(err, ErrAuthentication):
+		return instanceAuthenticationFailed, "管理连接认证或 TLS 校验失败；检查管理 Secret 和 CA/证书配置"
+	case errors.Is(err, ErrCredentialsInvalid):
+		return "InvalidCredentials", "管理 Secret 的用户名或密码字段缺失；检查引用字段映射"
+	case errors.Is(err, ErrCredentialsChanged):
+		return "CredentialsChanged", "观察期间管理凭据变化，已丢弃结果并关闭旧连接；等待重新验证"
+	case errors.Is(err, ErrCredentialsUnavailable):
+		return instanceDependencyUnavailable, "无法读取管理 Secret；检查其是否存在及 controller namespace 内的读取权限"
+	default:
+		return instanceDependencyUnavailable, "管理连接或能力查询失败；检查 PostgreSQL 可达性、catalog 读取权限和超时"
+	}
+}
+
+func managementResult(failure instance.Failure) (string, string) {
+	switch failure {
+	case instance.NoFailure:
+		return "ManagementReady", "当前管理能力检查通过；具体资源授权和扩展安装仍需执行时验证"
+	case instance.InsufficientPrivileges:
+		return "InsufficientPrivileges", "原生管理要求非 superuser 且具备 CREATEDB/CREATEROLE；不会自动修改账号权限"
+	case instance.DependencyUnavailable:
+		return instanceDependencyUnavailable, "当前 PostgreSQL 不可写或所需管理能力暂不可用"
+	case instance.AuthenticationFailed:
+		return instanceAuthenticationFailed, "当前管理能力检查未通过认证"
+	default:
+		return "ObservationIncomplete", "管理能力检查尚有缺项，不能仅凭 metadata 查询成功标记 Ready"
+	}
+}

internal/database/application/instance_reconciliation_test.go

@@ -0,0 +1,79 @@
+package application
+
+import (
+	"context"
+	"errors"
+	"testing"
+
+	"git.ddupan.top/panxiao81/ayatori/internal/database/domain/instance"
+)
+
+func TestInstanceFailurePresentation(t *testing.T) {
+	cases := []struct {
+		err    error
+		reason string
+	}{
+		{ErrAuthentication, "AuthenticationFailed"},
+		{ErrCredentialsInvalid, "InvalidCredentials"},
+		{ErrCredentialsChanged, "CredentialsChanged"},
+		{ErrCredentialsUnavailable, instanceDependencyUnavailable},
+		{context.DeadlineExceeded, instanceDependencyUnavailable},
+		{errors.New("private backend detail"), instanceDependencyUnavailable},
+	}
+	for _, test := range cases {
+		reason, message := observationFailure(test.err)
+		if reason != test.reason || message == "" || message == test.err.Error() {
+			t.Fatal("观察失败没有安全且可诊断的状态")
+		}
+	}
+	for _, failure := range []instance.Failure{
+		instance.NoFailure, instance.ObservationIncomplete, instance.DependencyUnavailable,
+		instance.AuthenticationFailed, instance.InsufficientPrivileges,
+	} {
+		reason, message := managementResult(failure)
+		if reason == "" || message == "" || (reason == "ManagementReady") != (failure == instance.NoFailure) {
+			t.Fatal("领域能力判定与状态不一致")
+		}
+	}
+}
+
+func TestMetadataCannotEstablishManagementReadiness(t *testing.T) {
+	source := &sourceStub{}
+	source.credentials, _ = NewCredentials("test", serviceTestPassword)
+	connector := &connectorStub{}
+	service, err := NewInstanceService(source, connector)
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer service.Close()
+	target := serviceTarget(t, "uid", "postgres.test", "management", 1)
+	if _, err := service.ObserveMetadata(t.Context(), target); err != nil {
+		t.Fatal(err)
+	}
+	connector.databases[0].metadata.Management = instance.ManagementChecks{
+		Connection: instance.CheckPassed, Metadata: instance.CheckPassed,
+		Roles: instance.CheckPassed, Databases: instance.CheckPassed,
+		Grants: instance.CheckPassed, Extensions: instance.CheckPassed,
+	}
+	observation, err := service.ObserveMetadata(t.Context(), target)
+	if err != nil {
+		t.Fatal(err)
+	}
+	capabilities, err := observation.Capabilities()
+	if err != nil {
+		t.Fatal(err)
+	}
+	aggregate, err := instance.Reconstitute(target, instance.Snapshot{}, false)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if err := aggregate.BeginValidation(); err != nil {
+		t.Fatal(err)
+	}
+	if err := aggregate.AssessManagement(capabilities); err != nil {
+		t.Fatal(err)
+	}
+	if aggregate.Snapshot().Failure != instance.ObservationIncomplete {
+		t.Fatal("metadata 入口不应携带完整管理检查")
+	}
+}

internal/database/application/instance_service.go

@@ -36,6 +36,7 @@ var (
 // Metadata 只查询版本与可用扩展，不能产生领域 Ready。
 type Database interface {
 	InspectMetadata(context.Context) (DatabaseMetadata, error)
+	InspectManagement(context.Context) (DatabaseMetadata, error)
 	Close()
 }
 
@@ -82,17 +83,26 @@ func (s *InstanceService) ObserveVersion(ctx context.Context, target instance.Ob
 
 // ObserveMetadata 返回当前目标和凭据下的版本与扩展；任何失败均丢弃全部结果。
 // 调用者仍需使用 CR resourceVersion 保存前提防止 spec 并发修改；本方法不建立跨系统事务。
-func (s *InstanceService) ObserveMetadata(ctx context.Context, target instance.ObservationTarget) (MetadataObservation, error) {
+func (s *InstanceService) ObserveMetadata(ctx context.Context, target instance.ObservationTarget) (InstanceObservation, error) {
+	return s.observe(ctx, target, false)
+}
+
+// ObserveManagement 复用同一凭据刷新与回读边界，但每轮重新检查原生管理能力。
+func (s *InstanceService) ObserveManagement(ctx context.Context, target instance.ObservationTarget) (InstanceObservation, error) {
+	return s.observe(ctx, target, true)
+}
+
+func (s *InstanceService) observe(ctx context.Context, target instance.ObservationTarget, management bool) (InstanceObservation, error) {
 	if err := target.Validate(); err != nil {
-		return MetadataObservation{}, err
+		return InstanceObservation{}, err
 	}
 	s.mu.Lock()
 	defer s.mu.Unlock()
 	if s.closed {
-		return MetadataObservation{}, ErrClosed
+		return InstanceObservation{}, ErrClosed
 	}
 	if err := ctx.Err(); err != nil {
-		return MetadataObservation{}, err
+		return InstanceObservation{}, err
 	}
 
 	// 先读取有效凭据。读取失败时不得继续使用缓存中的旧连接。
@@ -100,11 +110,11 @@ func (s *InstanceService) ObserveMetadata(ctx context.Context, target instance.O
 	credentials, err := s.source.Read(ctx, target.Definition().AdminCredential())
 	if err != nil {
 		s.release(name)
-		return MetadataObservation{}, credentialError(err)
+		return InstanceObservation{}, credentialError(err)
 	}
 	if credentials.username == "" || credentials.password == "" {
 		s.release(name)
-		return MetadataObservation{}, ErrCredentialsInvalid
+		return InstanceObservation{}, ErrCredentialsInvalid
 	}
 
 	// 连接身份与有效值均未变化时复用 pgxpool；generation 本身不要求换池。
@@ -117,7 +127,7 @@ func (s *InstanceService) ObserveMetadata(ctx context.Context, target instance.O
 	if current == nil {
 		database, err := s.connector.Connect(ctx, target.Definition().Endpoint(), credentials)
 		if err != nil {
-			return MetadataObservation{}, err
+			return InstanceObservation{}, err
 		}
 		current = &entry{
 			target:      target,
@@ -127,30 +137,38 @@ func (s *InstanceService) ObserveMetadata(ctx context.Context, target instance.O
 		s.entries[name] = current
 	}
 
-	metadata, err := current.database.InspectMetadata(ctx)
+	var metadata DatabaseMetadata
+	if management {
+		metadata, err = current.database.InspectManagement(ctx)
+	} else {
+		metadata, err = current.database.InspectMetadata(ctx)
+		// 即使 adapter 误填权限，也不能把只读 metadata 入口升级为 Ready。
+		metadata.Management = instance.ManagementChecks{}
+	}
 	if err != nil {
 		s.release(name)
-		return MetadataObservation{}, err
+		return InstanceObservation{}, err
 	}
 	if metadata.Version == "" {
 		s.release(name)
-		return MetadataObservation{}, ErrObservation
+		return InstanceObservation{}, ErrObservation
 	}
 
 	// 回读后再检查凭据，避免把轮换前取得的结果交给新凭据的调用链。
 	latest, err := s.source.Read(ctx, target.Definition().AdminCredential())
 	if err != nil {
 		s.release(name)
-		return MetadataObservation{}, credentialError(err)
+		return InstanceObservation{}, credentialError(err)
 	}
 	if latest != credentials {
 		s.release(name)
-		return MetadataObservation{}, ErrCredentialsChanged
+		return InstanceObservation{}, ErrCredentialsChanged
 	}
-	return MetadataObservation{
+	return InstanceObservation{
 		target:     target,
 		version:    metadata.Version,
 		extensions: instance.ObserveExtensionSupport(metadata.AvailableExtensions),
+		management: metadata.Management,
 	}, nil
 }
 

internal/database/application/instance_service_test.go

@@ -42,6 +42,10 @@ type databaseStub struct {
 	metadata DatabaseMetadata
 }
 
+func (d *databaseStub) InspectManagement(ctx context.Context) (DatabaseMetadata, error) {
+	return d.InspectMetadata(ctx)
+}
+
 func (d *databaseStub) InspectMetadata(context.Context) (DatabaseMetadata, error) {
 	return d.metadata, d.err
 }

internal/database/application/metadata.go

@@ -18,23 +18,31 @@ package application
 
 import "git.ddupan.top/panxiao81/ayatori/internal/database/domain/instance"
 
-// DatabaseMetadata 是一次只读查询的事实，不包含管理权限或完整就绪结论。
+// DatabaseMetadata 是一次只读查询的事实，可附带原生管理检查，但不包含就绪结论。
 // AvailableExtensions 是服务器提供的可用列表，不是已安装列表或安装授权。
 type DatabaseMetadata struct {
 	Version             string
 	AvailableExtensions []string
+	// Management 仅由 InspectManagement 填充；metadata 查询必须保持未观察。
+	Management instance.ManagementChecks
 }
 
-// MetadataObservation 只在查询成功且有效凭据再次核对一致后产生。
+// InstanceObservation 只在查询成功且有效凭据再次核对一致后产生。
 // target 绑定本次调用，而非连接最初创建时的 generation；零值表示没有观察。
-type MetadataObservation struct {
+type InstanceObservation struct {
 	target     instance.ObservationTarget
 	version    string
 	extensions instance.ExtensionSupport
+	management instance.ManagementChecks
 }
 
-func (o MetadataObservation) Target() instance.ObservationTarget { return o.target }
+// Capabilities 保留缺项为未观察；不能从 metadata 的成功补齐管理检查。
-func (o MetadataObservation) Version() string                    { return o.version }
+func (o InstanceObservation) Capabilities() (instance.CapabilityObservation, error) {
-func (o MetadataObservation) Extensions() instance.ExtensionSupport {
+	return instance.NewCapabilityObservation(o.target, o.version, o.management)
+}
+
+func (o InstanceObservation) Target() instance.ObservationTarget { return o.target }
+func (o InstanceObservation) Version() string                    { return o.version }
+func (o InstanceObservation) Extensions() instance.ExtensionSupport {
 	return o.extensions
 }

internal/database/controller/instance_controller.go

@@ -0,0 +1,42 @@
+package controller
+
+import (
+	"context"
+	"time"
+
+	"git.ddupan.top/panxiao81/ayatori/internal/database/adapter/kubernetes"
+	"git.ddupan.top/panxiao81/ayatori/internal/database/application"
+	ctrl "sigs.k8s.io/controller-runtime"
+	"sigs.k8s.io/controller-runtime/pkg/client"
+)
+
+type InstanceReconciler struct {
+	Client          client.Client
+	Reader          client.Reader
+	Observer        application.InstanceObserver
+	SecretNamespace string
+}
+
+// +kubebuilder:rbac:groups=database.ayatori.ddupan.top,resources=postgresqlinstances,verbs=get;list;watch;update;patch
+// +kubebuilder:rbac:groups=database.ayatori.ddupan.top,resources=postgresqlinstances/status,verbs=get;update;patch
+// +kubebuilder:rbac:groups=database.ayatori.ddupan.top,resources=postgresqlinstances/finalizers,verbs=update
+// Secret 权限单独声明为 namespace Role，不放入生成的 ClusterRole。
+
+func (r *InstanceReconciler) Reconcile(ctx context.Context, request ctrl.Request) (ctrl.Result, error) {
+	resources := &kubernetes.InstanceResources{Client: r.Client, Reader: r.Reader}
+	service := application.InstanceReconciliation{Resources: resources, Observer: r.Observer}
+	observationContext, cancel := context.WithTimeout(ctx, 15*time.Second)
+	defer cancel()
+	result, err := service.Reconcile(observationContext, request.Name)
+	if err != nil {
+		return ctrl.Result{}, err
+	}
+	// 查询超时后仍用 worker context 保存安全失败结果；manager 停止时不强行写入。
+	if err := resources.PresentInstance(ctx, result); err != nil {
+		return ctrl.Result{}, err
+	}
+	if result.Record == nil || result.RemoveProtection {
+		return ctrl.Result{}, nil
+	}
+	return ctrl.Result{RequeueAfter: dependencyRetry}, nil
+}

internal/database/controller/instance_controller_test.go

@@ -0,0 +1,192 @@
+package controller
+
+import (
+	"context"
+	"errors"
+	"testing"
+
+	databasev1alpha1 "git.ddupan.top/panxiao81/ayatori/api/database/v1alpha1"
+	"git.ddupan.top/panxiao81/ayatori/internal/database/adapter/kubernetes"
+	"git.ddupan.top/panxiao81/ayatori/internal/database/application"
+	"git.ddupan.top/panxiao81/ayatori/internal/database/domain/instance"
+	apierrors "k8s.io/apimachinery/pkg/api/errors"
+	"k8s.io/apimachinery/pkg/api/meta"
+	metav1 "k8s.io/apimachinery/pkg/apis/meta/v1"
+	ctrl "sigs.k8s.io/controller-runtime"
+	"sigs.k8s.io/controller-runtime/pkg/client"
+	"sigs.k8s.io/controller-runtime/pkg/controller/controllerutil"
+)
+
+type instanceBackend struct {
+	checks  instance.ManagementChecks
+	err     error
+	inspect func()
+	closed  int
+}
+
+func (b *instanceBackend) Read(context.Context, instance.CredentialReference) (application.Credentials, error) {
+	return application.NewCredentials("fixture", "test-only-instance-password")
+}
+
+func (b *instanceBackend) Connect(context.Context, instance.Endpoint, application.Credentials) (application.Database, error) {
+	return b, nil
+}
+
+func (b *instanceBackend) InspectMetadata(context.Context) (application.DatabaseMetadata, error) {
+	return application.DatabaseMetadata{Version: "18"}, nil
+}
+
+func (b *instanceBackend) InspectManagement(context.Context) (application.DatabaseMetadata, error) {
+	if b.inspect != nil {
+		b.inspect()
+	}
+	return application.DatabaseMetadata{Version: "18", Management: b.checks}, b.err
+}
+
+func (b *instanceBackend) Close() { b.closed++ }
+
+func newInstanceReconciler(t *testing.T, apiClient client.Client, backend *instanceBackend) *InstanceReconciler {
+	t.Helper()
+	service, err := application.NewInstanceService(backend, backend)
+	if err != nil {
+		t.Fatal(err)
+	}
+	t.Cleanup(service.Close)
+	return &InstanceReconciler{Client: apiClient, Reader: apiClient, Observer: service}
+}
+
+func reconcileInstance(t *testing.T, reconciler *InstanceReconciler, object *databasev1alpha1.PostgreSQLInstance) {
+	t.Helper()
+	if _, err := reconciler.Reconcile(t.Context(), ctrl.Request{NamespacedName: client.ObjectKeyFromObject(object)}); err != nil {
+		t.Fatal(err)
+	}
+}
+
+func assertInstanceReason(t *testing.T, object *databasev1alpha1.PostgreSQLInstance, reason string) {
+	t.Helper()
+	condition := meta.FindStatusCondition(object.Status.Conditions, "Ready")
+	if condition == nil || condition.Reason != reason || condition.ObservedGeneration != object.Generation {
+		t.Fatalf("Instance 状态不是当前 generation 的 %s", reason)
+	}
+	if reason != "ManagementReady" && condition.Status != metav1.ConditionFalse {
+		t.Fatal("失败状态仍为 Ready")
+	}
+}
+
+func TestInstanceObservationAPI(t *testing.T) {
+	apiClient, _, _ := bindingEnvironment(t)
+	backend := &instanceBackend{checks: instance.ManagementChecks{
+		Connection: instance.CheckPassed, Metadata: instance.CheckPassed,
+		Roles: instance.CheckPassed, Databases: instance.CheckPassed,
+		Grants: instance.CheckPassed, Extensions: instance.CheckPassed,
+	}}
+	reconciler := newInstanceReconciler(t, apiClient, backend)
+	object := readyInstance(t, apiClient, "observed-instance")
+	backend.inspect = func() {
+		current := &databasev1alpha1.PostgreSQLInstance{}
+		current.Name = object.Name
+		reload(t, apiClient, current)
+		if !controllerutil.ContainsFinalizer(current, kubernetes.InstanceFinalizer) {
+			t.Fatal("观察早于 finalizer 持久化")
+		}
+	}
+	reconcileInstance(t, reconciler, object)
+	reload(t, apiClient, object)
+	assertInstanceReason(t, object, "ManagementReady")
+	if object.Status.Phase != string(instance.PhaseReady) || object.Status.PostgreSQLVersion != "18" {
+		t.Fatal("当前成功观察未呈现")
+	}
+	before := object.ResourceVersion
+	reconcileInstance(t, reconciler, object)
+	reload(t, apiClient, object)
+	if object.ResourceVersion != before {
+		t.Fatal("相同观察不应反复写入 status")
+	}
+	backend.err = application.ErrAuthentication
+	reconcileInstance(t, reconciler, object)
+	reload(t, apiClient, object)
+	assertInstanceReason(t, object, "AuthenticationFailed")
+	if backend.closed != 1 || object.Status.PostgreSQLVersion != "" {
+		t.Fatal("观察失败应释放连接并清除旧版本结果")
+	}
+	backend.err = nil
+	backend.checks.Grants = instance.CheckUnobserved
+	reconcileInstance(t, reconciler, object)
+	reload(t, apiClient, object)
+	assertInstanceReason(t, object, "ObservationIncomplete")
+	backend.checks.Grants = instance.CheckPassed
+	// 用新 service/reconciler 恢复；不依赖上轮领域对象或 Ready。
+	reconciler = newInstanceReconciler(t, apiClient, backend)
+	reconcileInstance(t, reconciler, object)
+	reload(t, apiClient, object)
+	assertInstanceReason(t, object, "ManagementReady")
+
+	backend.inspect = func() {
+		reload(t, apiClient, object)
+		object.Annotations = map[string]string{"concurrent": "kept-by-instance-test"}
+		if err := apiClient.Update(t.Context(), object); err != nil {
+			t.Fatal(err)
+		}
+	}
+	_, err := reconciler.Reconcile(t.Context(), ctrl.Request{NamespacedName: client.ObjectKeyFromObject(object)})
+	if !apierrors.IsConflict(err) {
+		t.Fatal("旧观察不应覆盖在途 API 修改")
+	}
+	backend.inspect = nil
+	reconcileInstance(t, reconciler, object)
+	reload(t, apiClient, object)
+	if object.Annotations["concurrent"] != "kept-by-instance-test" {
+		t.Fatal("重试覆盖了其他字段")
+	}
+}
+
+type failedReferenceReader struct{ client.Reader }
+
+func (*failedReferenceReader) List(context.Context, client.ObjectList, ...client.ListOption) error {
+	return errors.New("injected reference list failure")
+}
+
+func TestInstanceDeletionProtection(t *testing.T) {
+	apiClient, _, _ := bindingEnvironment(t)
+	backend := &instanceBackend{}
+	reconciler := newInstanceReconciler(t, apiClient, backend)
+	object := readyInstance(t, apiClient, "protected-instance")
+	reconcileInstance(t, reconciler, object)
+	reload(t, apiClient, object)
+	database := availableDatabase(t, apiClient, "retained-database", object)
+	database.Status.Phase = "Released"
+	if err := apiClient.Status().Update(t.Context(), database); err != nil {
+		t.Fatal(err)
+	}
+	tenant := provisionTenant("pending-request", object.Name)
+	requireCreate(t, apiClient, tenant)
+	if err := apiClient.Delete(t.Context(), object); err != nil {
+		t.Fatal(err)
+	}
+	backend.inspect = func() { t.Fatal("删除中不应连接 PostgreSQL") }
+	reconciler.Reader = &failedReferenceReader{Reader: apiClient}
+	reconcileInstance(t, reconciler, object)
+	reload(t, apiClient, object)
+	assertInstanceReason(t, object, reasonDependency)
+	reconciler.Reader = apiClient
+	reconcileInstance(t, reconciler, object)
+	reload(t, apiClient, object)
+	assertInstanceReason(t, object, "InstanceInUse")
+	if object.Status.Phase != string(instance.PhaseDeleting) || backend.closed != 1 {
+		t.Fatal("删除没有停止本地观察")
+	}
+	if err := apiClient.Delete(t.Context(), database); err != nil {
+		t.Fatal(err)
+	}
+	reconcileInstance(t, reconciler, object)
+	reload(t, apiClient, object)
+	assertInstanceReason(t, object, "InstanceInUse")
+	if err := apiClient.Delete(t.Context(), tenant); err != nil {
+		t.Fatal(err)
+	}
+	reconcileInstance(t, reconciler, object)
+	if err := apiClient.Get(t.Context(), client.ObjectKeyFromObject(object), object); !apierrors.IsNotFound(err) {
+		t.Fatal("最后一个引用解除后 Instance 应可删除")
+	}
+	reconcileInstance(t, reconciler, object)
+}

internal/database/controller/instance_setup.go

@@ -0,0 +1,77 @@
+package controller
+
+import (
+	"context"
+	"errors"
+
+	databasev1alpha1 "git.ddupan.top/panxiao81/ayatori/api/database/v1alpha1"
+	corev1 "k8s.io/api/core/v1"
+	"k8s.io/apimachinery/pkg/util/validation"
+	ctrl "sigs.k8s.io/controller-runtime"
+	"sigs.k8s.io/controller-runtime/pkg/cache"
+	"sigs.k8s.io/controller-runtime/pkg/client"
+	"sigs.k8s.io/controller-runtime/pkg/handler"
+)
+
+// InstanceCacheOptions 必须在创建 manager 时使用；只 watch 固定 namespace 的 Secret metadata。
+// SecretCredentials 始终直读 API，不会令共享 cache 保存密码。
+func InstanceCacheOptions(namespace string) cache.Options {
+	return cache.Options{ByObject: map[client.Object]cache.ByObject{
+		&corev1.Secret{}: {Namespaces: map[string]cache.Config{namespace: {}}},
+	}}
+}
+
+func (r *InstanceReconciler) SetupWithManager(manager ctrl.Manager) error {
+	if r.Observer == nil || len(validation.IsDNS1123Label(r.SecretNamespace)) != 0 {
+		return errors.New("instance observer and valid management Secret namespace required")
+	}
+	if r.Client == nil {
+		r.Client = manager.GetClient()
+	}
+	if r.Reader == nil {
+		r.Reader = manager.GetAPIReader()
+	}
+	return ctrl.NewControllerManagedBy(manager).
+		Named("database-instance").
+		For(&databasev1alpha1.PostgreSQLInstance{}).
+		WatchesMetadata(&corev1.Secret{}, handler.EnqueueRequestsFromMapFunc(r.instancesForSecret)).
+		Watches(&databasev1alpha1.PostgreSQLDatabase{}, handler.EnqueueRequestsFromMapFunc(r.instanceForReference)).
+		Watches(&databasev1alpha1.PostgreSQLTenant{}, handler.EnqueueRequestsFromMapFunc(r.instanceForReference)).
+		Complete(r)
+}
+
+func (r *InstanceReconciler) instancesForSecret(ctx context.Context, object client.Object) []ctrl.Request {
+	if object.GetNamespace() != r.SecretNamespace {
+		return nil
+	}
+	instances := &databasev1alpha1.PostgreSQLInstanceList{}
+	if err := r.Client.List(ctx, instances); err != nil {
+		ctrl.LoggerFrom(ctx).Error(err, "无法映射管理 Secret 事件；等待低频重试")
+		return nil
+	}
+	var requests []ctrl.Request
+	for _, item := range instances.Items {
+		if string(item.Spec.AdminCredentialRef.Name) == object.GetName() {
+			request := ctrl.Request{Name: item.Name}
+			requests = append(requests, request)
+		}
+	}
+	return requests
+}
+
+func (r *InstanceReconciler) instanceForReference(_ context.Context, object client.Object) []ctrl.Request {
+	var name string
+	switch item := object.(type) {
+	case *databasev1alpha1.PostgreSQLDatabase:
+		name = string(item.Spec.InstanceRef.Name)
+	case *databasev1alpha1.PostgreSQLTenant:
+		if item.Spec.Provision != nil {
+			name = string(item.Spec.Provision.InstanceRef.Name)
+		}
+	}
+	if name == "" {
+		return nil
+	}
+	request := ctrl.Request{Name: name}
+	return []ctrl.Request{request}
+}