feat(backend): admin 手动运维接真实 lifecycle/provision

管理后台节点页的「拉黑/恢复(draining/up)」与「一键换机(replace)」此前
注入的是 StubLifecycle/StubProvision(Ready()=false,UI 置灰、操作返回
服务不可用)。本次接真实实现:

- RealLifecycle:复用 scheduler 的乐观锁状态流转(UPDATE nodes WHERE
  status=from + node_events + directory_version bump),手动运维与调度
  自动循环共用同一条 canonical 路径;读当前状态作 from 守卫,幂等安全。
- RealProvision:委托 provision.Service.Replace 的 make-before-break 换机
  (先拉起新节点再 drain 旧节点,容量不掉)。
- BuildServices:lifecycle 恒为真实;provision 在服务初始化成功时为真实,
  否则回退 not-ready stub(无厂商凭证时点击换机给出明确错误)。

补 services_real_test.go:换机委托(空 UUID 起新编排)、错误透传、
非法目标状态拒绝、Ready 标志。server 全量 23 包测试通过。
e2e(真实换机)待机群 + 厂商凭证。

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

server/internal/admin/router.go

@@ -10,6 +10,8 @@ import (
 	"github.com/go-chi/chi/v5/middleware"
 	"github.com/redis/go-redis/v9"
 	"github.com/wangjia/pangolin/server/internal/codes"
+	"github.com/wangjia/pangolin/server/internal/provision"
+	"github.com/wangjia/pangolin/server/internal/provision/providers"
 )
 
 // NewRouter wires the full admin HTTP handler chain:
@@ -46,17 +48,28 @@ func NewRouter(h *Handlers, sessions *SessionStore, cfg *Config, sec *SecurityLo
 	return r
 }
 
-// BuildServices assembles the downstream services. The codes service (#3) is
+// BuildServices assembles the downstream services. Codes (#3) and lifecycle (#5)
-// real; lifecycle (#5) and provisioning (#14) are stubs until those modules
+// are always real (DB-backed). Provisioning (#14) is real when the provision
-// land — the UI greys out their controls (Ready()==false).
+// service initializes; otherwise it falls back to the not-ready stub and the UI
+// greys out the replace control (Ready()==false).
 func BuildServices(db *sql.DB, rdb *redis.Client, failMax int, lockDur time.Duration) Services {
 	codeStore := codes.NewStore(db)
 	codeSvc := codes.NewService(codeStore, rdb, failMax, lockDur)
-	return Services{
+	svc := Services{
 		Codes:     NewCodesAdapter(codeSvc, codeStore),
-		Lifecycle: NewStubLifecycle(),
+		Lifecycle: NewRealLifecycle(db),
 		Provision: NewStubProvision(),
 	}
+	provSvc, err := provision.NewService(provision.Config{
+		Store:    provision.NewMySQLStore(db),
+		Adapters: providers.NewRegistry(),
+	})
+	if err != nil {
+		log.Printf("admin: provision service init failed (%v) — replace control disabled", err)
+	} else {
+		svc.Provision = NewRealProvision(provSvc)
+	}
+	return svc
 }
 
 // NewHandler builds the complete admin http.Handler from its dependencies.

server/internal/admin/services.go

@@ -2,10 +2,16 @@ package admin
 
 import (
 	"context"
+	"database/sql"
 	"errors"
+	"fmt"
+	"strconv"
 	"time"
 
 	"github.com/wangjia/pangolin/server/internal/codes"
+	"github.com/wangjia/pangolin/server/internal/provision"
+	"github.com/wangjia/pangolin/server/internal/scheduler"
+	"github.com/wangjia/pangolin/server/internal/scheduler/orchestrate"
 )
 
 // ErrServiceUnavailable is returned by stub services whose real implementation
@@ -134,6 +140,55 @@ func (StubLifecycle) TransitionStatus(context.Context, int64, string, string) er
 // Ready reports false.
 func (StubLifecycle) Ready() bool { return false }
 
+// RealLifecycle is the SQL-backed lifecycle service. It reuses the scheduler's
+// canonical optimistic-lock transition (UPDATE nodes ... WHERE status=from, plus
+// a node_events row and a directory_version bump so clients re-fetch), so manual
+// admin transitions and the autonomous scheduler loop share one code path.
+type RealLifecycle struct {
+	db *sql.DB
+	lc orchestrate.LifecycleService
+}
+
+// NewRealLifecycle wires the lifecycle adapter over db.
+func NewRealLifecycle(db *sql.DB) *RealLifecycle {
+	return &RealLifecycle{
+		db: db,
+		lc: scheduler.NewOrchestrateLifecycle(scheduler.NewSQLLifecycle(db, nil)),
+	}
+}
+
+// TransitionStatus moves nodeID to target ("draining" | "up"). It reads the
+// current status as the optimistic-lock guard, so the transition is a safe
+// no-op if the node already moved.
+func (l *RealLifecycle) TransitionStatus(ctx context.Context, nodeID int64, target, actor string) error {
+	if target != "draining" && target != "up" {
+		return fmt.Errorf("admin: unsupported target status %q", target)
+	}
+	var from string
+	switch err := l.db.QueryRowContext(ctx, `SELECT status FROM nodes WHERE id = ?`, nodeID).Scan(&from); err {
+	case nil:
+	case sql.ErrNoRows:
+		return fmt.Errorf("admin: node %d not found", nodeID)
+	default:
+		return err
+	}
+	if from == target {
+		return nil // already in target state — idempotent no-op
+	}
+	n, err := l.lc.TransitionStatus(ctx, strconv.FormatInt(nodeID, 10), from, target,
+		map[string]any{"actor": actor, "source": "admin"})
+	if err != nil {
+		return err
+	}
+	if n == 0 {
+		return fmt.Errorf("admin: node %d status changed concurrently", nodeID)
+	}
+	return nil
+}
+
+// Ready reports true — the lifecycle service only needs the database.
+func (*RealLifecycle) Ready() bool { return true }
+
 // --------------------------------------------------------------------------
 // Node provisioning service (#14) — not yet implemented
 // --------------------------------------------------------------------------
@@ -159,6 +214,34 @@ func (StubProvision) Replace(context.Context, int64, string) error { return ErrS
 // Ready reports false.
 func (StubProvision) Ready() bool { return false }
 
+// provReplacer is the slice of provision.Service the admin replace control needs
+// (declared as an interface so tests can fake it without a database).
+type provReplacer interface {
+	Replace(ctx context.Context, nodeID int64, replacementUUID string) (*provision.ReplaceResult, error)
+}
+
+// RealProvision drives a one-click make-before-break node replacement via the
+// real provision.Service (doc/04 §4.1): a fresh node is brought UP before the
+// old one drains, so capacity never dips.
+type RealProvision struct {
+	svc provReplacer
+}
+
+// NewRealProvision wires the provision adapter over the real service.
+func NewRealProvision(svc provReplacer) *RealProvision { return &RealProvision{svc: svc} }
+
+// Replace provisions a replacement for nodeID. An empty replacementUUID starts a
+// fresh orchestration (the service mints its own idempotency key). The actor is
+// recorded in the admin audit log by the caller.
+func (p *RealProvision) Replace(ctx context.Context, nodeID int64, _ string) error {
+	_, err := p.svc.Replace(ctx, nodeID, "")
+	return err
+}
+
+// Ready reports true — the provision service is wired. Without vendor
+// credentials the Replace call fails with a clear error when invoked.
+func (*RealProvision) Ready() bool { return true }
+
 // Services bundles the three downstream services the admin UI depends on.
 type Services struct {
 	Codes     CodesService

server/internal/admin/services_real_test.go

@@ -0,0 +1,60 @@
+package admin
+
+import (
+	"context"
+	"errors"
+	"testing"
+
+	"github.com/wangjia/pangolin/server/internal/provision"
+)
+
+// fakeReplacer records the arguments passed to Replace and returns canned values.
+type fakeReplacer struct {
+	gotNodeID int64
+	gotUUID   string
+	res       *provision.ReplaceResult
+	err       error
+}
+
+func (f *fakeReplacer) Replace(_ context.Context, nodeID int64, uuid string) (*provision.ReplaceResult, error) {
+	f.gotNodeID, f.gotUUID = nodeID, uuid
+	return f.res, f.err
+}
+
+func TestRealProvision_Replace_DelegatesWithFreshUUID(t *testing.T) {
+	f := &fakeReplacer{res: &provision.ReplaceResult{NewNodeID: 99}}
+	p := NewRealProvision(f)
+
+	if err := p.Replace(context.Background(), 42, "operator"); err != nil {
+		t.Fatalf("Replace: %v", err)
+	}
+	if f.gotNodeID != 42 {
+		t.Errorf("nodeID = %d, want 42", f.gotNodeID)
+	}
+	// Admin always starts a fresh orchestration: the service mints its own key.
+	if f.gotUUID != "" {
+		t.Errorf("replacementUUID = %q, want empty (fresh start)", f.gotUUID)
+	}
+	if !p.Ready() {
+		t.Error("RealProvision.Ready() = false, want true")
+	}
+}
+
+func TestRealProvision_Replace_PropagatesError(t *testing.T) {
+	sentinel := errors.New("no vendor credentials")
+	p := NewRealProvision(&fakeReplacer{err: sentinel})
+	if err := p.Replace(context.Background(), 7, "operator"); !errors.Is(err, sentinel) {
+		t.Errorf("Replace error = %v, want %v", err, sentinel)
+	}
+}
+
+func TestRealLifecycle_RejectsUnsupportedTarget(t *testing.T) {
+	// Target validation happens before any DB access, so a nil db is safe here.
+	l := &RealLifecycle{}
+	if err := l.TransitionStatus(context.Background(), 1, "destroyed", "operator"); err == nil {
+		t.Error("TransitionStatus with target=destroyed should error, got nil")
+	}
+	if !l.Ready() {
+		t.Error("RealLifecycle.Ready() = false, want true")
+	}
+}