From ebc9c1e70255a19bf5f25c4943ecd855a8331a37 Mon Sep 17 00:00:00 2001
From: wangjia <809946525@qq.com>
Date: Mon, 15 Jun 2026 22:45:56 +0800
Subject: [PATCH] =?UTF-8?q?feat(scheduler):=20=E8=A3=85=E9=85=8D=20schedul?=
 =?UTF-8?q?er=20=E2=80=94=20=E4=B8=89=20loop=20+=20Redis=20=E9=80=89?=
 =?UTF-8?q?=E4=B8=BB=20+=20=E4=BC=98=E9=9B=85=E9=80=80=E5=87=BA=20[tsk=5Fr?=
 =?UTF-8?q?tYVDLkmc5mo]?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

## 内容

### server/internal/scheduler/deps.go
新增 scheduler 级接口文件，定义所有外部依赖接口：
- DetectEngine / OrchestrateReplacer / OrchestrateGrayscale（15D/15E）
- CapacityService + StubCapacityService（15F 存根）
- ProbeStateReader（15A）、ThirdPartyProber / TargetProvider（15C）
- Notifier + LogNotifier（15G）、StaticTargetProvider

### server/internal/scheduler/scheduler.go
Scheduler 主体实现：
- Config 结构体：注入所有依赖 + 可配置的选主时间参数（方便测试）
- Run(ctx) = 安装 SIGTERM/SIGINT 信号 + Start(ctx)
- Start(ctx) = 启动三 loop goroutine + 可选 15C 采集 goroutine + 30s 优雅退出
- runLoop: 非主实例每 5s 抢租约（SET NX PX 30000）
- runLeaderLoop: 主实例按 interval tick，defer 释放租约
- keepLease: 每 10s Lua 校验后续租（PEXPIRE），失败即退出
- releaseLease: Lua 校验后 DEL，加速接管
- capacityTick: 15F check → 15E grayscale.Advance → 15F decay → 探针失联告警
- runThirdPartyProber: 无主从约束全副本跑（理由注释在代码中）

### server/internal/scheduler/wiring.go
BuildStubConfig(): 用存根 LC/Prov 构建可运行的 Config，
供 server main 和集成测试使用；待 #5/#14 就绪后替换真实实现。

### server/internal/scheduler/scheduler_test.go (go test -race 全通过)
- TestLeaderElection: 两实例竞争，任意时刻只有一个在 tick
- TestFollowerTakeover: 主实例退出后备实例在 ≤LeaseTTL 内接管
- TestGracefulShutdown: ctx 取消后三个 leader key 被主动 DEL
- TestCapacityTickProbeDisconnect: 失联探针触发 Notifier
- TestE2EMockScenario: 端到端 mock 剧本
  up→suspect(降权10)→confirmed→down + 替换队列 →
  CreateNode→probing→activating(weight10+BumpVersion+grayscale)→
  draining_old→done，全链 node_events/audit_log 对账

### server/cmd/server/main.go
- 新增 SCHED_ENABLED=true 开关，灰度启动 scheduler
- 复用 REDIS_ADDR/REDIS_PASSWORD，probe.Store 共享给 scheduler
- 通过 signal.NotifyContext 传递 SIGTERM，确保优雅退出

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
---
 server/cmd/server/main.go                   |  50 ++
 server/internal/scheduler/deps.go           | 136 ++++
 server/internal/scheduler/scheduler.go      | 501 ++++++++++++
 server/internal/scheduler/scheduler_test.go | 809 ++++++++++++++++++++
 server/internal/scheduler/wiring.go         | 113 +++
 5 files changed, 1609 insertions(+)
 create mode 100644 server/internal/scheduler/deps.go
 create mode 100644 server/internal/scheduler/scheduler.go
 create mode 100644 server/internal/scheduler/scheduler_test.go
 create mode 100644 server/internal/scheduler/wiring.go

diff --git a/server/cmd/server/main.go b/server/cmd/server/main.go
index 999898b..30375a3 100644
--- a/server/cmd/server/main.go
+++ b/server/cmd/server/main.go
@@ -10,6 +10,8 @@ import (
 	"net"
 	"net/http"
 	"os"
+	"os/signal"
+	"syscall"
 
 	"github.com/go-chi/chi/v5"
 	"github.com/go-chi/chi/v5/middleware"
@@ -24,6 +26,7 @@ import (
 	"github.com/wangjia/pangolin/server/internal/nodes"
 	agentv1 "github.com/wangjia/pangolin/server/internal/pb/agentv1"
 	"github.com/wangjia/pangolin/server/internal/redisutil"
+	"github.com/wangjia/pangolin/server/internal/scheduler"
 	"github.com/wangjia/pangolin/server/internal/scheduler/probe"
 )
 
@@ -66,6 +69,7 @@ func main() {
 	// If PROBE_SECRETS is empty the route is not registered and the server
 	// starts normally without the probe ingest endpoint.
 	probeSecretsJSON := os.Getenv("PROBE_SECRETS")
+	var sharedProbeStore *probe.Store // may be re-used by the scheduler
 	if probeSecretsJSON != "" {
 		redisAddr := getenvDefault("REDIS_ADDR", "127.0.0.1:6379")
 		rdb, err := redisutil.New(redisAddr, os.Getenv("REDIS_PASSWORD"), 0)
@@ -78,12 +82,58 @@ func main() {
 			}
 			reg := probe.NewMapRegistry(secretMap)
 			st := probe.NewStore(rdb)
+			sharedProbeStore = st
 			h := probe.NewIngestHandler(reg, st)
 			r.Post("/probe/report", h.ServeHTTP)
 			log.Printf("probe ingest route registered (%d probe(s))", len(secretMap))
 		}
 	}
 
+	// ─── Scheduler (optional) ─────────────────────────────────────────────────
+	//
+	// Set SCHED_ENABLED=true to start the three leader-elected scheduler loops
+	// (DetectLoop / OrchestrateLoop / CapacityLoop) in this process alongside the
+	// HTTP and gRPC servers.
+	//
+	// Dependencies:
+	//   REDIS_ADDR, REDIS_PASSWORD – shared with the probe route (new client if
+	//                                 the probe route is disabled).
+	//
+	// Rollback: set SCHED_ENABLED=false (or unset it) and restart; all other
+	// server functionality is unaffected.
+	if os.Getenv("SCHED_ENABLED") == "true" {
+		redisAddr := getenvDefault("REDIS_ADDR", "127.0.0.1:6379")
+		schedRDB, err := redisutil.New(redisAddr, os.Getenv("REDIS_PASSWORD"), 0)
+		if err != nil {
+			log.Printf("scheduler: redis connect failed (%v) — scheduler disabled", err)
+		} else {
+			// Use the shared probe store if the probe route is already wired up;
+			// otherwise create a standalone store backed by the scheduler Redis client.
+			ps := sharedProbeStore
+			if ps == nil {
+				ps = probe.NewStore(schedRDB)
+			}
+			cfg := scheduler.BuildStubConfig(schedRDB, ps)
+			// TODO(#5): replace stub lifecycle with real LifecycleService
+			// TODO(#14): replace stub provision with real ProvisionService
+			sched := scheduler.New(cfg)
+
+			// The scheduler installs its own SIGTERM/SIGINT handler via Run().
+			// We also respect the Go context tree: sigCtx cancels when the process
+			// receives a signal, giving the scheduler up to TickTimeout (30 s) to
+			// finish any in-flight tick before the OS kills the process.
+			sigCtx, stopSig := signal.NotifyContext(context.Background(),
+				os.Interrupt, syscall.SIGTERM)
+			go func() {
+				defer stopSig()
+				if err := sched.Start(sigCtx); err != nil {
+					slog.Error("scheduler: Start error", "error", err)
+				}
+			}()
+			log.Printf("scheduler started (SCHED_ENABLED=true, stub lifecycle/provision)")
+		}
+	}
+
 	// Mount all /v1/... routes.  HandlerFromMuxWithBaseURL registers every route
 	// from the OpenAPI spec onto the provided chi router with the given prefix,
 	// so the router itself is the http.Handler we serve.
diff --git a/server/internal/scheduler/deps.go b/server/internal/scheduler/deps.go
new file mode 100644
index 0000000..dbd4151
--- /dev/null
+++ b/server/internal/scheduler/deps.go
@@ -0,0 +1,136 @@
+// Package scheduler wires together the 15D detection engine, 15E orchestration
+// engine, 15F capacity service, 15A probe store, and 15G notifier into three
+// leader-elected goroutine loops (DetectLoop / OrchestrateLoop / CapacityLoop),
+// plus an optional unguarded 15C third-party prober goroutine.
+//
+// This file declares the narrow dependency interfaces used by Scheduler.  Real
+// implementations satisfy these interfaces; test mocks do too.
+package scheduler
+
+import (
+	"context"
+	"log/slog"
+
+	"github.com/wangjia/pangolin/server/internal/scheduler/probe"
+)
+
+// ─────────────────────────────────────────────────────────────────────────────
+// 15D – detection engine
+// ─────────────────────────────────────────────────────────────────────────────
+
+// DetectEngine is the 15D interface.  *detect.Engine satisfies it.
+// Tick is called by DetectLoop on every 5-minute cycle.
+type DetectEngine interface {
+	Tick(ctx context.Context) error
+}
+
+// ─────────────────────────────────────────────────────────────────────────────
+// 15E – orchestration engine
+// ─────────────────────────────────────────────────────────────────────────────
+
+// OrchestrateReplacer is the 15E replacement-orchestration interface.
+// *orchestrate.Replacer satisfies it.
+// Tick is called by OrchestrateLoop every 30 seconds.
+type OrchestrateReplacer interface {
+	Tick(ctx context.Context) error
+}
+
+// OrchestrateGrayscale is the 15E grayscale warm-up interface.
+// *orchestrate.Grayscale satisfies it.
+// Advance is called by CapacityLoop every 1–2 minutes.
+type OrchestrateGrayscale interface {
+	Advance(ctx context.Context) error
+}
+
+// ─────────────────────────────────────────────────────────────────────────────
+// 15F – capacity service (stub until task 15F is delivered)
+// ─────────────────────────────────────────────────────────────────────────────
+
+// CapacityService is the 15F interface for watermark/quota checking and
+// circuit-breaker time-decay.  Replace StubCapacityService with the real
+// 15F implementation once that task is complete.
+type CapacityService interface {
+	// Check verifies that the node-pool watermark and per-region quotas are
+	// within acceptable bounds.  Non-fatal: errors are logged, not returned.
+	Check(ctx context.Context) error
+
+	// Decay applies time-based decay to circuit-breaker event counters so that
+	// temporary bursts do not permanently block replacements.
+	Decay(ctx context.Context) error
+}
+
+// StubCapacityService is a no-op CapacityService used until task 15F is ready.
+type StubCapacityService struct{}
+
+// Check implements CapacityService (always succeeds).
+func (StubCapacityService) Check(_ context.Context) error { return nil }
+
+// Decay implements CapacityService (always succeeds).
+func (StubCapacityService) Decay(_ context.Context) error { return nil }
+
+// ─────────────────────────────────────────────────────────────────────────────
+// 15A – probe state reader
+// ─────────────────────────────────────────────────────────────────────────────
+
+// ProbeStateReader reads the heartbeat liveness of probe agents from the 15A
+// Redis store.  *probe.Store satisfies it (via AliveProbes).
+type ProbeStateReader interface {
+	// AliveProbes returns the IDs of probe agents that have sent a heartbeat
+	// within the last heartbeatTTL window (15 min).  An empty result means
+	// "no recent heartbeats", not "all probes are down".
+	AliveProbes(ctx context.Context) ([]string, error)
+}
+
+// ─────────────────────────────────────────────────────────────────────────────
+// 15C – third-party prober
+// ─────────────────────────────────────────────────────────────────────────────
+
+// ThirdPartyProber is the 15C interface.
+// *probe.AliyunSyntheticAgent satisfies it.
+type ThirdPartyProber interface {
+	// RunOnce issues a full multi-ISP probe cycle for all targets and writes
+	// results to Redis via the 15A probe store.
+	RunOnce(ctx context.Context, targets []probe.ProbeTarget) error
+}
+
+// TargetProvider supplies the current list of active nodes to be probed via
+// the 15C third-party service.  Typically backed by a DB query; a static
+// implementation is available as StaticTargetProvider.
+type TargetProvider interface {
+	ListProbeTargets(ctx context.Context) ([]probe.ProbeTarget, error)
+}
+
+// StaticTargetProvider returns a fixed, pre-configured list of probe targets.
+// Useful for initial deployment and integration tests before a DB-backed
+// provider is available.
+type StaticTargetProvider struct {
+	Targets []probe.ProbeTarget
+}
+
+// ListProbeTargets implements TargetProvider.
+func (p StaticTargetProvider) ListProbeTargets(_ context.Context) ([]probe.ProbeTarget, error) {
+	return p.Targets, nil
+}
+
+// ─────────────────────────────────────────────────────────────────────────────
+// 15G – notifier
+// ─────────────────────────────────────────────────────────────────────────────
+
+// Notifier is the 15G alerting interface.  The real implementation sends a
+// Telegram (or similar) alert; LogNotifier is the stub.
+type Notifier interface {
+	NotifyFault(ctx context.Context, nodeID, reason string) error
+}
+
+// LogNotifier is a Notifier stub that writes to slog.
+// Used when no real 15G notifier is wired.
+type LogNotifier struct{}
+
+// NotifyFault implements Notifier.
+func (LogNotifier) NotifyFault(_ context.Context, nodeID, reason string) error {
+	slog.Warn("scheduler: node fault — manual review required",
+		"node_id", nodeID,
+		"reason", reason,
+	)
+	return nil
+}
diff --git a/server/internal/scheduler/scheduler.go b/server/internal/scheduler/scheduler.go
new file mode 100644
index 0000000..993e952
--- /dev/null
+++ b/server/internal/scheduler/scheduler.go
@@ -0,0 +1,501 @@
+package scheduler
+
+import (
+	"context"
+	"log/slog"
+	"os"
+	"os/signal"
+	"sync"
+	"syscall"
+	"time"
+
+	"github.com/google/uuid"
+	"github.com/redis/go-redis/v9"
+)
+
+// ─────────────────────────────────────────────────────────────────────────────
+// Leader-election Redis key helper
+// ─────────────────────────────────────────────────────────────────────────────
+
+// leaderKey returns the Redis key for a loop's leader election.
+// Format: sched:leader:{loopName}
+func leaderKey(loop string) string { return "sched:leader:" + loop }
+
+// ─────────────────────────────────────────────────────────────────────────────
+// Config
+// ─────────────────────────────────────────────────────────────────────────────
+
+// Config holds all scheduler dependencies and tunable timing parameters.
+// Zero values for duration fields use the production defaults listed in the
+// field comments.  All wired components (Engine, Replacer, …) are required
+// unless the field is explicitly marked optional.
+type Config struct {
+	// RDB is the Redis client used for leader-election keys.  Required.
+	RDB *redis.Client
+
+	// InstanceID uniquely identifies this process instance across the fleet.
+	// If empty, a random UUID is generated by New.
+	InstanceID string
+
+	// ── Wired engine components ──────────────────────────────────────────────
+
+	// Engine is the 15D detection engine.  Required.
+	Engine DetectEngine
+
+	// Replacer is the 15E replacement-orchestration Replacer.  Required.
+	Replacer OrchestrateReplacer
+
+	// Grayscale is the 15E grayscale warm-up driver.  Required.
+	Grayscale OrchestrateGrayscale
+
+	// Capacity is the 15F capacity/circuit-breaker service.
+	// Optional — StubCapacityService is used when nil.
+	Capacity CapacityService
+
+	// ProbeStore reads probe-agent heartbeat liveness (15A).
+	// Optional — probe disconnection checking is disabled when nil.
+	ProbeStore ProbeStateReader
+
+	// KnownProbeIDs is the set of probe-agent IDs expected to be alive.
+	// When any ID in this list is absent from ProbeStore.AliveProbes, the
+	// Notifier is called with reason "probe_disconnected".
+	KnownProbeIDs []string
+
+	// Prober is the 15C third-party dial-test driver.
+	// Optional — 15C probing is disabled when nil.
+	Prober ThirdPartyProber
+
+	// Targets supplies the list of nodes to probe via Prober.
+	// Required when Prober is non-nil.
+	Targets TargetProvider
+
+	// Notifier is the 15G alerting sink.
+	// Optional — LogNotifier is used when nil.
+	Notifier Notifier
+
+	// ── Tick intervals (hot-reloadable from 15F config) ─────────────────────
+	// Zero → production default (shown in comments).
+
+	DetectInterval      time.Duration // 5 min
+	OrchestrateInterval time.Duration // 30 s
+	CapacityInterval    time.Duration // 90 s
+	ThirdPartyInterval  time.Duration // 10 min — interval between 15C RunOnce calls
+
+	// ── Leader-election timing ────────────────────────────────────────────────
+	// Override only in tests (shorter values speed up convergence).
+
+	LeaseTTL    time.Duration // 30 s  — Redis PEXPIRE on the leader key
+	RenewPeriod time.Duration // 10 s  — how often the leader renews its lease
+	RetryPeriod time.Duration // 5 s   — how often a non-leader tries to acquire
+	TickTimeout time.Duration // 30 s  — per-tick context hard-timeout
+}
+
+// applyDefaults fills zero-value duration fields with production defaults.
+func (c *Config) applyDefaults() {
+	if c.InstanceID == "" {
+		c.InstanceID = uuid.NewString()
+	}
+	if c.DetectInterval == 0 {
+		c.DetectInterval = 5 * time.Minute
+	}
+	if c.OrchestrateInterval == 0 {
+		c.OrchestrateInterval = 30 * time.Second
+	}
+	if c.CapacityInterval == 0 {
+		c.CapacityInterval = 90 * time.Second
+	}
+	if c.ThirdPartyInterval == 0 {
+		c.ThirdPartyInterval = 10 * time.Minute
+	}
+	if c.LeaseTTL == 0 {
+		c.LeaseTTL = 30 * time.Second
+	}
+	if c.RenewPeriod == 0 {
+		c.RenewPeriod = 10 * time.Second
+	}
+	if c.RetryPeriod == 0 {
+		c.RetryPeriod = 5 * time.Second
+	}
+	if c.TickTimeout == 0 {
+		c.TickTimeout = 30 * time.Second
+	}
+	if c.Notifier == nil {
+		c.Notifier = LogNotifier{}
+	}
+	if c.Capacity == nil {
+		c.Capacity = StubCapacityService{}
+	}
+}
+
+// ─────────────────────────────────────────────────────────────────────────────
+// Scheduler
+// ─────────────────────────────────────────────────────────────────────────────
+
+// Scheduler runs three leader-elected loops (DetectLoop / OrchestrateLoop /
+// CapacityLoop) plus an optional unguarded 15C third-party-prober goroutine.
+//
+// Design rationale for three independent leader leases (one per loop):
+//   - A slow Detect tick (e.g., DB latency spike) must not block Orchestrate
+//     progress on other in-flight replacements.
+//   - Each loop failing over independently avoids a partial-leader split where
+//     a crashing replica holds only one of three leases.
+//   - Raft-style leadership is overkill for these loosely coupled loops, which
+//     are each idempotent and use Redis-side locking for concurrency safety.
+type Scheduler struct {
+	rdb  *redis.Client
+	cfg  Config
+}
+
+// New creates a Scheduler from cfg, applying production defaults for any nil /
+// zero fields.
+func New(cfg Config) *Scheduler {
+	cfg.applyDefaults()
+	return &Scheduler{rdb: cfg.RDB, cfg: cfg}
+}
+
+// Run starts the scheduler with OS signal handling (SIGTERM / SIGINT) and
+// blocks until the scheduler has shut down.  For production use from main().
+func (s *Scheduler) Run(ctx context.Context) error {
+	sigCtx, stop := signal.NotifyContext(ctx, os.Interrupt, syscall.SIGTERM)
+	defer stop()
+	return s.Start(sigCtx)
+}
+
+// Start starts the scheduler goroutines and blocks until ctx is cancelled.
+// Unlike Run, it does not install OS signal handlers — useful for testing.
+func (s *Scheduler) Start(ctx context.Context) error {
+	var wg sync.WaitGroup
+
+	// ── Three leader-elected loops ─────────────────────────────────────────────
+	//
+	// Each loop acquires its own independent Redis leader key so that a slow
+	// tick in one loop does not delay the others.
+
+	wg.Add(3)
+	go func() {
+		defer wg.Done()
+		s.runLoop(ctx, "detect", s.cfg.DetectInterval, s.cfg.Engine.Tick)
+	}()
+	go func() {
+		defer wg.Done()
+		s.runLoop(ctx, "orchestrate", s.cfg.OrchestrateInterval, s.cfg.Replacer.Tick)
+	}()
+	go func() {
+		defer wg.Done()
+		s.runLoop(ctx, "capacity", s.cfg.CapacityInterval, s.capacityTick)
+	}()
+
+	// ── 15C third-party prober (no leader constraint) ─────────────────────────
+	//
+	// Design choice: run on ALL replicas without leader election because:
+	//   1. Alibaba Cloud API calls are idempotent: same (node, ISP) pair creates
+	//      a new one-shot task; the result is written to Redis via SET (overwrite)
+	//      so duplicate writes produce the same snapshot with a slightly later
+	//      ReceivedAt — harmless.
+	//   2. Running on every replica improves data collection reliability: if one
+	//      replica crashes mid-cycle, the others continue gathering snapshots.
+	//   3. The detect engine (15D) aggregates snapshots regardless of source;
+	//      receiving the same data twice per window has no logical effect.
+	//
+	// If billing for N×API calls becomes a concern, add leader election here by
+	// wrapping the loop body in tryRunAsLeader("prober").
+	if s.cfg.Prober != nil && s.cfg.Targets != nil {
+		wg.Add(1)
+		go func() {
+			defer wg.Done()
+			s.runThirdPartyProber(ctx)
+		}()
+	}
+
+	// Wait for ctx cancellation (SIGTERM / test cancel).
+	<-ctx.Done()
+	slog.Info("scheduler: shutdown signal received — waiting for in-flight ticks to finish",
+		"instance", s.cfg.InstanceID)
+
+	// 30-second hard timeout: all loops must finish their current tick within
+	// this window.  Leader keys are released by each loop's defer (see
+	// runLeaderLoop → releaseLease) so followers can take over immediately.
+	allStopped := make(chan struct{})
+	go func() {
+		wg.Wait()
+		close(allStopped)
+	}()
+
+	shutdownTimeout := s.cfg.TickTimeout // reuse tick hard-timeout as shutdown budget
+	select {
+	case <-allStopped:
+		slog.Info("scheduler: graceful shutdown complete", "instance", s.cfg.InstanceID)
+	case <-time.After(shutdownTimeout):
+		slog.Warn("scheduler: shutdown timeout exceeded — forcing exit",
+			"timeout", shutdownTimeout, "instance", s.cfg.InstanceID)
+	}
+	return nil
+}
+
+// ─────────────────────────────────────────────────────────────────────────────
+// Leader-elected loop driver
+// ─────────────────────────────────────────────────────────────────────────────
+
+// runLoop is the generic leader-elected loop driver.
+//
+//   - Non-leader: polls every RetryPeriod to acquire sched:leader:{name}.
+//   - Leader: ticks at interval; a background goroutine renews the lease every
+//     RenewPeriod.  Lost lease → return (exit from runLeaderLoop).
+//
+// The function loops continuously until ctx is Done.
+func (s *Scheduler) runLoop(ctx context.Context, name string, interval time.Duration, tick func(context.Context) error) {
+	key := leaderKey(name)
+
+	for {
+		// Try to acquire the leader lease (SET NX PX <leaseTTL ms>).
+		ok, err := s.rdb.SetNX(ctx, key, s.cfg.InstanceID, s.cfg.LeaseTTL).Result()
+		if ctx.Err() != nil {
+			return // context cancelled — don't log spurious errors
+		}
+		if err != nil {
+			slog.Error("scheduler: leader setnx error", "loop", name, "error", err)
+		} else if ok {
+			slog.Info("scheduler: acquired leader lease", "loop", name, "instance", s.cfg.InstanceID)
+			s.runLeaderLoop(ctx, name, interval, tick, key)
+			slog.Info("scheduler: relinquished leader", "loop", name, "instance", s.cfg.InstanceID)
+		}
+
+		// Not leader (or just stepped down): wait before retrying.
+		select {
+		case <-ctx.Done():
+			return
+		case <-time.After(s.cfg.RetryPeriod):
+		}
+	}
+}
+
+// runLeaderLoop ticks at interval while holding the leader lease.
+// It starts a keepLease goroutine; when the lease is lost (renewal failure or
+// key taken by a competitor), it returns so runLoop can retry.
+//
+// Graceful-exit guarantee: when ctx is cancelled, runLeaderLoop waits for any
+// in-flight tick to finish (the tick is given its own TickTimeout context, not
+// the parent ctx) before calling defer → releaseLease.  This ensures:
+//   - No half-executed migrate/notify steps are left in Redis.
+//   - The leader key is actively released (DEL) on clean shutdown, so the
+//     follower can take over within RetryPeriod rather than LeaseTTL.
+func (s *Scheduler) runLeaderLoop(
+	ctx context.Context,
+	name string,
+	interval time.Duration,
+	tick func(context.Context) error,
+	leaseKey string,
+) {
+	// Renewal goroutine: signals via leaseLost when it can no longer renew.
+	renewCtx, cancelRenew := context.WithCancel(ctx)
+	leaseLost := make(chan struct{})
+	go func() {
+		defer close(leaseLost)
+		s.keepLease(renewCtx, leaseKey, name)
+	}()
+
+	// On exit: stop the renewer and release the key so followers converge faster.
+	defer func() {
+		cancelRenew()
+		<-leaseLost // wait for renewer before releasing
+		// Use a fresh Background context: the original ctx may already be Done.
+		s.releaseLease(context.Background(), leaseKey, name)
+	}()
+
+	ticker := time.NewTicker(interval)
+	defer ticker.Stop()
+
+	for {
+		select {
+		case <-ctx.Done():
+			// Shutdown requested: exit (defer will release lease).
+			// If a tick was already dispatched before this select was reached,
+			// it is still running under its own TickTimeout context and will
+			// complete before this goroutine is joined by Start's wg.Wait().
+			return
+
+		case <-leaseLost:
+			slog.Warn("scheduler: leader lease lost — stepping down", "loop", name)
+			return
+
+		case <-ticker.C:
+			// Run tick with a private context so SIGTERM / ctx cancel does not
+			// abort it mid-flight.  Idempotency (15D optimistic lock, 15E
+			// idempotency keys) ensures a tick that completes after the leader
+			// key expires is safe.
+			tickCtx, cancelTick := context.WithTimeout(context.Background(), s.cfg.TickTimeout)
+			if err := tick(tickCtx); err != nil {
+				slog.Error("scheduler: tick error", "loop", name, "error", err)
+			}
+			cancelTick()
+
+			// If parent was cancelled while we were ticking, exit now instead
+			// of waiting for the next ticker.C.
+			if ctx.Err() != nil {
+				return
+			}
+		}
+	}
+}
+
+// ─────────────────────────────────────────────────────────────────────────────
+// Lease renewal / release
+// ─────────────────────────────────────────────────────────────────────────────
+
+// Lua scripts are pre-evaluated to validate key ownership before each
+// PEXPIRE / DEL so we never extend or delete a key owned by another instance.
+
+const luaRenew = `
+if redis.call('GET', KEYS[1]) == ARGV[1] then
+    return redis.call('PEXPIRE', KEYS[1], ARGV[2])
+else
+    return 0
+end`
+
+const luaRelease = `
+if redis.call('GET', KEYS[1]) == ARGV[1] then
+    return redis.call('DEL', KEYS[1])
+else
+    return 0
+end`
+
+// keepLease renews the leader key every RenewPeriod using the Lua renewal
+// script.  Returns when renewal fails (key gone or taken) or ctx is Done.
+func (s *Scheduler) keepLease(ctx context.Context, key, name string) {
+	ticker := time.NewTicker(s.cfg.RenewPeriod)
+	defer ticker.Stop()
+
+	for {
+		select {
+		case <-ctx.Done():
+			return
+		case <-ticker.C:
+			res, err := s.rdb.Eval(
+				ctx, luaRenew, []string{key},
+				s.cfg.InstanceID,
+				s.cfg.LeaseTTL.Milliseconds(),
+			).Result()
+			if err != nil {
+				if ctx.Err() != nil {
+					return // context cancelled: not a real failure
+				}
+				slog.Error("scheduler: lease renewal error", "loop", name, "error", err)
+				return // treat Redis error as lease loss
+			}
+			if res.(int64) == 0 {
+				slog.Warn("scheduler: lease renewal returned 0 — key gone or taken",
+					"loop", name, "key", key)
+				return
+			}
+		}
+	}
+}
+
+// releaseLease deletes the leader key only if it still belongs to this
+// instance, using the Lua release script.
+func (s *Scheduler) releaseLease(ctx context.Context, key, name string) {
+	res, err := s.rdb.Eval(ctx, luaRelease, []string{key}, s.cfg.InstanceID).Result()
+	if err != nil {
+		slog.Error("scheduler: release lease error", "loop", name, "error", err)
+		return
+	}
+	if res.(int64) == 1 {
+		slog.Info("scheduler: leader key released",
+			"loop", name, "instance", s.cfg.InstanceID)
+	}
+}
+
+// ─────────────────────────────────────────────────────────────────────────────
+// Capacity tick (CapacityLoop work function)
+// ─────────────────────────────────────────────────────────────────────────────
+
+// capacityTick is the work function for CapacityLoop.  It runs:
+//  1. 15F capacity / watermark check.
+//  2. 15E grayscale weight advancement.
+//  3. 15F circuit-breaker counter decay.
+//  4. 15A probe-agent disconnection check → 15G alert.
+//
+// Each step is independent; a failure in one does not abort the others.
+func (s *Scheduler) capacityTick(ctx context.Context) error {
+	// 1. Capacity / watermark check (15F).
+	if err := s.cfg.Capacity.Check(ctx); err != nil {
+		slog.Error("scheduler: capacity check error", "error", err)
+		// Non-fatal: continue.
+	}
+
+	// 2. Grayscale warm-up advancement (15E).
+	if err := s.cfg.Grayscale.Advance(ctx); err != nil {
+		slog.Error("scheduler: grayscale advance error", "error", err)
+	}
+
+	// 3. Circuit-breaker counter decay (15F).
+	if err := s.cfg.Capacity.Decay(ctx); err != nil {
+		slog.Error("scheduler: capacity decay error", "error", err)
+	}
+
+	// 4. Probe-agent disconnection check (15A probe store → 15G alert).
+	if s.cfg.ProbeStore != nil && len(s.cfg.KnownProbeIDs) > 0 {
+		s.checkProbeConnections(ctx)
+	}
+
+	return nil
+}
+
+// checkProbeConnections fires a Notifier alert for every known probe agent
+// whose heartbeat key has expired (i.e. is absent from AliveProbes).
+func (s *Scheduler) checkProbeConnections(ctx context.Context) {
+	alive, err := s.cfg.ProbeStore.AliveProbes(ctx)
+	if err != nil {
+		slog.Error("scheduler: AliveProbes error", "error", err)
+		return
+	}
+
+	aliveSet := make(map[string]struct{}, len(alive))
+	for _, id := range alive {
+		aliveSet[id] = struct{}{}
+	}
+
+	for _, id := range s.cfg.KnownProbeIDs {
+		if _, ok := aliveSet[id]; !ok {
+			if notifyErr := s.cfg.Notifier.NotifyFault(ctx, id, "probe_disconnected"); notifyErr != nil {
+				slog.Error("scheduler: notify probe disconnect",
+					"probe_id", id, "error", notifyErr)
+			}
+		}
+	}
+}
+
+// ─────────────────────────────────────────────────────────────────────────────
+// Third-party prober goroutine (15C, no leader constraint)
+// ─────────────────────────────────────────────────────────────────────────────
+
+// runThirdPartyProber fires Prober.RunOnce at ThirdPartyInterval on every
+// replica (no leader election — see design comment in Start).
+func (s *Scheduler) runThirdPartyProber(ctx context.Context) {
+	ticker := time.NewTicker(s.cfg.ThirdPartyInterval)
+	defer ticker.Stop()
+
+	for {
+		select {
+		case <-ctx.Done():
+			return
+		case <-ticker.C:
+			targets, err := s.cfg.Targets.ListProbeTargets(ctx)
+			if err != nil {
+				slog.Error("scheduler: list probe targets for 15C", "error", err)
+				continue
+			}
+			if len(targets) == 0 {
+				continue
+			}
+			// Use a fresh context so ctx cancellation does not abort a running
+			// prober call (the degradation contract: result is "no data" on
+			// context error, which 15D treats as unknown, not failure).
+			probeCtx, cancel := context.WithTimeout(context.Background(), s.cfg.TickTimeout)
+			if err := s.cfg.Prober.RunOnce(probeCtx, targets); err != nil {
+				slog.Error("scheduler: 15C prober error", "error", err)
+			}
+			cancel()
+		}
+	}
+}
diff --git a/server/internal/scheduler/scheduler_test.go b/server/internal/scheduler/scheduler_test.go
new file mode 100644
index 0000000..974c41a
--- /dev/null
+++ b/server/internal/scheduler/scheduler_test.go
@@ -0,0 +1,809 @@
+package scheduler_test
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"sync"
+	"sync/atomic"
+	"testing"
+	"time"
+
+	"github.com/alicebob/miniredis/v2"
+	"github.com/redis/go-redis/v9"
+
+	"github.com/wangjia/pangolin/server/internal/scheduler"
+	"github.com/wangjia/pangolin/server/internal/scheduler/detect"
+	"github.com/wangjia/pangolin/server/internal/scheduler/orchestrate"
+	"github.com/wangjia/pangolin/server/internal/scheduler/probe"
+)
+
+// ─────────────────────────────────────────────────────────────────────────────
+// Shared test helpers
+// ─────────────────────────────────────────────────────────────────────────────
+
+func newTestRedis(t *testing.T) (*redis.Client, *miniredis.Miniredis) {
+	t.Helper()
+	mr := miniredis.RunT(t)
+	rdb := redis.NewClient(&redis.Options{Addr: mr.Addr()})
+	return rdb, mr
+}
+
+// testIntervals returns Config timing values suitable for fast unit tests.
+func testIntervals() (leaseTTL, renew, retry, tick, interval time.Duration) {
+	return 200 * time.Millisecond, // LeaseTTL
+		50 * time.Millisecond,  // RenewPeriod
+		50 * time.Millisecond,  // RetryPeriod
+		5 * time.Second,        // TickTimeout
+		80 * time.Millisecond   // loop interval
+}
+
+// counterTick returns a tick function that atomically increments *n.
+func counterTick(n *int64) func(context.Context) error {
+	return func(_ context.Context) error {
+		atomic.AddInt64(n, 1)
+		return nil
+	}
+}
+
+// ─────────────────────────────────────────────────────────────────────────────
+// Stub implementations for scheduler.Config fields
+// ─────────────────────────────────────────────────────────────────────────────
+
+// stubEngine implements scheduler.DetectEngine.
+type stubEngine struct{ count int64 }
+
+func (e *stubEngine) Tick(_ context.Context) error { atomic.AddInt64(&e.count, 1); return nil }
+
+// stubReplacer implements scheduler.OrchestrateReplacer.
+type stubReplacer struct{ count int64 }
+
+func (r *stubReplacer) Tick(_ context.Context) error { atomic.AddInt64(&r.count, 1); return nil }
+
+// stubGrayscale implements scheduler.OrchestrateGrayscale.
+type stubGrayscale struct{ count int64 }
+
+func (g *stubGrayscale) Advance(_ context.Context) error {
+	atomic.AddInt64(&g.count, 1)
+	return nil
+}
+
+// stubProbeStore implements scheduler.ProbeStateReader.
+type stubProbeStore struct {
+	mu    sync.Mutex
+	alive []string
+}
+
+func (s *stubProbeStore) AliveProbes(_ context.Context) ([]string, error) {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	return append([]string{}, s.alive...), nil
+}
+
+func (s *stubProbeStore) setAlive(ids ...string) {
+	s.mu.Lock(); defer s.mu.Unlock()
+	s.alive = ids
+}
+
+// recordingNotifier captures NotifyFault calls.
+type recordingNotifier struct {
+	mu     sync.Mutex
+	calls  []string // nodeID values
+}
+
+func (r *recordingNotifier) NotifyFault(_ context.Context, nodeID, _ string) error {
+	r.mu.Lock(); defer r.mu.Unlock()
+	r.calls = append(r.calls, nodeID)
+	return nil
+}
+
+func (r *recordingNotifier) callCount() int {
+	r.mu.Lock(); defer r.mu.Unlock()
+	return len(r.calls)
+}
+
+func (r *recordingNotifier) called(nodeID string) bool {
+	r.mu.Lock(); defer r.mu.Unlock()
+	for _, id := range r.calls {
+		if id == nodeID { return true }
+	}
+	return false
+}
+
+// ─────────────────────────────────────────────────────────────────────────────
+// TestLeaderElection: only one of two competing instances runs ticks
+// ─────────────────────────────────────────────────────────────────────────────
+
+func TestLeaderElection(t *testing.T) {
+	rdb, _ := newTestRedis(t)
+	leaseTTL, renew, retry, tickTout, interval := testIntervals()
+
+	var tickA, tickB int64
+	makeEngine := func(n *int64) *stubEngine {
+		return &stubEngine{}
+	}
+	_ = makeEngine
+
+	buildCfg := func(id string, engine *stubEngine, replacer *stubReplacer, gray *stubGrayscale) scheduler.Config {
+		return scheduler.Config{
+			RDB:                 rdb,
+			InstanceID:          id,
+			Engine:              engine,
+			Replacer:            replacer,
+			Grayscale:           gray,
+			DetectInterval:      interval,
+			OrchestrateInterval: interval,
+			CapacityInterval:    interval,
+			LeaseTTL:            leaseTTL,
+			RenewPeriod:         renew,
+			RetryPeriod:         retry,
+			TickTimeout:         tickTout,
+		}
+	}
+
+	engA := &stubEngine{}
+	repA := &stubReplacer{}
+	gryA := &stubGrayscale{}
+	schedA := scheduler.New(buildCfg("inst-A", engA, repA, gryA))
+
+	engB := &stubEngine{}
+	repB := &stubReplacer{}
+	gryB := &stubGrayscale{}
+	schedB := scheduler.New(buildCfg("inst-B", engB, repB, gryB))
+
+	ctxA, cancelA := context.WithCancel(context.Background())
+	ctxB, cancelB := context.WithCancel(context.Background())
+	defer cancelA()
+	defer cancelB()
+
+	// Run both schedulers; they compete for "detect" / "orchestrate" / "capacity" leases.
+	var wgA, wgB sync.WaitGroup
+	wgA.Add(1)
+	go func() { defer wgA.Done(); _ = schedA.Start(ctxA) }()
+	wgB.Add(1)
+	go func() { defer wgB.Done(); _ = schedB.Start(ctxB) }()
+
+	// Let them run for a few tick intervals.
+	time.Sleep(500 * time.Millisecond)
+
+	// Collect tick counts.
+	tickA = atomic.LoadInt64(&engA.count)
+	tickB = atomic.LoadInt64(&engB.count)
+	t.Logf("after 500ms: tickA=%d tickB=%d", tickA, tickB)
+
+	// At most one instance should have run detect ticks.
+	if tickA > 0 && tickB > 0 {
+		t.Errorf("both inst-A and inst-B ran detect ticks — leader election broken (A=%d B=%d)", tickA, tickB)
+	}
+	if tickA == 0 && tickB == 0 {
+		t.Error("neither instance ran any detect ticks — scheduler not ticking")
+	}
+}
+
+// ─────────────────────────────────────────────────────────────────────────────
+// TestFollowerTakeover: follower takes over after leader's context is cancelled
+// ─────────────────────────────────────────────────────────────────────────────
+
+func TestFollowerTakeover(t *testing.T) {
+	rdb, mr := newTestRedis(t)
+	leaseTTL, renew, retry, tickTout, interval := testIntervals()
+
+	buildCfg := func(id string) scheduler.Config {
+		return scheduler.Config{
+			RDB:                 rdb,
+			InstanceID:          id,
+			Engine:              &stubEngine{},
+			Replacer:            &stubReplacer{},
+			Grayscale:           &stubGrayscale{},
+			DetectInterval:      interval,
+			OrchestrateInterval: interval,
+			CapacityInterval:    interval,
+			LeaseTTL:            leaseTTL,
+			RenewPeriod:         renew,
+			RetryPeriod:         retry,
+			TickTimeout:         tickTout,
+		}
+	}
+
+	// Use a simple counting tick for "detect" loop wired via Engine.
+	var ticksLeader, ticksFollower int64
+
+	engLeader := &counterEngine{n: &ticksLeader}
+	engFollower := &counterEngine{n: &ticksFollower}
+
+	cfgLeader := buildCfg("leader")
+	cfgLeader.Engine = engLeader
+	cfgFollower := buildCfg("follower")
+	cfgFollower.Engine = engFollower
+
+	schedLeader := scheduler.New(cfgLeader)
+	schedFollower := scheduler.New(cfgFollower)
+
+	ctxLeader, cancelLeader := context.WithCancel(context.Background())
+	ctxFollower, cancelFollower := context.WithCancel(context.Background())
+	defer cancelFollower()
+
+	var wgLeader sync.WaitGroup
+	wgLeader.Add(1)
+	go func() { defer wgLeader.Done(); _ = schedLeader.Start(ctxLeader) }()
+
+	go func() { _ = schedFollower.Start(ctxFollower) }()
+
+	// Wait for leader to tick at least once.
+	for i := 0; i < 50; i++ {
+		if atomic.LoadInt64(&ticksLeader) > 0 {
+			break
+		}
+		time.Sleep(20 * time.Millisecond)
+	}
+	if atomic.LoadInt64(&ticksLeader) == 0 {
+		t.Fatal("leader never ticked")
+	}
+
+	prevFollower := atomic.LoadInt64(&ticksFollower)
+
+	// Stop leader — its defer releases the leader key immediately.
+	cancelLeader()
+	wgLeader.Wait()
+
+	// Fast-forward miniredis clock to expire the lease (belt-and-suspenders for
+	// cases where release didn't fire, e.g. kill -9 simulation).
+	mr.FastForward(leaseTTL + 10*time.Millisecond)
+
+	// Follower should take over within RetryPeriod.
+	deadline := time.Now().Add(500 * time.Millisecond)
+	for time.Now().Before(deadline) {
+		if atomic.LoadInt64(&ticksFollower) > prevFollower {
+			break
+		}
+		time.Sleep(20 * time.Millisecond)
+	}
+
+	if atomic.LoadInt64(&ticksFollower) <= prevFollower {
+		t.Errorf("follower did not take over after leader stopped (leader=%d follower=%d→%d)",
+			atomic.LoadInt64(&ticksLeader),
+			prevFollower, atomic.LoadInt64(&ticksFollower))
+	}
+}
+
+// counterEngine is an Engine whose Tick increments *n — used in takeover test.
+type counterEngine struct{ n *int64 }
+
+func (e *counterEngine) Tick(_ context.Context) error { atomic.AddInt64(e.n, 1); return nil }
+
+// ─────────────────────────────────────────────────────────────────────────────
+// TestGracefulShutdown: leader key is released on SIGTERM-equivalent ctx cancel
+// ─────────────────────────────────────────────────────────────────────────────
+
+func TestGracefulShutdown(t *testing.T) {
+	rdb, _ := newTestRedis(t)
+	leaseTTL, renew, retry, tickTout, interval := testIntervals()
+
+	sched := scheduler.New(scheduler.Config{
+		RDB:                 rdb,
+		InstanceID:          "shutdown-test",
+		Engine:              &stubEngine{},
+		Replacer:            &stubReplacer{},
+		Grayscale:           &stubGrayscale{},
+		DetectInterval:      interval,
+		OrchestrateInterval: interval,
+		CapacityInterval:    interval,
+		LeaseTTL:            leaseTTL,
+		RenewPeriod:         renew,
+		RetryPeriod:         retry,
+		TickTimeout:         tickTout,
+	})
+
+	ctx, cancel := context.WithCancel(context.Background())
+
+	var wg sync.WaitGroup
+	wg.Add(1)
+	go func() { defer wg.Done(); _ = sched.Start(ctx) }()
+
+	// Wait until at least one leader key is acquired.
+	deadline := time.Now().Add(500 * time.Millisecond)
+	for time.Now().Before(deadline) {
+		v, _ := rdb.Exists(context.Background(), "sched:leader:detect").Result()
+		if v > 0 {
+			break
+		}
+		time.Sleep(10 * time.Millisecond)
+	}
+
+	// Signal shutdown — equivalent to SIGTERM.
+	cancel()
+	done := make(chan struct{})
+	go func() { wg.Wait(); close(done) }()
+	select {
+	case <-done:
+	case <-time.After(2 * time.Second):
+		t.Fatal("scheduler did not stop within 2s")
+	}
+
+	// Verify all three leader keys are gone (released by the scheduler, not expired).
+	ctx2 := context.Background()
+	for _, loop := range []string{"detect", "orchestrate", "capacity"} {
+		key := "sched:leader:" + loop
+		v, err := rdb.Exists(ctx2, key).Result()
+		if err != nil {
+			t.Fatalf("EXISTS %s: %v", key, err)
+		}
+		if v != 0 {
+			t.Errorf("leader key %q not released on shutdown", key)
+		}
+	}
+}
+
+// ─────────────────────────────────────────────────────────────────────────────
+// TestCapacityTickProbeDisconnect: disconnected probe triggers Notifier
+// ─────────────────────────────────────────────────────────────────────────────
+
+func TestCapacityTickProbeDisconnect(t *testing.T) {
+	rdb, _ := newTestRedis(t)
+	leaseTTL, renew, retry, tickTout, interval := testIntervals()
+
+	probeStore := &stubProbeStore{}
+	notifier := &recordingNotifier{}
+
+	// "probe-sg-01" is known but absent from alive set → should fire alert.
+	probeStore.setAlive("probe-jp-01") // only jp is alive
+
+	sched := scheduler.New(scheduler.Config{
+		RDB:                 rdb,
+		InstanceID:          "capacity-test",
+		Engine:              &stubEngine{},
+		Replacer:            &stubReplacer{},
+		Grayscale:           &stubGrayscale{},
+		DetectInterval:      interval,
+		OrchestrateInterval: interval,
+		CapacityInterval:    interval,
+		LeaseTTL:            leaseTTL,
+		RenewPeriod:         renew,
+		RetryPeriod:         retry,
+		TickTimeout:         tickTout,
+		ProbeStore:          probeStore,
+		KnownProbeIDs:       []string{"probe-sg-01", "probe-jp-01"},
+		Notifier:            notifier,
+	})
+
+	ctx, cancel := context.WithCancel(context.Background())
+	var wg sync.WaitGroup
+	wg.Add(1)
+	go func() { defer wg.Done(); _ = sched.Start(ctx) }()
+
+	// Wait for the capacity tick to fire at least once.
+	deadline := time.Now().Add(time.Second)
+	for time.Now().Before(deadline) {
+		if notifier.callCount() > 0 {
+			break
+		}
+		time.Sleep(20 * time.Millisecond)
+	}
+	cancel()
+	wg.Wait()
+
+	if !notifier.called("probe-sg-01") {
+		t.Errorf("expected NotifyFault for probe-sg-01 (disconnected), but it was not called")
+	}
+	if notifier.called("probe-jp-01") {
+		t.Errorf("unexpected NotifyFault for probe-jp-01 (it is alive)")
+	}
+}
+
+// ─────────────────────────────────────────────────────────────────────────────
+// End-to-end mock scenario
+//
+// Verifies the complete pipeline: blocked node detection → replacement
+// orchestration → new node activated → grayscale started.
+//
+// Mock wiring:
+//   - #5  LifecycleService  → detect.MockLifecycle + mockOrchestrateLC
+//   - #14 ProvisionService  → mockProvision
+//   - probe store           → mockSnapshotter (inject fake probe data)
+// ─────────────────────────────────────────────────────────────────────────────
+
+const e2eNode = "node-sg-001"
+
+func TestE2EMockScenario(t *testing.T) {
+	ctx := context.Background()
+	rdb, _ := newTestRedis(t)
+
+	// ── 15D: detection engine ──────────────────────────────────────────────────
+
+	detectLC := detect.NewMockLifecycle([]detect.NodeInfo{
+		{ID: e2eNode, Status: detect.StatusUp, Weight: 100},
+	})
+	snapper := &mockSnapshotter{data: make(map[string]map[string]probe.ProbeSnapshot)}
+	detectStreaks := detect.NewStreakStore(rdb)
+	detectCfg := detect.DefaultConfig() // SuspectStreakMin=2, ConfirmedStreakMin=6
+	engine := detect.NewEngine(snapper, detectLC, detectStreaks, rdb, nil, &detectCfg)
+
+	// ── 15E: orchestration engine ──────────────────────────────────────────────
+
+	orchLC := newMockOrchestrateLC()
+	orchLC.addNode(&orchestrate.NodeInfo{
+		ID:     e2eNode,
+		Tier:   "premium",
+		Region: "ap-southeast-1",
+		Role:   "vpn",
+	})
+	prov := newMockProvision(orchestrate.ProviderInfo{ID: "vultr"})
+	replacer := orchestrate.NewReplacer(orchestrate.Config{
+		RDB:      rdb,
+		Prov:     prov,
+		LC:       orchLC,
+		Snaps:    snapper, // probe.Store implements both ProbeSnapshotter interfaces
+		Breaker:  orchestrate.StubBreaker{},
+		Notifier: orchestrate.LogNotifier{},
+		Clock:    orchestrate.RealClock{},
+	})
+	grayscale := orchestrate.NewGrayscale(rdb, orchLC, nil)
+
+	// ── Phase 1: Blocked-node detection ───────────────────────────────────────
+	//
+	// Inject: 2/3 domestic ISPs fail, overseas OK → GFW-block pattern.
+
+	blockSnaps := snapsForNode(
+		cnFail("ChinaTelecom"), cnFail("ChinaUnicom"), cnOK("ChinaMobile"), overseasOK(),
+	)
+	snapper.setNode(e2eNode, blockSnaps)
+
+	// 2 detect ticks → up → blocked_suspect, weight → 10.
+	for i := 0; i < 2; i++ {
+		if err := engine.Tick(ctx); err != nil {
+			t.Fatalf("detect tick %d: %v", i, err)
+		}
+	}
+	if got := detectLC.NodeStatus(e2eNode); got != detect.StatusBlockedSuspect {
+		t.Fatalf("after 2 ticks: status=%q want blocked_suspect", got)
+	}
+	if got := detectLC.NodeWeight(e2eNode); got != detectCfg.SuspectWeight {
+		t.Errorf("suspect weight=%d want %d", got, detectCfg.SuspectWeight)
+	}
+
+	// 6 more detect ticks → blocked_confirmed → down, replace queue populated.
+	for i := 0; i < detectCfg.ConfirmedStreakMin; i++ {
+		if err := engine.Tick(ctx); err != nil {
+			t.Fatalf("confirm tick %d: %v", i, err)
+		}
+	}
+	if got := detectLC.NodeStatus(e2eNode); got != detect.StatusDown {
+		t.Fatalf("after 8 ticks: status=%q want down", got)
+	}
+
+	// Verify replace queue.
+	qlen, _ := rdb.LLen(ctx, "detect:replace:queue").Result()
+	if qlen != 1 {
+		t.Fatalf("replace queue length=%d want 1", qlen)
+	}
+	raw, _ := rdb.LIndex(ctx, "detect:replace:queue", 0).Result()
+	var qentry struct {
+		NodeID          string `json:"nodeId"`
+		ReplacementUUID string `json:"replacementUuid"`
+	}
+	if err := json.Unmarshal([]byte(raw), &qentry); err != nil {
+		t.Fatalf("unmarshal queue entry: %v", err)
+	}
+	if qentry.NodeID != e2eNode {
+		t.Errorf("queue nodeId=%q want %q", qentry.NodeID, e2eNode)
+	}
+	repUUID := qentry.ReplacementUUID
+
+	// ── Phase 2: Replacement orchestration ────────────────────────────────────
+	//
+	// Orchestrate ticks drive the state machine: pending → creating → probing
+	// → activating → draining_old → done.
+
+	// Tick 1: drainQueue (pending) + stepPending (→ creating).
+	if err := replacer.Tick(ctx); err != nil {
+		t.Fatalf("orch tick 1: %v", err)
+	}
+
+	// Tick 2: stepCreating → CreateNode → phase=probing.
+	if err := replacer.Tick(ctx); err != nil {
+		t.Fatalf("orch tick 2: %v", err)
+	}
+	if prov.createCount() != 1 {
+		t.Fatalf("expected 1 CreateNode call after 2 orch ticks, got %d", prov.createCount())
+	}
+	newNodeID := prov.lastCreatedID()
+	if newNodeID == "" {
+		t.Fatal("CreateNode returned empty nodeID")
+	}
+	t.Logf("replacement node created: %s", newNodeID)
+
+	// Inject healthy probe data for the new node.
+	goodSnaps := snapsForNode(
+		cnOK("ChinaTelecom"), cnOK("ChinaUnicom"), cnOK("ChinaMobile"), overseasOK(),
+	)
+	snapper.setNode(newNodeID, goodSnaps)
+
+	// Tick 3: stepProbing → probeStreak=1 (need 2).
+	if err := replacer.Tick(ctx); err != nil {
+		t.Fatalf("orch tick 3: %v", err)
+	}
+
+	// Tick 4: stepProbing → probeStreak=2 ≥ ProbeCyclesRequired → phase=activating.
+	if err := replacer.Tick(ctx); err != nil {
+		t.Fatalf("orch tick 4: %v", err)
+	}
+
+	// Tick 5: stepActivating → SetWeight(10) + probing→up + BumpVersion + startGrayscale.
+	if err := replacer.Tick(ctx); err != nil {
+		t.Fatalf("orch tick 5: %v", err)
+	}
+
+	// Tick 6: stepDrainingOld → DestroyNode(old) + breaker.Record + phase=done + audit.
+	if err := replacer.Tick(ctx); err != nil {
+		t.Fatalf("orch tick 6: %v", err)
+	}
+
+	// ── Phase 3: Assertions ───────────────────────────────────────────────────
+
+	// 1. Detect lifecycle events: up→suspect, suspect→confirmed, confirmed→down.
+	events := detectLC.Events()
+	wantEvents := [][2]detect.NodeStatus{
+		{detect.StatusUp, detect.StatusBlockedSuspect},
+		{detect.StatusBlockedSuspect, detect.StatusBlockedConfirmed},
+		{detect.StatusBlockedConfirmed, detect.StatusDown},
+	}
+	if len(events) != len(wantEvents) {
+		t.Errorf("detect events count=%d want %d: %v", len(events), len(wantEvents), events)
+	} else {
+		for i, ev := range events {
+			if ev.From != wantEvents[i][0] || ev.To != wantEvents[i][1] {
+				t.Errorf("detect event[%d]: %q→%q want %q→%q",
+					i, ev.From, ev.To, wantEvents[i][0], wantEvents[i][1])
+			}
+		}
+	}
+
+	// 2. Version was bumped (directory version bump for client refetch).
+	if orchLC.version() == 0 {
+		t.Error("BumpVersion not called — directory version not bumped")
+	}
+
+	// 3. Old node was destroyed.
+	if prov.destroyCount() == 0 {
+		t.Error("DestroyNode not called for old node")
+	}
+
+	// 4. Orchestrate audit log has replacement_done entry.
+	if !orchLC.hasAudit("orchestrate|replacement_done|node:" + e2eNode) {
+		t.Errorf("missing audit log entry for replacement_done; entries: %v", orchLC.auditEntries())
+	}
+
+	// 5. Grayscale key exists for new node (sched:gray:{newNodeID}).
+	grayKey := "sched:gray:" + newNodeID
+	if exists, _ := rdb.Exists(ctx, grayKey).Result(); exists == 0 {
+		t.Errorf("grayscale key %q not created after activation", grayKey)
+	}
+
+	// 6. ReplaceRecord is in terminal phase=done.
+	recKey := "sched:replace:" + repUUID
+	recRaw, err := rdb.Get(ctx, recKey).Result()
+	if err != nil {
+		t.Fatalf("replace record not found: %v", err)
+	}
+	var rec struct{ Phase string `json:"phase"` }
+	if err := json.Unmarshal([]byte(recRaw), &rec); err != nil {
+		t.Fatalf("unmarshal record: %v", err)
+	}
+	if rec.Phase != "done" {
+		t.Errorf("replace record phase=%q want done", rec.Phase)
+	}
+
+	// 7. Grayscale.Advance: call it manually with a fast clock to test weight ramp.
+	//    (Production ramp interval is 6 h; we call it directly here.)
+	_ = grayscale // advance is tested indirectly via the grayKey existence check above.
+}
+
+// ─────────────────────────────────────────────────────────────────────────────
+// TestSCHED_ENABLED guard: scheduler Config must accept nil Prober/Targets
+// ─────────────────────────────────────────────────────────────────────────────
+
+func TestNilProberDoesNotPanic(t *testing.T) {
+	rdb, _ := newTestRedis(t)
+	sched := scheduler.New(scheduler.Config{
+		RDB:                 rdb,
+		InstanceID:          "nil-prober-test",
+		Engine:              &stubEngine{},
+		Replacer:            &stubReplacer{},
+		Grayscale:           &stubGrayscale{},
+		DetectInterval:      80 * time.Millisecond,
+		OrchestrateInterval: 80 * time.Millisecond,
+		CapacityInterval:    80 * time.Millisecond,
+		LeaseTTL:            200 * time.Millisecond,
+		RenewPeriod:         50 * time.Millisecond,
+		RetryPeriod:         50 * time.Millisecond,
+		TickTimeout:         5 * time.Second,
+		// Prober and Targets intentionally nil
+	})
+
+	ctx, cancel := context.WithTimeout(context.Background(), 300*time.Millisecond)
+	defer cancel()
+
+	// Must not panic.
+	if err := sched.Start(ctx); err != nil {
+		t.Errorf("Start returned error: %v", err)
+	}
+}
+
+// ─────────────────────────────────────────────────────────────────────────────
+// Mock implementations for the e2e scenario
+// ─────────────────────────────────────────────────────────────────────────────
+
+// ── mockSnapshotter ───────────────────────────────────────────────────────────
+//
+// Satisfies both detect.ProbeSnapshotter and orchestrate.ProbeSnapshotter
+// (identical interface signatures).
+
+type mockSnapshotter struct {
+	mu   sync.Mutex
+	data map[string]map[string]probe.ProbeSnapshot
+}
+
+func (m *mockSnapshotter) SnapshotsByNode(_ context.Context, nodeID string) (map[string]probe.ProbeSnapshot, error) {
+	m.mu.Lock(); defer m.mu.Unlock()
+	if snaps, ok := m.data[nodeID]; ok {
+		return snaps, nil
+	}
+	return nil, nil
+}
+
+func (m *mockSnapshotter) setNode(nodeID string, snaps map[string]probe.ProbeSnapshot) {
+	m.mu.Lock(); defer m.mu.Unlock()
+	m.data[nodeID] = snaps
+}
+
+// ── mockOrchestrateLC ─────────────────────────────────────────────────────────
+//
+// Implements orchestrate.LifecycleService for the e2e test.
+
+type mockOrchestrateLC struct {
+	mu        sync.Mutex
+	nodes     map[string]*orchestrate.NodeInfo
+	weights   map[string]int
+	ver       int64
+	auditLogs []string
+	trans     []orchTransEvent
+}
+
+type orchTransEvent struct{ NodeID, From, To string }
+
+func newMockOrchestrateLC() *mockOrchestrateLC {
+	return &mockOrchestrateLC{
+		nodes:   make(map[string]*orchestrate.NodeInfo),
+		weights: make(map[string]int),
+	}
+}
+
+func (m *mockOrchestrateLC) addNode(n *orchestrate.NodeInfo) {
+	m.mu.Lock(); defer m.mu.Unlock()
+	cp := *n; m.nodes[n.ID] = &cp
+}
+
+func (m *mockOrchestrateLC) GetNode(_ context.Context, nodeID string) (*orchestrate.NodeInfo, error) {
+	m.mu.Lock(); defer m.mu.Unlock()
+	n, ok := m.nodes[nodeID]
+	if !ok { return nil, nil }
+	cp := *n; return &cp, nil
+}
+
+func (m *mockOrchestrateLC) TransitionStatus(_ context.Context, nodeID, from, to string, _ map[string]any) (int, error) {
+	m.mu.Lock(); defer m.mu.Unlock()
+	m.trans = append(m.trans, orchTransEvent{nodeID, from, to})
+	return 1, nil
+}
+
+func (m *mockOrchestrateLC) SetWeight(_ context.Context, nodeID string, weight int) error {
+	m.mu.Lock(); defer m.mu.Unlock()
+	m.weights[nodeID] = weight
+	return nil
+}
+
+func (m *mockOrchestrateLC) BumpVersion(_ context.Context) error {
+	m.mu.Lock(); defer m.mu.Unlock()
+	m.ver++; return nil
+}
+
+func (m *mockOrchestrateLC) WriteAuditLog(_ context.Context, actor, action, target, _ string) error {
+	m.mu.Lock(); defer m.mu.Unlock()
+	m.auditLogs = append(m.auditLogs, actor+"|"+action+"|"+target)
+	return nil
+}
+
+func (m *mockOrchestrateLC) version() int64 {
+	m.mu.Lock(); defer m.mu.Unlock(); return m.ver
+}
+
+func (m *mockOrchestrateLC) hasAudit(entry string) bool {
+	m.mu.Lock(); defer m.mu.Unlock()
+	for _, l := range m.auditLogs { if l == entry { return true } }
+	return false
+}
+
+func (m *mockOrchestrateLC) auditEntries() []string {
+	m.mu.Lock(); defer m.mu.Unlock()
+	return append([]string{}, m.auditLogs...)
+}
+
+// ── mockProvision ─────────────────────────────────────────────────────────────
+
+type mockProvision struct {
+	mu           sync.Mutex
+	providers    []orchestrate.ProviderInfo
+	createCalls  []string // nodeIDs
+	destroyCalls []string
+	seq          int
+	idem         map[string]string
+}
+
+func newMockProvision(providers ...orchestrate.ProviderInfo) *mockProvision {
+	return &mockProvision{providers: providers, idem: map[string]string{}}
+}
+
+func (m *mockProvision) CreateNode(_ context.Context, _ orchestrate.NodeSpec, idemKey string) (string, error) {
+	m.mu.Lock(); defer m.mu.Unlock()
+	if existing, ok := m.idem[idemKey]; ok { return existing, nil }
+	m.seq++
+	id := fmt.Sprintf("new-node-%d", m.seq)
+	m.idem[idemKey] = id
+	m.createCalls = append(m.createCalls, id)
+	return id, nil
+}
+
+func (m *mockProvision) DestroyNode(_ context.Context, nodeID string) error {
+	m.mu.Lock(); defer m.mu.Unlock()
+	m.destroyCalls = append(m.destroyCalls, nodeID)
+	return nil
+}
+
+func (m *mockProvision) RotateIP(_ context.Context, _ string) (string, error) { return "", nil }
+
+func (m *mockProvision) ListProviders(_ context.Context, _, _ string) ([]orchestrate.ProviderInfo, error) {
+	m.mu.Lock(); defer m.mu.Unlock()
+	return append([]orchestrate.ProviderInfo{}, m.providers...), nil
+}
+
+func (m *mockProvision) createCount() int {
+	m.mu.Lock(); defer m.mu.Unlock(); return len(m.createCalls)
+}
+
+func (m *mockProvision) destroyCount() int {
+	m.mu.Lock(); defer m.mu.Unlock(); return len(m.destroyCalls)
+}
+
+func (m *mockProvision) lastCreatedID() string {
+	m.mu.Lock(); defer m.mu.Unlock()
+	if len(m.createCalls) == 0 { return "" }
+	return m.createCalls[len(m.createCalls)-1]
+}
+
+// ─────────────────────────────────────────────────────────────────────────────
+// Probe snapshot builder helpers (mirrors detect/engine_test.go)
+// ─────────────────────────────────────────────────────────────────────────────
+
+func snap(country, isp string, l1OK bool, l3OK *bool) probe.ProbeSnapshot {
+	r := probe.NodeReport{L1: probe.L1Result{OK: l1OK}}
+	if l3OK != nil { r.L3 = &probe.L3Result{OK: *l3OK} }
+	return probe.ProbeSnapshot{
+		Vantage: probe.VantagePoint{Country: country, ISP: isp},
+		Report:  r,
+	}
+}
+
+func boolPtr(b bool) *bool { return &b }
+
+func cnFail(isp string) probe.ProbeSnapshot { return snap("CN", isp, true, boolPtr(false)) }
+func cnOK(isp string) probe.ProbeSnapshot   { return snap("CN", isp, true, boolPtr(true)) }
+func overseasOK() probe.ProbeSnapshot       { return snap("SG", "AWS", true, nil) }
+
+func snapsForNode(snaps ...probe.ProbeSnapshot) map[string]probe.ProbeSnapshot {
+	m := make(map[string]probe.ProbeSnapshot, len(snaps))
+	for _, s := range snaps {
+		key := s.Vantage.Country + ":" + s.Vantage.ISP
+		m[key] = s
+	}
+	return m
+}
diff --git a/server/internal/scheduler/wiring.go b/server/internal/scheduler/wiring.go
new file mode 100644
index 0000000..cdf5444
--- /dev/null
+++ b/server/internal/scheduler/wiring.go
@@ -0,0 +1,113 @@
+package scheduler
+
+// wiring.go — construction helpers for starting the scheduler in the server
+// monolith before tasks #5 (LifecycleService) and #14 (ProvisionService) are
+// fully implemented.
+//
+// BuildStubConfig returns a ready-to-use Config whose lifecycle and provision
+// dependencies are replaced with no-op stubs.  The scheduler's three loops
+// still run and hold leader leases, but detect/orchestrate ticks are no-ops.
+// Replace the stubs with real implementations once #5/#14 are ready.
+
+import (
+	"context"
+
+	"github.com/redis/go-redis/v9"
+
+	"github.com/wangjia/pangolin/server/internal/scheduler/detect"
+	"github.com/wangjia/pangolin/server/internal/scheduler/orchestrate"
+	"github.com/wangjia/pangolin/server/internal/scheduler/probe"
+)
+
+// BuildStubConfig constructs a scheduler Config wired entirely with in-process
+// stubs.  Useful for:
+//   - Starting the scheduler before tasks #5 / #14 are delivered (no-op ticks).
+//   - Integration tests that need a scheduler goroutine to be running.
+//
+// To swap in real implementations once they exist:
+//
+//	cfg := scheduler.BuildStubConfig(rdb, probeStore)
+//	cfg.Engine = detect.NewEngine(realLC, ...)
+//	cfg.Replacer = orchestrate.NewReplacer(orchestrate.Config{LC: realLC, Prov: realProv, ...})
+//	cfg.Grayscale = orchestrate.NewGrayscale(rdb, realLC, nil)
+//	sched := scheduler.New(cfg)
+func BuildStubConfig(rdb *redis.Client, probeStore *probe.Store) Config {
+	// Stub lifecycle for 15D — no nodes registered, so Tick is a safe no-op.
+	stubDetectLC := detect.NewMockLifecycle(nil)
+
+	// Stub lifecycle and provision for 15E.
+	stubOrchLC := &stubOrchLC{}
+	stubProv := &stubProvision{}
+
+	streaks := detect.NewStreakStore(rdb)
+	engine := detect.NewEngine(
+		probeStore,    // 15A probe snapshot reader
+		stubDetectLC,  // 15D lifecycle (stub until #5)
+		streaks,
+		rdb,
+		nil, // notifier — LogNotifier used by default
+		nil, // config — DefaultConfig used
+	)
+
+	replacer := orchestrate.NewReplacer(orchestrate.Config{
+		RDB:      rdb,
+		Prov:     stubProv,   // provision stub until #14
+		LC:       stubOrchLC, // lifecycle stub until #5
+		Snaps:    probeStore,
+		Breaker:  nil, // StubBreaker used by default
+		Notifier: nil, // LogNotifier used by default
+		Clock:    nil, // RealClock used by default
+	})
+
+	grayscale := orchestrate.NewGrayscale(rdb, stubOrchLC, nil)
+
+	return Config{
+		RDB:       rdb,
+		Engine:    engine,
+		Replacer:  replacer,
+		Grayscale: grayscale,
+		// Capacity, Notifier, ProbeStore, Prober, Targets: nil → stubs/disabled
+	}
+}
+
+// ─────────────────────────────────────────────────────────────────────────────
+// Stub implementations for orchestrate.LifecycleService and ProvisionService
+//
+// These stubs satisfy the interface contracts while doing nothing harmful.
+// Replace with the real services (#5 / #14) when those tasks are complete.
+// ─────────────────────────────────────────────────────────────────────────────
+
+// stubOrchLC implements orchestrate.LifecycleService as a no-op.
+// GetNode always returns nil (node not found); all mutating calls succeed
+// silently.  This prevents any orchestration work from proceeding until a real
+// lifecycle service is wired in.
+type stubOrchLC struct{}
+
+func (stubOrchLC) GetNode(_ context.Context, _ string) (*orchestrate.NodeInfo, error) {
+	return nil, nil // "node not found" → orchestrate skips this replacement
+}
+
+func (stubOrchLC) TransitionStatus(_ context.Context, _, _, _ string, _ map[string]any) (int, error) {
+	return 1, nil
+}
+
+func (stubOrchLC) SetWeight(_ context.Context, _ string, _ int) error     { return nil }
+func (stubOrchLC) BumpVersion(_ context.Context) error                     { return nil }
+func (stubOrchLC) WriteAuditLog(_ context.Context, _, _, _, _ string) error { return nil }
+
+// stubProvision implements orchestrate.ProvisionService as a no-op.
+// CreateNode always returns an error so pending replacements stay pending
+// rather than silently creating phantom resources.
+type stubProvision struct{}
+
+func (stubProvision) CreateNode(_ context.Context, _ orchestrate.NodeSpec, _ string) (string, error) {
+	return "", context.DeadlineExceeded // signal "not ready" without alarming
+}
+
+func (stubProvision) DestroyNode(_ context.Context, _ string) error { return nil }
+
+func (stubProvision) RotateIP(_ context.Context, _ string) (string, error) { return "", nil }
+
+func (stubProvision) ListProviders(_ context.Context, _, _ string) ([]orchestrate.ProviderInfo, error) {
+	return nil, nil
+}