fix(scheduler): resolve SetSnapshot race conditions and remove usage throttle

Backend: Fix three race conditions in SetSnapshot that caused account
scheduling anomalies and broken sticky sessions:
- Use Lua CAS script for atomic version activation, preventing version
  rollback when concurrent goroutines write snapshots simultaneously
- Add UnlockBucket to release rebuild lock immediately after completion
  instead of waiting 30s TTL expiry
- Replace immediate DEL of old snapshots with 60s EXPIRE grace period,
  preventing readers from hitting empty ZRANGE during version switches

Frontend: Remove serial queue throttle (1-2s delay per request) from
usage loading since backend now uses passive sampling. All usage
requests execute immediately in parallel.

backend/internal/repository/scheduler_cache.go

@@ -24,6 +24,49 @@ const (
 
 	defaultSchedulerSnapshotMGetChunkSize  = 128
 	defaultSchedulerSnapshotWriteChunkSize = 256
+
+	// snapshotGraceTTLSeconds 旧快照过期的宽限期（秒）。
+	// 替代立即 DEL，让正在读取旧版本的 reader 有足够时间完成 ZRANGE。
+	snapshotGraceTTLSeconds = 60
+)
+
+var (
+	// activateSnapshotScript 原子 CAS 切换快照版本。
+	// 仅当新版本号 >= 当前激活版本时才切换，防止并发写入导致版本回滚。
+	// 旧快照使用 EXPIRE 设置宽限期而非立即 DEL，避免与 reader 竞态。
+	//
+	// KEYS[1] = activeKey     (sched:active:{bucket})
+	// KEYS[2] = readyKey      (sched:ready:{bucket})
+	// KEYS[3] = bucketSetKey  (sched:buckets)
+	// KEYS[4] = snapshotKey   (新写入的快照 key)
+	// ARGV[1] = 新版本号字符串
+	// ARGV[2] = bucket 字符串 (用于 SADD)
+	// ARGV[3] = 快照 key 前缀 (用于构造旧快照 key)
+	// ARGV[4] = 宽限期 TTL 秒数
+	//
+	// 返回 1 = 已激活, 0 = 版本过旧未激活
+	activateSnapshotScript = redis.NewScript(`
+local currentActive = redis.call('GET', KEYS[1])
+local newVersion = tonumber(ARGV[1])
+
+if currentActive ~= false then
+	local curVersion = tonumber(currentActive)
+	if curVersion and newVersion < curVersion then
+		redis.call('DEL', KEYS[4])
+		return 0
+	end
+end
+
+redis.call('SET', KEYS[1], ARGV[1])
+redis.call('SET', KEYS[2], '1')
+redis.call('SADD', KEYS[3], ARGV[2])
+
+if currentActive ~= false and currentActive ~= ARGV[1] then
+	redis.call('EXPIRE', ARGV[3] .. currentActive, tonumber(ARGV[4]))
+end
+
+return 1
+`)
 )
 
 type schedulerCache struct {
@@ -108,9 +151,9 @@ func (c *schedulerCache) GetSnapshot(ctx context.Context, bucket service.Schedul
 }
 
 func (c *schedulerCache) SetSnapshot(ctx context.Context, bucket service.SchedulerBucket, accounts []service.Account) error {
-	activeKey := schedulerBucketKey(schedulerActivePrefix, bucket)
-	oldActive, _ := c.rdb.Get(ctx, activeKey).Result()
-
+	// Phase 1: 分配新版本号并写入快照数据。
+	// INCR 保证每个调用方获得唯一递增版本号。
+	// 写入的 snapshotKey 是新的版本化 key，reader 尚不知晓，因此无竞态。
 	versionKey := schedulerBucketKey(schedulerVersionPrefix, bucket)
 	version, err := c.rdb.Incr(ctx, versionKey).Result()
 	if err != nil {
@@ -124,7 +167,6 @@ func (c *schedulerCache) SetSnapshot(ctx context.Context, bucket service.Schedul
 		return err
 	}
 
-	pipe := c.rdb.Pipeline()
 	if len(accounts) > 0 {
 		// 使用序号作为 score，保持数据库返回的排序语义。
 		members := make([]redis.Z, 0, len(accounts))
@@ -134,6 +176,7 @@ func (c *schedulerCache) SetSnapshot(ctx context.Context, bucket service.Schedul
 				Member: strconv.FormatInt(account.ID, 10),
 			})
 		}
+		pipe := c.rdb.Pipeline()
 		for start := 0; start < len(members); start += c.writeChunkSize {
 			end := start + c.writeChunkSize
 			if end > len(members) {
@@ -141,18 +184,25 @@ func (c *schedulerCache) SetSnapshot(ctx context.Context, bucket service.Schedul
 			}
 			pipe.ZAdd(ctx, snapshotKey, members[start:end]...)
 		}
-	} else {
-		pipe.Del(ctx, snapshotKey)
-	}
-	pipe.Set(ctx, activeKey, versionStr, 0)
-	pipe.Set(ctx, schedulerBucketKey(schedulerReadyPrefix, bucket), "1", 0)
-	pipe.SAdd(ctx, schedulerBucketSetKey, bucket.String())
-	if _, err := pipe.Exec(ctx); err != nil {
-		return err
+		if _, err := pipe.Exec(ctx); err != nil {
+			return err
+		}
 	}
 
-	if oldActive != "" && oldActive != versionStr {
-		_ = c.rdb.Del(ctx, schedulerSnapshotKey(bucket, oldActive)).Err()
+	// Phase 2: 原子 CAS 激活版本。
+	// Lua 脚本保证：仅当新版本 >= 当前激活版本时才切换 active 指针，
+	// 防止并发写入导致版本回滚。
+	// 旧快照使用 EXPIRE 宽限期而非立即 DEL，避免 reader 竞态。
+	activeKey := schedulerBucketKey(schedulerActivePrefix, bucket)
+	readyKey := schedulerBucketKey(schedulerReadyPrefix, bucket)
+	snapshotKeyPrefix := fmt.Sprintf("%s%d:%s:%s:v", schedulerSnapshotPrefix, bucket.GroupID, bucket.Platform, bucket.Mode)
+
+	keys := []string{activeKey, readyKey, schedulerBucketSetKey, snapshotKey}
+	args := []any{versionStr, bucket.String(), snapshotKeyPrefix, snapshotGraceTTLSeconds}
+
+	_, err = activateSnapshotScript.Run(ctx, c.rdb, keys, args...).Result()
+	if err != nil {
+		return err
 	}
 
 	return nil
@@ -232,6 +282,11 @@ func (c *schedulerCache) TryLockBucket(ctx context.Context, bucket service.Sched
 	return c.rdb.SetNX(ctx, key, time.Now().UnixNano(), ttl).Result()
 }
 
+func (c *schedulerCache) UnlockBucket(ctx context.Context, bucket service.SchedulerBucket) error {
+	key := schedulerBucketKey(schedulerLockPrefix, bucket)
+	return c.rdb.Del(ctx, key).Err()
+}
+
 func (c *schedulerCache) ListBuckets(ctx context.Context) ([]service.SchedulerBucket, error) {
 	raw, err := c.rdb.SMembers(ctx, schedulerBucketSetKey).Result()
 	if err != nil {