perf(worker): 按实例内存上限扩展异步容量

将策略推导出的集群目标与单 Worker 内存安全上限分离，避免失活实例的全部容量转移到单个 2 GiB Pod。

生产环境单实例上限和媒体物化并发设为 24；两实例正常时总容量由 16 提升至 48，后续可通过增加 Worker 节点继续扩展。新增 0093 迁移保存实例容量上限，并兼容滚动发布中的旧实例。

风险：更高媒体并发会增加 Worker 和节点内存压力，保留单实例 24 的硬边界并由持续监控观察 RSS、OOM、队列与节点余量。

验证：go test ./... -count=1；go vet ./...；真实 PostgreSQL 分配与故障转移测试；迁移安全检查；kubectl kustomize；gofmt -l；git diff --check。

README.md

@@ -162,7 +162,7 @@ AI_GATEWAY_DATABASE_URL=postgresql://easyai:easyai2025@localhost:5432/easyai_ai_
 
 如果现有 `easyai-pgvector` 没有把 `5432` 映射到宿主机，就需要补端口映射，或者把 AI Gateway 后端容器化后接入同一个 `easyai` Docker network。
 
-异步队列 worker 不使用固定并发。服务分别汇总启用平台模型和活跃用户组的有效 `concurrent` 策略，采用两者中更严格的容量，默认每 5 秒在线调整 River 执行容量。策略解析兼容历史 `platformLimits/modelLimits.max_concurrent_requests`，运行时统一转换为 `rules`。`AI_GATEWAY_ASYNC_WORKER_HARD_LIMIT` 默认 `2048`，仅作为单进程资源安全边界；平台模型和用户组的 PostgreSQL concurrency lease 才是业务并发真值。可通过 `AI_GATEWAY_ASYNC_WORKER_REFRESH_INTERVAL_SECONDS` 调整刷新周期。
+异步队列 worker 不使用固定业务并发。服务分别汇总启用平台模型和活跃用户组的有效 `concurrent` 策略，采用两者中更严格的集群容量，默认每 5 秒在线调整 River 执行容量。策略解析兼容历史 `platformLimits/modelLimits.max_concurrent_requests`，运行时统一转换为 `rules`。`AI_GATEWAY_ASYNC_WORKER_HARD_LIMIT` 默认 `2048`，限制策略推导出的集群目标；`AI_GATEWAY_ASYNC_WORKER_INSTANCE_HARD_LIMIT` 默认 `32`，按单 Worker 的内存安全容量限制实例分配，即使其他实例失活也不会突破。平台模型和用户组的 PostgreSQL concurrency lease 仍是业务并发真值。可通过 `AI_GATEWAY_ASYNC_WORKER_REFRESH_INTERVAL_SECONDS` 调整刷新周期。
 
 ## 迁移原则
 

apps/api/internal/config/config.go

@@ -63,6 +63,7 @@ type Config struct {
 	MediaMaterializationConcurrency               int
 	AsyncQueueWorkerEnabled                       bool
 	AsyncWorkerHardLimit                          int
+	AsyncWorkerInstanceHardLimit                  int
 	AsyncWorkerRefreshIntervalSeconds             int
 }
 
@@ -123,6 +124,7 @@ func Load() Config {
 		MediaMaterializationConcurrency:   envIntValidated("AI_GATEWAY_MEDIA_MATERIALIZATION_CONCURRENCY", 8),
 		AsyncQueueWorkerEnabled:           env("AI_GATEWAY_ASYNC_QUEUE_WORKER_ENABLED", "true") == "true",
 		AsyncWorkerHardLimit:              envIntValidated("AI_GATEWAY_ASYNC_WORKER_HARD_LIMIT", 2048),
+		AsyncWorkerInstanceHardLimit:      envIntValidated("AI_GATEWAY_ASYNC_WORKER_INSTANCE_HARD_LIMIT", 32),
 		AsyncWorkerRefreshIntervalSeconds: envIntValidated("AI_GATEWAY_ASYNC_WORKER_REFRESH_INTERVAL_SECONDS", 5),
 	}
 }
@@ -150,6 +152,9 @@ func (c Config) Validate() error {
 	if c.AsyncWorkerHardLimit < 1 || c.AsyncWorkerHardLimit > 10000 {
 		return errors.New("AI_GATEWAY_ASYNC_WORKER_HARD_LIMIT must be between 1 and 10000")
 	}
+	if c.AsyncWorkerInstanceHardLimit < 1 || c.AsyncWorkerInstanceHardLimit > 10000 {
+		return errors.New("AI_GATEWAY_ASYNC_WORKER_INSTANCE_HARD_LIMIT must be between 1 and 10000")
+	}
 	if c.AsyncWorkerRefreshIntervalSeconds < 1 {
 		return errors.New("AI_GATEWAY_ASYNC_WORKER_REFRESH_INTERVAL_SECONDS must be positive")
 	}

apps/api/internal/config/config_test.go

@@ -26,7 +26,12 @@ func TestLoadIdentitySecretStoreUsesNewEnvironmentNamesAndIgnoresLegacyBusinessV
 }
 
 func TestValidateIdentityFileSecretStoreRequiresDirectory(t *testing.T) {
-	cfg := Config{IdentitySecretStore: "file", AsyncWorkerHardLimit: 2048, AsyncWorkerRefreshIntervalSeconds: 5}
+	cfg := Config{
+		IdentitySecretStore:               "file",
+		AsyncWorkerHardLimit:              2048,
+		AsyncWorkerInstanceHardLimit:      32,
+		AsyncWorkerRefreshIntervalSeconds: 5,
+	}
 	if err := cfg.Validate(); err == nil || !strings.Contains(err.Error(), "IDENTITY_SECRET_DIR") {
 		t.Fatalf("Validate() error = %v, want missing identity secret directory", err)
 	}
@@ -41,6 +46,7 @@ func TestValidateIdentityKubernetesSecretStore(t *testing.T) {
 		IdentitySecretStore:               "kubernetes",
 		IdentityKubernetesSecretName:      "easyai-gateway-identity",
 		AsyncWorkerHardLimit:              2048,
+		AsyncWorkerInstanceHardLimit:      32,
 		AsyncWorkerRefreshIntervalSeconds: 5,
 	}
 	if err := cfg.Validate(); err == nil || !strings.Contains(err.Error(), "namespace") {
@@ -58,6 +64,7 @@ func TestValidateIdentitySecurityEventTiming(t *testing.T) {
 		IdentitySecurityEventStaleAfterSeconds:        60,
 		IdentitySecurityEventClockSkewSeconds:         60,
 		AsyncWorkerHardLimit:                          2048,
+		AsyncWorkerInstanceHardLimit:                  32,
 		AsyncWorkerRefreshIntervalSeconds:             5,
 	}
 	if err := cfg.Validate(); err == nil || !strings.Contains(err.Error(), "heartbeat") {
@@ -66,11 +73,20 @@ func TestValidateIdentitySecurityEventTiming(t *testing.T) {
 }
 
 func TestValidateAsyncWorkerSettings(t *testing.T) {
-	cfg := Config{AsyncWorkerHardLimit: 10001, AsyncWorkerRefreshIntervalSeconds: 5}
+	cfg := Config{
+		AsyncWorkerHardLimit:              10001,
+		AsyncWorkerInstanceHardLimit:      32,
+		AsyncWorkerRefreshIntervalSeconds: 5,
+	}
 	if err := cfg.Validate(); err == nil || !strings.Contains(err.Error(), "HARD_LIMIT") {
 		t.Fatalf("Validate() error = %v, want invalid hard limit", err)
 	}
 	cfg.AsyncWorkerHardLimit = 2048
+	cfg.AsyncWorkerInstanceHardLimit = 10001
+	if err := cfg.Validate(); err == nil || !strings.Contains(err.Error(), "INSTANCE_HARD_LIMIT") {
+		t.Fatalf("Validate() error = %v, want invalid instance hard limit", err)
+	}
+	cfg.AsyncWorkerInstanceHardLimit = 32
 	cfg.AsyncWorkerRefreshIntervalSeconds = 0
 	if err := cfg.Validate(); err == nil || !strings.Contains(err.Error(), "REFRESH_INTERVAL") {
 		t.Fatalf("Validate() error = %v, want invalid refresh interval", err)
@@ -80,6 +96,12 @@ func TestValidateAsyncWorkerSettings(t *testing.T) {
 	if err := loaded.Validate(); err == nil || !strings.Contains(err.Error(), "HARD_LIMIT") {
 		t.Fatalf("Validate() error = %v, want invalid non-integer hard limit", err)
 	}
+	t.Setenv("AI_GATEWAY_ASYNC_WORKER_HARD_LIMIT", "2048")
+	t.Setenv("AI_GATEWAY_ASYNC_WORKER_INSTANCE_HARD_LIMIT", "not-an-integer")
+	loaded = Load()
+	if err := loaded.Validate(); err == nil || !strings.Contains(err.Error(), "INSTANCE_HARD_LIMIT") {
+		t.Fatalf("Validate() error = %v, want invalid non-integer instance hard limit", err)
+	}
 }
 
 func TestLoadAsyncQueueWorkerEnabled(t *testing.T) {

apps/api/internal/runner/queue_worker.go

@@ -220,6 +220,10 @@ func (s *Service) startRiverQueue(ctx context.Context, workerEnabled bool) error
 	s.logger.Info("async worker capacity initialized",
 		"capacity", snapshot.Capacity,
 		"desiredCapacity", snapshot.Desired,
+		"activeInstances", snapshot.ActiveInstances,
+		"globalCapacity", snapshot.GlobalCapacity,
+		"instanceID", snapshot.InstanceID,
+		"instanceHardLimit", s.cfg.AsyncWorkerInstanceHardLimit,
 		"hardLimit", snapshot.HardLimit,
 		"enabledModels", snapshot.EnabledModels,
 		"unlimitedModels", snapshot.UnlimitedModels,
@@ -289,13 +293,14 @@ func (s *Service) loadAsyncWorkerCapacity(ctx context.Context) (store.AsyncWorke
 		Site:                strings.TrimSpace(os.Getenv("EASYAI_SITE")),
 		Revision:            strings.TrimSpace(os.Getenv("AI_GATEWAY_REVISION")),
 		DesiredCapacity:     snapshot.Capacity,
+		CapacityLimit:       s.cfg.AsyncWorkerInstanceHardLimit,
 		HeartbeatStaleAfter: time.Duration(s.cfg.AsyncWorkerRefreshIntervalSeconds) * 6 * time.Second,
 	})
 	if err != nil {
 		return store.AsyncWorkerCapacitySnapshot{}, err
 	}
 	snapshot.Capacity = allocation.Allocated
-	snapshot.GlobalCapacity = allocation.DesiredCapacity
+	snapshot.GlobalCapacity = allocation.GlobalAllocated
 	snapshot.ActiveInstances = allocation.ActiveInstances
 	snapshot.InstanceID = allocation.InstanceID
 	return snapshot, nil
@@ -372,7 +377,9 @@ func (s *Service) resizeAsyncWorkerCapacity(ctx context.Context) {
 		"capacity", snapshot.Capacity,
 		"desiredCapacity", snapshot.Desired,
 		"activeInstances", snapshot.ActiveInstances,
+		"globalCapacity", snapshot.GlobalCapacity,
 		"instanceID", snapshot.InstanceID,
+		"instanceHardLimit", s.cfg.AsyncWorkerInstanceHardLimit,
 		"hardLimit", snapshot.HardLimit,
 		"modelDesired", snapshot.ModelDesired,
 		"groupDesired", snapshot.GroupDesired,

apps/api/internal/runner/service.go

@@ -93,6 +93,9 @@ func New(cfg config.Config, db *store.Store, logger *slog.Logger, observers ...b
 	if cfg.AsyncWorkerHardLimit == 0 {
 		cfg.AsyncWorkerHardLimit = 2048
 	}
+	if cfg.AsyncWorkerInstanceHardLimit == 0 {
+		cfg.AsyncWorkerInstanceHardLimit = 32
+	}
 	if cfg.AsyncWorkerRefreshIntervalSeconds == 0 {
 		cfg.AsyncWorkerRefreshIntervalSeconds = 5
 	}

apps/api/internal/store/admission_queue_integration_test.go

@@ -625,4 +625,17 @@ WHERE instance_id = $1`, secondID, (workerHeartbeatStaleAfter + time.Second).Str
 	if err != nil || first.Allocated != 5 || first.ActiveInstances != 1 {
 		t.Fatalf("single-worker failover allocation = %+v, err=%v", first, err)
 	}
+
+	first, err = db.RegisterWorkerInstance(ctx, WorkerRegistrationInput{
+		InstanceID: firstID, DesiredCapacity: 100, CapacityLimit: 3,
+	})
+	if err != nil || first.Allocated != 3 || first.GlobalAllocated != 3 {
+		t.Fatalf("bounded single-worker allocation = %+v, err=%v", first, err)
+	}
+	second, err = db.RegisterWorkerInstance(ctx, WorkerRegistrationInput{
+		InstanceID: secondID, DesiredCapacity: 100, CapacityLimit: 2,
+	})
+	if err != nil || second.Allocated != 2 || second.GlobalAllocated != 5 || second.ActiveInstances != 2 {
+		t.Fatalf("bounded two-worker allocation = %+v, err=%v", second, err)
+	}
 }

apps/api/internal/store/worker_registry.go

@@ -18,6 +18,7 @@ type WorkerRegistrationInput struct {
 	Site                string
 	Revision            string
 	DesiredCapacity     int
+	CapacityLimit       int
 	HeartbeatStaleAfter time.Duration
 }
 
@@ -25,10 +26,16 @@ type WorkerAllocation struct {
 	InstanceID      string
 	DesiredCapacity int
 	Allocated       int
+	GlobalAllocated int
 	ActiveInstances int
 	HeartbeatAt     time.Time
 }
 
+type activeWorkerCapacity struct {
+	InstanceID    string
+	CapacityLimit int
+}
+
 // ActiveWorkerCapacity returns the cluster-wide execution capacity currently
 // owned by live Worker instances.
 func (s *Store) ActiveWorkerCapacity(ctx context.Context) (int, error) {
@@ -49,6 +56,12 @@ func (s *Store) RegisterWorkerInstance(ctx context.Context, input WorkerRegistra
 	if input.DesiredCapacity < 0 {
 		return WorkerAllocation{}, errors.New("worker desired capacity cannot be negative")
 	}
+	if input.CapacityLimit < 0 {
+		return WorkerAllocation{}, errors.New("worker capacity limit cannot be negative")
+	}
+	if input.CapacityLimit == 0 {
+		input.CapacityLimit = input.DesiredCapacity
+	}
 	staleAfter := input.HeartbeatStaleAfter
 	if staleAfter < workerHeartbeatStaleAfter {
 		staleAfter = workerHeartbeatStaleAfter
@@ -70,9 +83,9 @@ func (s *Store) RegisterWorkerInstance(ctx context.Context, input WorkerRegistra
 	if _, err := tx.Exec(ctx, `
 INSERT INTO gateway_worker_instances (
   instance_id, pod_uid, pod_name, site, revision, status,
-  desired_capacity, allocated_capacity, started_at, heartbeat_at, updated_at
+  desired_capacity, capacity_limit, allocated_capacity, started_at, heartbeat_at, updated_at
 )
-VALUES ($1, $2, $3, $4, $5, 'active', $6, 0, now(), now(), now())
+VALUES ($1, $2, $3, $4, $5, 'active', $6, $7, 0, now(), now(), now())
 ON CONFLICT (instance_id) DO UPDATE
 SET pod_uid = EXCLUDED.pod_uid,
     pod_name = EXCLUDED.pod_name,
@@ -80,6 +93,7 @@ SET pod_uid = EXCLUDED.pod_uid,
     revision = EXCLUDED.revision,
     status = 'active',
     desired_capacity = EXCLUDED.desired_capacity,
+    capacity_limit = EXCLUDED.capacity_limit,
     heartbeat_at = now(),
     updated_at = now()`,
 		input.InstanceID,
@@ -88,6 +102,7 @@ SET pod_uid = EXCLUDED.pod_uid,
 		strings.TrimSpace(input.Site),
 		strings.TrimSpace(input.Revision),
 		input.DesiredCapacity,
+		input.CapacityLimit,
 	); err != nil {
 		return WorkerAllocation{}, err
 	}
@@ -101,7 +116,7 @@ WHERE status <> 'active'
 	}
 
 	rows, err := tx.Query(ctx, `
-SELECT instance_id
+SELECT instance_id, capacity_limit
 FROM gateway_worker_instances
 WHERE status = 'active'
   AND heartbeat_at > now() - $1::interval
@@ -110,41 +125,37 @@ FOR UPDATE`, staleAfter.String())
 	if err != nil {
 		return WorkerAllocation{}, err
 	}
-	activeIDs := make([]string, 0)
+	activeWorkers := make([]activeWorkerCapacity, 0)
 	for rows.Next() {
-		var instanceID string
-		if err := rows.Scan(&instanceID); err != nil {
+		var worker activeWorkerCapacity
+		if err := rows.Scan(&worker.InstanceID, &worker.CapacityLimit); err != nil {
 			rows.Close()
 			return WorkerAllocation{}, err
 		}
-		activeIDs = append(activeIDs, instanceID)
+		activeWorkers = append(activeWorkers, worker)
 	}
 	if err := rows.Err(); err != nil {
 		rows.Close()
 		return WorkerAllocation{}, err
 	}
 	rows.Close()
-	if len(activeIDs) == 0 {
+	if len(activeWorkers) == 0 {
 		return WorkerAllocation{}, errors.New("worker registration was not active after heartbeat")
 	}
 
-	base := input.DesiredCapacity / len(activeIDs)
-	remainder := input.DesiredCapacity % len(activeIDs)
+	allocations, globalAllocated := allocateWorkerCapacities(activeWorkers, input.DesiredCapacity)
 	allocated := 0
-	for index, instanceID := range activeIDs {
-		capacity := base
-		if index < remainder {
-			capacity++
-		}
+	for _, worker := range activeWorkers {
+		capacity := allocations[worker.InstanceID]
 		if _, err := tx.Exec(ctx, `
 UPDATE gateway_worker_instances
 SET desired_capacity = $2,
     allocated_capacity = $3,
     updated_at = now()
-WHERE instance_id = $1`, instanceID, input.DesiredCapacity, capacity); err != nil {
+WHERE instance_id = $1`, worker.InstanceID, input.DesiredCapacity, capacity); err != nil {
 			return WorkerAllocation{}, err
 		}
-		if instanceID == input.InstanceID {
+		if worker.InstanceID == input.InstanceID {
 			allocated = capacity
 		}
 	}
@@ -162,11 +173,42 @@ WHERE instance_id = $1`, input.InstanceID).Scan(&heartbeatAt); err != nil {
 		InstanceID:      input.InstanceID,
 		DesiredCapacity: input.DesiredCapacity,
 		Allocated:       allocated,
-		ActiveInstances: len(activeIDs),
+		GlobalAllocated: globalAllocated,
+		ActiveInstances: len(activeWorkers),
 		HeartbeatAt:     heartbeatAt,
 	}, nil
 }
 
+func allocateWorkerCapacities(workers []activeWorkerCapacity, desired int) (map[string]int, int) {
+	allocations := make(map[string]int, len(workers))
+	if desired <= 0 || len(workers) == 0 {
+		return allocations, 0
+	}
+	remaining := desired
+	for remaining > 0 {
+		progressed := false
+		for _, worker := range workers {
+			limit := worker.CapacityLimit
+			if limit <= 0 {
+				limit = desired
+			}
+			if allocations[worker.InstanceID] >= limit {
+				continue
+			}
+			allocations[worker.InstanceID]++
+			remaining--
+			progressed = true
+			if remaining == 0 {
+				break
+			}
+		}
+		if !progressed {
+			break
+		}
+	}
+	return allocations, desired - remaining
+}
+
 func (s *Store) MarkWorkerDraining(ctx context.Context, instanceID string) error {
 	result, err := s.pool.Exec(ctx, `
 UPDATE gateway_worker_instances

apps/api/internal/store/worker_registry_test.go

@@ -0,0 +1,44 @@
+package store
+
+import "testing"
+
+func TestAllocateWorkerCapacitiesHonorsInstanceLimits(t *testing.T) {
+	workers := []activeWorkerCapacity{
+		{InstanceID: "worker-a", CapacityLimit: 24},
+		{InstanceID: "worker-b", CapacityLimit: 24},
+	}
+	allocations, global := allocateWorkerCapacities(workers, 310)
+	if global != 48 || allocations["worker-a"] != 24 || allocations["worker-b"] != 24 {
+		t.Fatalf("allocations=%v global=%d, want 24/24 and 48", allocations, global)
+	}
+}
+
+func TestAllocateWorkerCapacitiesBalancesBelowLimits(t *testing.T) {
+	workers := []activeWorkerCapacity{
+		{InstanceID: "worker-a", CapacityLimit: 24},
+		{InstanceID: "worker-b", CapacityLimit: 24},
+	}
+	allocations, global := allocateWorkerCapacities(workers, 5)
+	if global != 5 || allocations["worker-a"] != 3 || allocations["worker-b"] != 2 {
+		t.Fatalf("allocations=%v global=%d, want 3/2 and 5", allocations, global)
+	}
+}
+
+func TestAllocateWorkerCapacitiesKeepsSurvivorBounded(t *testing.T) {
+	workers := []activeWorkerCapacity{{InstanceID: "worker-a", CapacityLimit: 24}}
+	allocations, global := allocateWorkerCapacities(workers, 310)
+	if global != 24 || allocations["worker-a"] != 24 {
+		t.Fatalf("allocations=%v global=%d, want survivor capped at 24", allocations, global)
+	}
+}
+
+func TestAllocateWorkerCapacitiesSupportsUnequalLimits(t *testing.T) {
+	workers := []activeWorkerCapacity{
+		{InstanceID: "worker-a", CapacityLimit: 1},
+		{InstanceID: "worker-b", CapacityLimit: 4},
+	}
+	allocations, global := allocateWorkerCapacities(workers, 5)
+	if global != 5 || allocations["worker-a"] != 1 || allocations["worker-b"] != 4 {
+		t.Fatalf("allocations=%v global=%d, want 1/4 and 5", allocations, global)
+	}
+}

apps/api/migrations/0093_worker_instance_capacity_limit.sql

@@ -0,0 +1,8 @@
+ALTER TABLE gateway_worker_instances
+  ADD COLUMN capacity_limit integer NOT NULL DEFAULT 0,
+  ADD CONSTRAINT gateway_worker_instances_capacity_limit_check
+    CHECK (capacity_limit >= 0);
+
+UPDATE gateway_worker_instances
+SET capacity_limit = GREATEST(allocated_capacity, 1)
+WHERE capacity_limit = 0;

deploy/kubernetes/production/application-config.yaml

@@ -35,5 +35,5 @@ data:
   AI_GATEWAY_LOCAL_RESULT_MAX_BYTES: "268435456"
   AI_GATEWAY_LOCAL_RESULT_MAX_TASK_BYTES: "536870912"
   AI_GATEWAY_ASYNC_QUEUE_WORKER_ENABLED: "true"
-  AI_GATEWAY_ASYNC_WORKER_HARD_LIMIT: "16"
+  AI_GATEWAY_ASYNC_WORKER_HARD_LIMIT: "2048"
   AI_GATEWAY_ASYNC_WORKER_REFRESH_INTERVAL_SECONDS: "5"

deploy/kubernetes/production/application.yaml

@@ -312,8 +312,10 @@ spec:
               value: "true"
             - name: AI_GATEWAY_DATABASE_MAX_CONNS
               value: "24"
+            - name: AI_GATEWAY_ASYNC_WORKER_INSTANCE_HARD_LIMIT
+              value: "24"
             - name: AI_GATEWAY_MEDIA_MATERIALIZATION_CONCURRENCY
-              value: "8"
+              value: "24"
             - name: POD_NAMESPACE
               valueFrom:
                 fieldRef:
@@ -439,8 +441,10 @@ spec:
               value: "true"
             - name: AI_GATEWAY_DATABASE_MAX_CONNS
               value: "24"
+            - name: AI_GATEWAY_ASYNC_WORKER_INSTANCE_HARD_LIMIT
+              value: "24"
             - name: AI_GATEWAY_MEDIA_MATERIALIZATION_CONCURRENCY
-              value: "8"
+              value: "24"
             - name: POD_NAMESPACE
               valueFrom:
                 fieldRef:

scripts/migration-safety-reviewed.json

@@ -8,6 +8,13 @@
         "non-null column addition"
       ],
       "reason": "The migration replaces two existing constraints and one index with expanded definitions. Production preflight confirmed six schema-v1 identity revisions, no OIDC sessions or revocation watermarks, and small affected relations; the non-null columns use constant defaults."
+    },
+    "apps/api/migrations/0093_worker_instance_capacity_limit.sql": {
+      "sha256": "4b6a39ef90e7032555739ccf0568bf7ae0bce23ef98c3bee9c2c8e3bfa13291e",
+      "allowedViolations": [
+        "non-null column addition"
+      ],
+      "reason": "The table stores ephemeral Worker heartbeats and currently has only two production rows. PostgreSQL adds the integer column with a constant default, then the migration backfills each live row from its existing allocation so a rolling deployment cannot assign unsafe failover capacity."
     }
   }
 }