fix(worker): 记住轮询间隔内的容量需求

异步视频任务在 River 轮询空档会释放本地 lease，采样瞬间可能观察不到已经触顶的领取需求，导致自适应容量长期停在初始值。本提交记录采样窗口内出现过的容量饱和，同时保持空闲 Worker 不盲目升容。\n\n移植时保留主线现有的单节点容量阶梯、请求下限和严格失败门禁，仅合入失败路径先记录并排空任务与回调、再删除模拟器资源的清理顺序，并补充脚本守卫。\n\n验证：go test ./... -count=1；go vet ./...；bash -n；ShellCheck；production-acceptance-script-test.sh。

apps/api/internal/workerload/controller.go

@@ -70,13 +70,14 @@ type Snapshot struct {
 type Controller struct {
 	mu sync.Mutex
 
-	mode           string
-	hardLimit      int
-	activeLimit    int
-	heavyLimit     int
-	claimLimit     int
-	healthySamples int
-	healthyCount   int
+	mode                string
+	hardLimit           int
+	activeLimit         int
+	heavyLimit          int
+	claimLimit          int
+	healthySamples      int
+	healthyCount        int
+	demandedSinceSample bool
 
 	preparing  int
 	waiting    int
@@ -179,10 +180,20 @@ func (c *Controller) TryStart() (*Lease, bool) {
 	c.mu.Lock()
 	defer c.mu.Unlock()
 	activeLimit := min(c.activeLimit, c.claimLimit)
-	if activeLimit <= 0 || c.activeLocked() >= activeLimit || c.preparing+c.finalizing >= c.heavyLimit {
+	if activeLimit <= 0 {
+		return nil, false
+	}
+	if c.activeLocked() >= activeLimit {
+		c.demandedSinceSample = true
+		return nil, false
+	}
+	if c.preparing+c.finalizing >= c.heavyLimit {
 		return nil, false
 	}
 	c.preparing++
+	if c.activeLocked() >= activeLimit {
+		c.demandedSinceSample = true
+	}
 	c.last = c.snapshotLocked(ResourceSample{SampledAt: c.last.SampledAt})
 	return &Lease{controller: c, phase: PhasePreparing}, true
 }
@@ -276,6 +287,9 @@ func (l *Lease) Phase() Phase {
 }
 
 func (c *Controller) adjustLocked(state PressureState, memory, cpu, database float64) {
+	activeLimit := min(c.activeLimit, c.claimLimit)
+	demanded := c.demandedSinceSample || (activeLimit > 0 && c.activeLocked() >= activeLimit)
+	c.demandedSinceSample = false
 	switch state {
 	case PressureCritical:
 		c.healthyCount = 0
@@ -287,7 +301,7 @@ func (c *Controller) adjustLocked(state PressureState, memory, cpu, database flo
 		c.activeLimit = max(1, c.activeLimit-step)
 		c.heavyLimit = min(c.heavyLimit, max(1, (c.activeLimit+3)/4))
 	default:
-		if memory > .60 || cpu > .65 || database > .65 || c.activeLocked() < min(c.activeLimit, c.claimLimit) {
+		if memory > .60 || cpu > .65 || database > .65 || !demanded {
 			c.healthyCount = 0
 			return
 		}