feat(queue): 增加非文本模型分布式准入队列

使用 PostgreSQL 统一同步与异步非文本任务的并发准入、持久化等待和 Worker 容量分配，并将生产 API 与独立 Worker 角色拆分。

补充策略管理、共享契约、OpenAPI、Kubernetes 双节点 Worker 清单及跨节点验收脚本；未默认启用任何生产 queue_size 策略。

已在原基线完成 Go、前端、迁移、Shell、Kustomize 与长任务容量验收；合入最新主干后将重新执行发布门禁。

apps/api/internal/store/rate_limit_status.go

@@ -42,6 +42,10 @@ type ModelRateLimitStatus struct {
 	ModelCooldownUntil        string                   `json:"modelCooldownUntil,omitempty"`
 	Concurrent                RateLimitMetricStatus    `json:"concurrent"`
 	QueuedTasks               float64                  `json:"queuedTasks"`
+	WaitingSyncTasks          int                      `json:"waitingSyncTasks,omitempty"`
+	WaitingAsyncTasks         int                      `json:"waitingAsyncTasks,omitempty"`
+	OldestQueueWaitSeconds    float64                  `json:"oldestQueueWaitSeconds,omitempty"`
+	QueueLimit                int                      `json:"queueLimit,omitempty"`
 	RPM                       RateLimitMetricStatus    `json:"rpm"`
 	TPM                       RateLimitMetricStatus    `json:"tpm"`
 	LoadRatio                 float64                  `json:"loadRatio"`
@@ -151,6 +155,9 @@ func (s *Store) ListModelRateLimitStatuses(ctx context.Context) ([]ModelRateLimi
 	       COALESCE(to_char(m.cooldown_until AT TIME ZONE 'UTC', 'YYYY-MM-DD"T"HH24:MI:SS.MS"Z"'), ''),
 	       COALESCE(con.active, 0)::float8,
 	       COALESCE(queued.waiting, 0)::float8,
+	       COALESCE(admission.waiting_sync, 0)::int,
+	       COALESCE(admission.waiting_async, 0)::int,
+	       COALESCE(admission.oldest_wait_seconds, 0)::float8,
 	       COALESCE(rpm.used_value, 0)::float8, COALESCE(rpm.reserved_value, 0)::float8, COALESCE(rpm.reset_at::text, ''),
 	       COALESCE(tpm.used_value, 0)::float8, COALESCE(tpm.reserved_value, 0)::float8, COALESCE(tpm.reset_at::text, '')
 FROM platform_models m
@@ -192,6 +199,15 @@ LEFT JOIN (
   ) queued_sources
   GROUP BY queued_sources.platform_model_id
 ) queued ON queued.platform_model_id = m.id::text
+LEFT JOIN (
+  SELECT platform_model_id::text AS platform_model_id,
+         COUNT(*) FILTER (WHERE mode = 'sync') AS waiting_sync,
+         COUNT(*) FILTER (WHERE mode = 'async') AS waiting_async,
+         EXTRACT(EPOCH FROM now() - MIN(enqueued_at)) AS oldest_wait_seconds
+  FROM gateway_task_admissions
+  WHERE status = 'waiting'
+  GROUP BY platform_model_id
+) admission ON admission.platform_model_id = m.id::text
 LEFT JOIN (
   SELECT DISTINCT ON (scope_key) scope_key, used_value, reserved_value, reset_at
   FROM gateway_rate_limit_counters
@@ -231,6 +247,9 @@ ORDER BY p.priority ASC, m.model_name ASC`)
 		var modelCooldownUntil string
 		var concurrentCurrent float64
 		var queuedTasks float64
+		var waitingSyncTasks int
+		var waitingAsyncTasks int
+		var oldestQueueWaitSeconds float64
 		var rpmUsed float64
 		var rpmReserved float64
 		var rpmResetAt string
@@ -263,6 +282,9 @@ ORDER BY p.priority ASC, m.model_name ASC`)
 			&modelCooldownUntil,
 			&concurrentCurrent,
 			&queuedTasks,
+			&waitingSyncTasks,
+			&waitingAsyncTasks,
+			&oldestQueueWaitSeconds,
 			&rpmUsed,
 			&rpmReserved,
 			&rpmResetAt,
@@ -287,6 +309,12 @@ ORDER BY p.priority ASC, m.model_name ASC`)
 		item.ModelCooldownUntil = modelCooldownUntil
 		item.RateLimitPolicy = policy
 		item.QueuedTasks = queuedTasks
+		item.WaitingSyncTasks = waitingSyncTasks
+		item.WaitingAsyncTasks = waitingAsyncTasks
+		item.OldestQueueWaitSeconds = oldestQueueWaitSeconds
+		if queueRule, ok := QueueRuleFromPolicy(policy); ok {
+			item.QueueLimit = queueRule.Limit
+		}
 		item.Concurrent = metricStatus(concurrentCurrent, concurrentCurrent, 0, rateLimitForMetric(policy, "concurrent"), "")
 		item.RPM = metricStatus(rpmUsed+rpmReserved, rpmUsed, rpmReserved, rateLimitForMetric(policy, "rpm"), rpmResetAt)
 		item.TPM = metricStatus(tpmUsed+tpmReserved, tpmUsed, tpmReserved, tpmLimit(policy), tpmResetAt)