fix(routing): 避免路由探测耗尽数据库连接

将跨实例探测协调从持有会话级 advisory lock 改为带过期时间的数据库短租约，避免网络探测期间占用关键 PostgreSQL 连接。新增可配置的探测并发上限和真实 PostgreSQL 回归测试，确保租约生效期间连接池仍可全部获取。\n\n验证：go test ./... -count=1；go vet ./...；真实 PostgreSQL 17 集成测试；迁移安全检查。

apps/api/internal/store/execution_pools.go

@@ -222,6 +222,49 @@ WHERE expires_at > now()`)
 	return items, rows.Err()
 }
 
+// TryClaimRouteProbe coordinates probes without holding a database connection
+// during DNS/TCP/TLS/HTTP I/O. The previous session advisory lock pinned one
+// critical-pool connection per in-flight probe and could starve readiness.
+func (s *Store) TryClaimRouteProbe(ctx context.Context, poolID, routeProfileKey string, ttl time.Duration) (string, bool, error) {
+	poolID = strings.TrimSpace(poolID)
+	routeProfileKey = strings.TrimSpace(routeProfileKey)
+	if poolID == "" || routeProfileKey == "" {
+		return "", false, errors.New("route probe pool and profile are required")
+	}
+	if ttl <= 0 {
+		ttl = 15 * time.Second
+	}
+	token := uuid.NewString()
+	var claimed string
+	err := s.pool.QueryRow(ctx, `
+INSERT INTO gateway_route_probe_leases (
+  pool_id, route_profile_key, lease_token, expires_at, updated_at
+)
+VALUES ($1, $2, $3::uuid, now() + $4::interval, now())
+ON CONFLICT (pool_id, route_profile_key) DO UPDATE
+SET lease_token = EXCLUDED.lease_token,
+    expires_at = EXCLUDED.expires_at,
+    updated_at = now()
+WHERE gateway_route_probe_leases.expires_at <= now()
+RETURNING lease_token::text`, poolID, routeProfileKey, token, ttl.String()).Scan(&claimed)
+	if errors.Is(err, pgx.ErrNoRows) {
+		return "", false, nil
+	}
+	if err != nil {
+		return "", false, err
+	}
+	return claimed, true, nil
+}
+
+func (s *Store) ReleaseRouteProbe(ctx context.Context, poolID, routeProfileKey, token string) error {
+	_, err := s.pool.Exec(ctx, `
+DELETE FROM gateway_route_probe_leases
+WHERE pool_id = $1
+  AND route_profile_key = $2
+  AND lease_token = $3::uuid`, strings.TrimSpace(poolID), strings.TrimSpace(routeProfileKey), strings.TrimSpace(token))
+	return err
+}
+
 func (s *Store) ListRouteHealth(ctx context.Context, routeProfileKey string, _ time.Time) ([]executionpool.RouteHealth, error) {
 	return s.listRouteHealth(ctx, "WHERE route_profile_key = $1", strings.TrimSpace(routeProfileKey))
 }

apps/api/internal/store/execution_pools_integration_test.go

@@ -4,6 +4,9 @@ import (
 	"context"
 	"testing"
 	"time"
+
+	"github.com/easyai/easyai-ai-gateway/apps/api/internal/executionpool"
+	"github.com/jackc/pgx/v5/pgxpool"
 )
 
 func TestExecutionPoolQueriesUseTimestampCutoff(t *testing.T) {
@@ -18,3 +21,42 @@ func TestExecutionPoolQueriesUseTimestampCutoff(t *testing.T) {
 		t.Fatalf("list capacity with timestamp cutoff: %v", err)
 	}
 }
+
+func TestRouteProbeLeaseDoesNotPinDatabaseConnection(t *testing.T) {
+	db := billingV2IntegrationStore(t)
+	ctx, cancel := context.WithTimeout(context.Background(), 10*time.Second)
+	defer cancel()
+
+	if err := db.UpsertExecutionPool(ctx, executionpool.ExecutionPool{ID: "probe-pool", State: executionpool.PoolActive}); err != nil {
+		t.Fatalf("upsert execution pool: %v", err)
+	}
+	token, claimed, err := db.TryClaimRouteProbe(ctx, "probe-pool", "route-profile", 15*time.Second)
+	if err != nil {
+		t.Fatalf("claim route probe: %v", err)
+	}
+	if !claimed || token == "" {
+		t.Fatal("route probe was not claimed")
+	}
+
+	connections := make([]*pgxpool.Conn, 0, db.Pool().Stat().MaxConns())
+	defer func() {
+		for _, connection := range connections {
+			connection.Release()
+		}
+	}()
+	for range db.Pool().Stat().MaxConns() {
+		connection, err := db.Pool().Acquire(ctx)
+		if err != nil {
+			t.Fatalf("acquire all pool connections while lease is active: %v", err)
+		}
+		connections = append(connections, connection)
+	}
+	for _, connection := range connections {
+		connection.Release()
+	}
+	connections = connections[:0]
+
+	if err := db.ReleaseRouteProbe(ctx, "probe-pool", "route-profile", token); err != nil {
+		t.Fatalf("release route probe: %v", err)
+	}
+}