perf(cmpp): isolate inbound transport capacity

gateway/cmd/gateway/main.go

@@ -50,13 +50,15 @@ func main() {
 
 	go func() {
 		log.Printf("cmpp gateway inbound server listening on %s", cmppAddr)
+		inboundConcurrency := positiveEnvInt("GATEWAY_CMPP_INBOUND_MAX_CONCURRENCY", 64)
 		if err := (inbound.Server{
 			Addr:                 cmppAddr,
 			APIBaseURL:           apiBaseURL,
+			HTTPClient:           inbound.NewAPIHTTPClient(inboundConcurrency),
 			PresenceStore:        presenceStore,
 			RecoveryStore:        recoveryStore,
 			GatewayInstanceID:    getenv("GATEWAY_INSTANCE_ID", hostname()),
-			MaxSubmitConcurrency: positiveEnvInt("GATEWAY_CMPP_INBOUND_MAX_CONCURRENCY", 64),
+			MaxSubmitConcurrency: inboundConcurrency,
 			SecurityEventToken:   os.Getenv("SECURITY_EVENT_TOKEN"),
 		}).ListenAndServe(); err != nil {
 			log.Fatalf("gateway inbound server stopped: %v", err)

gateway/internal/inbound/http_client.go

@@ -0,0 +1,26 @@
+package inbound
+
+import (
+	"net"
+	"net/http"
+	"time"
+)
+
+// NewAPIHTTPClient keeps enough loopback connections warm for the bounded CMPP
+// Submit window. Go's default of two idle connections per host otherwise causes
+// connection churn exactly when SubmitResp latency matters most.
+func NewAPIHTTPClient(maxConcurrency int) *http.Client {
+	if maxConcurrency < 1 {
+		maxConcurrency = 64
+	}
+	transport := http.DefaultTransport.(*http.Transport).Clone()
+	transport.MaxIdleConns = maxConcurrency + 16
+	transport.MaxIdleConnsPerHost = maxConcurrency
+	transport.MaxConnsPerHost = maxConcurrency
+	transport.IdleConnTimeout = 90 * time.Second
+	transport.DialContext = (&net.Dialer{
+		Timeout:   3 * time.Second,
+		KeepAlive: 30 * time.Second,
+	}).DialContext
+	return &http.Client{Transport: transport, Timeout: defaultHTTPTimeout}
+}

gateway/internal/inbound/http_client_test.go

@@ -0,0 +1,28 @@
+package inbound
+
+import (
+	"net/http"
+	"testing"
+)
+
+func TestNewAPIHTTPClientMatchesBoundedSubmitConcurrency(t *testing.T) {
+	client := NewAPIHTTPClient(48)
+	transport, ok := client.Transport.(*http.Transport)
+	if !ok {
+		t.Fatalf("expected *http.Transport, got %T", client.Transport)
+	}
+	if transport.MaxConnsPerHost != 48 || transport.MaxIdleConnsPerHost != 48 {
+		t.Fatalf("unexpected host connection bounds: max=%d idle=%d", transport.MaxConnsPerHost, transport.MaxIdleConnsPerHost)
+	}
+	if client.Timeout != defaultHTTPTimeout {
+		t.Fatalf("unexpected client timeout: %s", client.Timeout)
+	}
+}
+
+func TestNewAPIHTTPClientUsesSafeDefault(t *testing.T) {
+	client := NewAPIHTTPClient(0)
+	transport := client.Transport.(*http.Transport)
+	if transport.MaxConnsPerHost != 64 {
+		t.Fatalf("expected default max connections 64, got %d", transport.MaxConnsPerHost)
+	}
+}