Tiny support sgl-router http response status code metrics (#14689)

sgl-model-gateway/src/middleware.rs

@@ -332,9 +332,12 @@ impl Default for ResponseLogger {
 impl<B> OnResponse<B> for ResponseLogger {
     fn on_response(self, response: &Response<B>, latency: Duration, span: &Span) {
         let status = response.status();
+        let status_code = status.as_u16();
+
+        RouterMetrics::record_http_status_code(status_code);
 
         // Record these in the span for structured logging/observability tools
-        span.record("status_code", status.as_u16());
+        span.record("status_code", status_code);
         // Use microseconds as integer to avoid format! string allocation
         span.record("latency", latency.as_micros() as u64);
 

sgl-model-gateway/src/observability/metrics.rs

@@ -263,6 +263,11 @@ pub fn init_metrics() {
         "sgl_tokenizer_factory_load_duration_seconds",
         "Time to load and initialize tokenizer"
     );
+
+    describe_counter!(
+        "sgl_router_http_responses_total",
+        "Total number of HTTP responses by status code"
+    );
 }
 
 pub fn start_prometheus(config: PrometheusConfig) {
@@ -563,6 +568,13 @@ impl RouterMetrics {
     pub fn record_job_shutdown_rejected() {
         counter!("sgl_router_job_shutdown_rejected_total").increment(1);
     }
+
+    pub fn record_http_status_code(status_code: u16) {
+        counter!("sgl_router_http_responses_total",
+            "status_code" => status_code.to_string()
+        )
+        .increment(1);
+    }
 }
 
 impl TokenizerMetrics {