fix(replication): bin transfer summaries at 128 MiB and keep window refresh off the hot path

Second review round:

- update_xfer_rate split at 1 MiB while the minio-go transferSummary
  labels (and RustFS's own worker-pool split) mean >= 128 MiB for
  Large, so a 2 MiB replication reported under Large with Small stuck
  at zero. The producer now bins on MIN_LARGE_OBJ_SIZE; a MetricsV2
  assertion covers 2 MiB / 127 MiB / exactly 128 MiB.
- add_size no longer recomputes the rolling windows: two full
  one-hour-deque scans per failure under the bucket-stats write lock
  made failure bursts quadratic (30k events ~2.1s). The windows are
  stamped only at the collection point (get_latest_replication_stats,
  which serves both the local leg and the peer RPC); the aggregation
  regression now drives that path explicitly before the RPC round trip
  and merge.

crates/replication/src/stats.rs

@@ -543,13 +543,14 @@ impl FailStats {
         self.size = self.size.saturating_add(size);
         self.recent.push_back(FailureSample { observed_at, size });
         self.prune(observed_at);
-        self.refresh_windows();
     }
 
     /// Recompute the serializable rolling-window snapshots from the local
-    /// samples. Only meaningful on the live per-node struct: a deserialized
-    /// or merged struct has no samples, and refreshing it would wipe the
-    /// aggregated windows.
+    /// samples. Called at the collection point (per-node stats snapshot),
+    /// never on the failure hot path — the two deque scans are O(window) and
+    /// `add_size` runs under the bucket-stats write lock. Only meaningful on
+    /// the live per-node struct: a deserialized or merged struct has no
+    /// samples, and refreshing it would wipe the aggregated windows.
     pub fn refresh_windows(&mut self) {
         self.last_minute = self.recent_since(Duration::from_secs(60));
         self.last_hour = self.recent_since(Duration::from_secs(3600));
@@ -664,7 +665,9 @@ impl BucketReplicationStat {
     }
 
     pub fn update_xfer_rate(&mut self, size: i64, duration: Duration) {
-        if size > 1024 * 1024 {
+        // Same boundary as the worker-pool split and minio-go's
+        // Large/Small transfer-summary labels: >= 128 MiB is "large".
+        if size >= crate::runtime::MIN_LARGE_OBJ_SIZE {
             self.xfer_rate_lrg.add_size(size, duration);
         } else {
             self.xfer_rate_sml.add_size(size, duration);

rustfs/src/admin/replication_metrics_wire.rs

@@ -498,15 +498,49 @@ mod tests {
         assert_eq!(json["downtimeInfo"], serde_json::json!({}));
     }
 
+    /// minio-go's transferSummary labels mean >= 128 MiB for Large; the
+    /// producer must bin on the same boundary (MIN_LARGE_OBJ_SIZE, shared
+    /// with the worker-pool split), or a 2 MiB replication shows under Large
+    /// while Small stays zero.
+    #[test]
+    fn transfer_summary_bins_on_the_128_mib_boundary() {
+        const MIB: i64 = 1024 * 1024;
+        let mut stats = BucketStats::default();
+        let stat = stats
+            .replication_stats
+            .stats
+            .entry("arn:minio:replication::t:b".to_string())
+            .or_default();
+        stat.update_xfer_rate(2 * MIB, std::time::Duration::from_secs(1));
+        stat.update_xfer_rate(127 * MIB, std::time::Duration::from_secs(1));
+        stat.update_xfer_rate(128 * MIB, std::time::Duration::from_secs(1));
+
+        let json = serde_json::to_value(MetricsV2Wire::from_stats(&stats, "node-1")).expect("v2 wire should serialize");
+        let summary = &json["queueStats"]["nodes"][0]["tgtTransferStats"]["arn:minio:replication::t:b"];
+        let small_peak = summary["Small"]["peakRate"].as_f64().expect("Small peakRate");
+        let large_peak = summary["Large"]["peakRate"].as_f64().expect("Large peakRate");
+        assert!(
+            (small_peak - (127 * MIB) as f64).abs() < 1.0,
+            "2 MiB and 127 MiB transfers must bin as Small (peak {small_peak})"
+        );
+        assert!(
+            (large_peak - (128 * MIB) as f64).abs() < 1.0,
+            "exactly 128 MiB must bin as Large (peak {large_peak})"
+        );
+    }
+
     /// Review regression: both metrics endpoints aggregate first, and the
     /// FailStats merge drops the process-local samples — the rolling windows
     /// must survive a peer-RPC round trip plus aggregation and still reach
     /// the wire body.
     #[test]
     fn failure_windows_survive_aggregation_before_serialization() {
-        // Node A: live failure, windows stamped at the collection point.
+        // Node A: live failure; the windows are stamped at the collection
+        // point (get_latest_replication_stats calls refresh_windows before
+        // the stats cross the wire), never on the failure hot path.
         let mut node_a = crate::admin::storage_api::replication::BucketReplicationStat::default();
         node_a.fail_stats.add_size(512, None::<&std::io::Error>);
+        node_a.fail_stats.refresh_windows();
         node_a.failed = node_a.fail_stats.to_metric();
 
         // Node A's stats cross the peer RPC wire: the samples are dropped,