diff --git a/.config/nextest.toml b/.config/nextest.toml
index 191e546..35d6aea 100644
--- a/.config/nextest.toml
+++ b/.config/nextest.toml
@@ -5,6 +5,23 @@ junit = { path = "junit.xml" }
 # occasional msg1 under burst load even with the per-edge repair loop.
 # Allow a retry rather than failing the whole CI run on a single
 # dropped packet.
+#
+# Deliberately scoped to [profile.ci] and NOT applied locally, which makes
+# the two gates disagree: the hosted runner retries a flaky test twice, the
+# local sweep fails on the first failure. The asymmetry is intended and this
+# is the record of why, since an undocumented one is indistinguishable from
+# an oversight.
+#
+# It is here for shared-runner packet loss, a property of the hosted
+# environment and not of the code. Applying it locally would suppress a real
+# local flake, and a failure that only reproduces under load is a robustness
+# bug to fix rather than to retry past. Keeping the local sweep strict is
+# what makes it the sharper of the two gates.
+#
+# The cost, stated rather than hidden: a test that fails once and passes on
+# retry is reported green here with no separate signal, so a genuine
+# intermittent failure can be absorbed. If that starts mattering, the fix is
+# to surface retried-but-passed tests, not to drop the retries.
 retries = 2
 
 [test-groups]
diff --git a/src/transport/udp/socket.rs b/src/transport/udp/socket.rs
index 6ba77f9..08f3485 100644
--- a/src/transport/udp/socket.rs
+++ b/src/transport/udp/socket.rs
@@ -853,7 +853,7 @@ mod tests {
     /// find N packets already buffered, and one syscall reaps the burst.
     #[cfg(any(target_os = "linux", target_os = "macos"))]
     #[tokio::test(flavor = "multi_thread", worker_threads = 2)]
-    #[ignore]
+    #[ignore = "microbenchmark; run explicitly with --ignored --nocapture"]
     async fn bench_udp_recv_amortization() {
         use std::sync::Arc;
         use std::sync::atomic::{AtomicBool, Ordering};