diff --git a/src/reflector/raw_socket.cpp b/src/reflector/raw_socket.cpp
index c35e7cd..0340047 100644
--- a/src/reflector/raw_socket.cpp
+++ b/src/reflector/raw_socket.cpp
@@ -525,14 +525,13 @@ bool RawSocket::RejoinGroups() noexcept {
             continue;
         }
         // A fresh fd rather than a re-join on the old one: where a recreated interface is handed
-        // back the number it had, the kernel still has the old fd down for that (group, index) and
-        // refuses the join as a duplicate — and a membership it keeps for a dead index still counts
-        // against the socket's join cap (Linux igmp_max_memberships, 20 by default, and not
-        // raisable on a locked-down router), so kept sockets would exhaust it after a handful of
-        // recreations. Closed before the reopen rather than after, so the descriptor it frees is
-        // the one the reopen takes: at the process fd limit that is the difference between
-        // recovering and staying deaf. Refcounts are untouched, so the memberships already handed
-        // to reflectors stay valid.
+        // back the number it had, the kernel refuses the duplicate (group, index) join — and a
+        // membership it keeps for a dead index still counts against igmp_max_memberships, so a
+        // kept fd stops joining after 20 recreations (measured; see the recreation-cap test).
+        // Closed before the reopen rather than after, so the descriptor it frees is the one the
+        // reopen takes: at the process fd limit that is the difference between recovering and
+        // staying deaf. Refcounts are untouched, so the memberships already handed to reflectors
+        // stay valid.
         auto& join_fd = join_fds_.Get(family);
         join_fd.Reset();
         join_fd.Reset(socket(family == IpAddress::Family::V6 ? AF_INET6 : AF_INET, SOCK_DGRAM, 0));
diff --git a/tests/raw_socket_test.cpp b/tests/raw_socket_test.cpp
index 332ca26..3a1ad5a 100644
--- a/tests/raw_socket_test.cpp
+++ b/tests/raw_socket_test.cpp
@@ -18,6 +18,7 @@
 #include <cstdint>
 #include <arpa/inet.h>
 #include <cstdio>
+#include <cstring>
 #include <cstdlib>
 #include <format>
 #include <fstream>
@@ -1152,6 +1153,32 @@ TEST_F(RawSocketInterfacePairRequiresRootTest, RebindRestoresTheCaptureAfterRecr
     EXPECT_EQ(socket.Fd(), fd);  // same fd throughout, so dispatcher registrations survive
 }
 
+#if defined(__linux__)
+// A kept join fd would exhaust the socket's membership cap: a destroyed interface's memberships are
+// not scrubbed from a surviving socket, so every recreation's join adds another. Measured here at
+// the default igmp_max_memberships of 20 -- the 21st join on a kept fd fails with ENOBUFS -- which
+// is why the rebind reopens rather than re-joining in place.
+TEST_F(RawSocketInterfacePairRequiresRootTest, SurvivesRepeatedRecreationsWithoutExhaustingTheJoinCap) {
+    Interface iface{pair.InjectInterface()};
+    ASSERT_TRUE(iface.IsValid());
+    RawSocket socket{iface};
+    ASSERT_TRUE(socket.IsValid());
+    const auto group = IpAddress::MdnsGroupV4();
+    auto membership = socket.JoinMulticastGroup(group);
+    ASSERT_TRUE(membership.IsValid());
+
+    // Comfortably past the cap, so a rebind holding its join fd runs out partway through.
+    for (int i = 0; i < 25; ++i) {
+        ASSERT_TRUE(pair.Recreate()) << "recreation " << i;
+        ASSERT_NE(iface.Reidentify(), Interface::IdentityChange::Parked) << "recreation " << i;
+        ASSERT_TRUE(socket.Rebind()) << "recreation " << i;
+        ASSERT_TRUE(socket.GroupsJoined()) << "recreation " << i;
+    }
+
+    EXPECT_TRUE(KernelHasMembership(pair.InjectInterface(), group));
+}
+#endif
+
 #if defined(__linux__)
 // The capture re-attaching is not enough: without the re-join the socket comes back attached and
 // permanently deaf on every group it had.