diff --git a/server.go b/server.go
index 0ebf558..b306567 100644
--- a/server.go
+++ b/server.go
@@ -97,24 +97,42 @@ func (s *Server) removeSub(conn net.Conn) {
 	}
 }
 
+// publish fan-outs evt to every subscriber on evt.Topic and "*" (minus
+// the sender). The subscriber set is snapshotted under a brief RLock
+// and the lock is released before any WriteEvent I/O — otherwise a slow
+// subscriber (full pipe / TCP send buffer) would block WriteEvent while
+// the RLock is still held, starving every concurrent addSub/removeSub
+// caller waiting on mu.Lock(). See broker.publishWith in service.go for
+// the same pattern (P2-003).
 func (s *Server) publish(evt *Event, sender net.Conn) {
 	s.mu.RLock()
-	defer s.mu.RUnlock()
-
-	// Send to topic-specific subscribers
+	targets := make([]net.Conn, 0, len(s.subs[evt.Topic])+len(s.subs["*"]))
 	for _, conn := range s.subs[evt.Topic] {
 		if conn != sender {
-			WriteEvent(conn, evt)
+			targets = append(targets, conn)
 		}
 	}
-	// Send to wildcard subscribers
 	if evt.Topic != "*" {
 		for _, conn := range s.subs["*"] {
 			if conn != sender {
-				WriteEvent(conn, evt)
+				targets = append(targets, conn)
 			}
 		}
 	}
+	s.mu.RUnlock()
+
+	for _, conn := range targets {
+		if err := WriteEvent(conn, evt); err != nil {
+			slog.Debug("eventstream write failed",
+				"remote", conn.RemoteAddr(),
+				"topic", evt.Topic,
+				"error", err)
+		}
+	}
 
-	slog.Debug("eventstream published", "topic", evt.Topic, "bytes", len(evt.Payload), "from", sender.RemoteAddr())
+	slog.Debug("eventstream published",
+		"topic", evt.Topic,
+		"bytes", len(evt.Payload),
+		"from", sender.RemoteAddr(),
+		"targets", len(targets))
 }
diff --git a/zz_server_internals_test.go b/zz_server_internals_test.go
index a402ee0..9e9264d 100644
--- a/zz_server_internals_test.go
+++ b/zz_server_internals_test.go
@@ -4,7 +4,9 @@ package eventstream
 
 import (
 	"net"
+	"sync"
 	"testing"
+	"time"
 )
 
 // TestServer_AddRemoveSubDirect drives the bookkeeping methods on the
@@ -97,6 +99,207 @@ func TestServer_Publish_WildcardSubscriber(t *testing.T) {
 	}
 }
 
+// TestServer_PublishDoesNotBlockOnSlowSubscriber is a regression test
+// for the bug where Server.publish held mu.RLock() across WriteEvent
+// I/O. With the bug, a slow subscriber (full pipe buffer) wedged the
+// RLock, blocking every concurrent addSub/removeSub waiting on
+// mu.Lock(). The fix snapshots the subscriber list under a brief lock
+// then releases before iterating, so subscription mutations stay
+// responsive even while a peer is wedged.
+func TestServer_PublishDoesNotBlockOnSlowSubscriber(t *testing.T) {
+	t.Parallel()
+	s := &Server{subs: map[string][]net.Conn{}}
+
+	// Fast subscriber: drain immediately in a goroutine.
+	fast, fastPeer := net.Pipe()
+	defer fast.Close()
+	defer fastPeer.Close()
+	go func() {
+		for {
+			if _, err := ReadEvent(fastPeer); err != nil {
+				return
+			}
+		}
+	}()
+
+	// Slow subscriber: never read until the test releases it. With the
+	// old code, WriteEvent blocks on the pipe write, and the RLock is
+	// held the whole time.
+	slow, slowPeer := net.Pipe()
+	defer slow.Close()
+	defer slowPeer.Close()
+
+	release := make(chan struct{})
+	go func() {
+		// Drain the slow side only after the test signals.
+		<-release
+		for {
+			if _, err := ReadEvent(slowPeer); err != nil {
+				return
+			}
+		}
+	}()
+
+	s.addSub("topic", fast)
+	s.addSub("topic", slow)
+
+	// Drive a publish from a sender that isn't subscribed.
+	sender, _ := net.Pipe()
+	defer sender.Close()
+
+	publishDone := make(chan struct{})
+	go func() {
+		defer close(publishDone)
+		s.publish(&Event{Topic: "topic", Payload: []byte("payload")}, sender)
+	}()
+
+	// Give publish a moment to enter — with the buggy implementation
+	// it acquires RLock and starts writing to fast (which drains),
+	// then blocks on slow's WriteEvent while still holding RLock.
+	// With the fixed implementation it has already snapshotted and
+	// released RLock before either WriteEvent call.
+	time.Sleep(20 * time.Millisecond)
+
+	// Concurrently try to mutate the subscription set. With the bug,
+	// addSub blocks on mu.Lock() while the publish holds RLock.
+	extra, _ := net.Pipe()
+	defer extra.Close()
+
+	addDone := make(chan struct{})
+	addStart := time.Now()
+	go func() {
+		s.addSub("other-topic", extra)
+		close(addDone)
+	}()
+
+	select {
+	case <-addDone:
+		elapsed := time.Since(addStart)
+		if elapsed > 100*time.Millisecond {
+			t.Fatalf("addSub took %v with slow subscriber present; want <100ms", elapsed)
+		}
+	case <-time.After(100 * time.Millisecond):
+		t.Fatalf("addSub blocked >100ms behind slow subscriber — publish is holding RLock across I/O")
+	}
+
+	// Drain the slow side so publish can finish; the test goroutines clean up.
+	close(release)
+	select {
+	case <-publishDone:
+	case <-time.After(2 * time.Second):
+		t.Fatal("publish never completed after slow subscriber drained")
+	}
+}
+
+// TestServer_PublishHandlesSlowSubscriberWithoutDeadlock exercises the
+// happy-path semantics alongside the slow-subscriber scenario: a fast
+// subscriber still receives the event promptly, multiple publishers
+// can run concurrently against a wedged peer, and removeSub during the
+// wedge succeeds without deadlock.
+func TestServer_PublishHandlesSlowSubscriberWithoutDeadlock(t *testing.T) {
+	t.Parallel()
+	s := &Server{subs: map[string][]net.Conn{}}
+
+	fast, fastPeer := net.Pipe()
+	defer fast.Close()
+	defer fastPeer.Close()
+
+	slow, slowPeer := net.Pipe()
+	defer slow.Close()
+	defer slowPeer.Close()
+
+	// Fast peer drains continuously; it asserts the first event was
+	// delivered, then keeps reading to unblock the rest of the publishers.
+	received := make(chan *Event, 8)
+	fastDone := make(chan struct{})
+	go func() {
+		defer close(fastDone)
+		for {
+			evt, err := ReadEvent(fastPeer)
+			if err != nil {
+				return
+			}
+			select {
+			case received <- evt:
+			default:
+			}
+		}
+	}()
+
+	// Slow peer never reads — its pipe stays wedged for the lifetime
+	// of the test. We rely on Close() at the end to unblock anything
+	// still stuck inside WriteEvent.
+
+	s.addSub("t", fast)
+	s.addSub("t", slow)
+
+	sender, _ := net.Pipe()
+	defer sender.Close()
+
+	// Kick off three publishers in parallel. None of them must
+	// deadlock the broker; the fast peer must get all three events.
+	var wg sync.WaitGroup
+	for i := 0; i < 3; i++ {
+		wg.Add(1)
+		go func() {
+			defer wg.Done()
+			s.publish(&Event{Topic: "t", Payload: []byte("p")}, sender)
+		}()
+	}
+
+	// The fast peer must receive the first event quickly.
+	select {
+	case got := <-received:
+		if got.Topic != "t" || string(got.Payload) != "p" {
+			t.Errorf("fast peer got unexpected event: %+v", got)
+		}
+	case <-time.After(500 * time.Millisecond):
+		t.Fatal("fast peer never received event — broker is wedged by slow peer")
+	}
+
+	// removeSub on the slow peer must not deadlock even while
+	// publish goroutines are mid-flight blocked on the pipe.
+	removeDone := make(chan struct{})
+	go func() {
+		s.removeSub(slow)
+		close(removeDone)
+	}()
+	select {
+	case <-removeDone:
+	case <-time.After(200 * time.Millisecond):
+		t.Fatal("removeSub blocked while slow subscriber wedged the broker")
+	}
+
+	// Drain the slow side so the in-flight publishers can complete.
+	// We use a goroutine so closing slow/slowPeer in defer also
+	// works as a safety net.
+	drainDone := make(chan struct{})
+	go func() {
+		defer close(drainDone)
+		for {
+			if _, err := ReadEvent(slowPeer); err != nil {
+				return
+			}
+		}
+	}()
+
+	done := make(chan struct{})
+	go func() {
+		wg.Wait()
+		close(done)
+	}()
+	select {
+	case <-done:
+	case <-time.After(2 * time.Second):
+		t.Fatal("publishers never drained after slow peer started draining")
+	}
+	// Close both sides to terminate the drain + fast reader goroutines.
+	slowPeer.Close()
+	fastPeer.Close()
+	<-drainDone
+	<-fastDone
+}
+
 // TestNewServer covers the driver-less constructor branch.
 func TestNewServer_NilDriverPanicsOnUse(t *testing.T) {
 	t.Parallel()