diff --git a/replication_convergence_test.go b/replication_convergence_test.go
new file mode 100644
index 00000000..684da009
--- /dev/null
+++ b/replication_convergence_test.go
@@ -0,0 +1,135 @@
+package weshnet_test
+
+import (
+	"context"
+	"fmt"
+	"testing"
+	"time"
+
+	mocknet "github.com/libp2p/go-libp2p/p2p/net/mock"
+	"github.com/stretchr/testify/require"
+
+	weshnet "berty.tech/weshnet/v2"
+	"berty.tech/weshnet/v2/pkg/protocoltypes"
+	"berty.tech/weshnet/v2/pkg/testutil"
+)
+
+// TestReplicationConvergence_OfflinePeerCatchesUp asserts that a peer which
+// goes offline, misses messages and then reconnects catches up on every missed
+// message in a single orbit-db head exchange — with no application-level retry.
+//
+// It runs the full weshnet protocol service (NewTestingProtocolWithMockedPeers).
+// Note the one mocknet-imposed compromise: mocknet does not run the discovery
+// poll + BackoffConnector that re-dials a dropped peer on a real network (a
+// probe confirms a peer never reconnects on its own after the link is
+// restored), so the test re-establishes the connection itself with a single
+// ConnectPeers — modelling the one network reconnect that discovery would
+// produce. Everything after that is the real stack: the reconnect re-fires the
+// head exchange, which backfills the whole log. No retry loop is used.
+//
+// A single reconnect is reliable here because this isolated, clean reconnect
+// rarely trips the go-libp2p-pubsub connect/disconnect race that can drop a
+// peer's subscriptions on reconnect (the reconnect's new-peer event is deduped
+// against the not-yet-removed stale peer entry). It is low-probability, not
+// immune: the busier bertyreplication topology hits it ~1/100 and needs a
+// reconnect retry loop.
+func TestScenario_ReplicationConvergenceOfflinePeerCatchesUp(t *testing.T) {
+	testutil.FilterStabilityAndSpeed(t, testutil.Flappy, testutil.Slow)
+
+	ctx, cancel := context.WithCancel(context.Background())
+	defer cancel()
+
+	logger, cleanupLog := testutil.Logger(t)
+	defer cleanupLog()
+
+	mn := mocknet.New()
+	defer mn.Close()
+
+	opts := weshnet.TestingOpts{
+		Mocknet:     mn,
+		Logger:      logger,
+		ConnectFunc: weshnet.ConnectAll,
+	}
+	tps, cleanupTps := weshnet.NewTestingProtocolWithMockedPeers(ctx, t, &opts, nil, 2)
+	defer cleanupTps()
+
+	alice, bob := tps[0], tps[1]
+	aliceID, bobID := alice.Opts.Host.ID(), bob.Opts.Host.ID()
+
+	// Create, join and activate the group on both peers.
+	group := weshnet.CreateMultiMemberGroupInstance(ctx, t, alice, bob)
+	groupPK := group.PublicKey
+
+	// Baseline: bob receives a message from alice while connected, proving the
+	// group is live before we partition them.
+	require.NoError(t, sendMessages(ctx, alice, groupPK, []string{"online-1"}))
+	waitForMessages(ctx, t, bob, groupPK, []string{"online-1"}, time.Second*20)
+
+	// Take bob offline: sever the mocknet link (so no re-dial can succeed) and
+	// close the existing connection.
+	require.NoError(t, mn.UnlinkPeers(aliceID, bobID))
+	require.NoError(t, mn.DisconnectPeers(aliceID, bobID))
+
+	// Alice keeps sending while bob is offline.
+	const offlineCount = 20
+	offline := make([]string, offlineCount)
+	for i := range offline {
+		offline[i] = fmt.Sprintf("offline-%d", i)
+	}
+	require.NoError(t, sendMessages(ctx, alice, groupPK, offline))
+
+	// Bring bob back online: restore the link and re-establish the connection
+	// once. mocknet won't re-dial on its own, so this single ConnectPeers stands
+	// in for the discovery-driven reconnect; we do not retry it, and we never
+	// touch orbit-db. The reconnect re-fires the head exchange on its own.
+	_, err := mn.LinkPeers(aliceID, bobID)
+	require.NoError(t, err)
+	_, err = mn.ConnectPeers(aliceID, bobID)
+	require.NoError(t, err)
+
+	// Assert bob catches up on every message it missed, with no manual help
+	// beyond that single reconnect and no head-exchange retry.
+	waitForMessages(ctx, t, bob, groupPK, offline, time.Second*60)
+}
+
+func sendMessages(ctx context.Context, tp *weshnet.TestingProtocol, groupPK []byte, messages []string) error {
+	for _, m := range messages {
+		if _, err := tp.Client.AppMessageSend(ctx, &protocoltypes.AppMessageSend_Request{
+			GroupPk: groupPK,
+			Payload: []byte(m),
+		}); err != nil {
+			return err
+		}
+	}
+	return nil
+}
+
+// waitForMessages subscribes to the group and blocks until every wanted payload
+// has been observed or the timeout fires. GroupMessageList replays the store
+// and then tails new events, so messages backfilled after the subscription
+// starts are streamed too.
+func waitForMessages(ctx context.Context, t *testing.T, tp *weshnet.TestingProtocol, groupPK []byte, want []string, timeout time.Duration) {
+	t.Helper()
+
+	ctx, cancel := context.WithTimeout(ctx, timeout)
+	defer cancel()
+
+	remaining := map[string]struct{}{}
+	for _, m := range want {
+		remaining[m] = struct{}{}
+	}
+
+	sub, err := tp.Client.GroupMessageList(ctx, &protocoltypes.GroupMessageList_Request{
+		GroupPk: groupPK,
+	})
+	require.NoError(t, err)
+
+	for len(remaining) > 0 {
+		res, err := sub.Recv()
+		if err != nil {
+			require.NoError(t, err, "peer did not catch up, still missing %d/%d message(s)", len(remaining), len(want))
+			return
+		}
+		delete(remaining, string(res.Message))
+	}
+}