diff --git a/internal/e2e/local_soak_test.go b/internal/e2e/local_soak_test.go
new file mode 100644
index 0000000..bfbba18
--- /dev/null
+++ b/internal/e2e/local_soak_test.go
@@ -0,0 +1,350 @@
+package e2e
+
+import (
+	"bufio"
+	"bytes"
+	"context"
+	"errors"
+	"flag"
+	"fmt"
+	"io"
+	"net"
+	"sync"
+	"sync/atomic"
+	"testing"
+	"time"
+
+	"github.com/openlibrecommunity/olcrtc/internal/client"
+	"github.com/openlibrecommunity/olcrtc/internal/server"
+)
+
+// Local throughput soak: pump as much traffic as the selected transport
+// can sustain, locally, for an arbitrary duration.
+//
+// The tunnel is built on the in-memory carrier (no real provider, no
+// network), so this measures the upper bound of what the
+// SOCKS+muxconn+transport stack can do on this machine. Useful to:
+//
+//   - leave running for hours and watch for goroutine / memory growth
+//   - reproduce slow-leak corruption with the byte-pattern verifier
+//   - get a feel for raw transport throughput before touching real WebRTC
+//
+// Quick start:
+//
+//	go test -count=1 -v ./internal/e2e \
+//	    -run '^TestLocalThroughputSoak$' \
+//	    -olcrtc.local-soak \
+//	    -olcrtc.local-soak-duration=12h \
+//	    -timeout=13h
+//
+// The test is gated by -olcrtc.local-soak so it never runs in regular CI.
+
+var (
+	localSoakEnabled = flag.Bool( //nolint:gochecknoglobals // package-level state intentional
+		"olcrtc.local-soak",
+		false,
+		"run TestLocalThroughputSoak (long-running local throughput pump)",
+	)
+	localSoakDuration = flag.Duration( //nolint:gochecknoglobals // package-level state intentional
+		"olcrtc.local-soak-duration",
+		30*time.Second,
+		"how long to keep pumping traffic (e.g. 12h, 30m, 90s)",
+	)
+	localSoakTransport = flag.String( //nolint:gochecknoglobals // package-level state intentional
+		"olcrtc.local-soak-transport",
+		transportData,
+		"transport to pump through: datachannel|videochannel|seichannel|vp8channel",
+	)
+	localSoakChunk = flag.Int( //nolint:gochecknoglobals // package-level state intentional
+		"olcrtc.local-soak-chunk",
+		64*1024,
+		"write/read chunk size in bytes",
+	)
+	localSoakProgress = flag.Duration( //nolint:gochecknoglobals // package-level state intentional
+		"olcrtc.local-soak-progress",
+		30*time.Second,
+		"how often to log throughput progress lines",
+	)
+	localSoakVerify = flag.Bool( //nolint:gochecknoglobals // package-level state intentional
+		"olcrtc.local-soak-verify",
+		true,
+		"verify echoed bytes match the sent pattern (slower, but catches corruption)",
+	)
+)
+
+var errLocalSoakPayloadMismatch = errors.New("local soak payload mismatch")
+
+// TestLocalThroughputSoak pumps a deterministic byte pattern through a
+// locally-built tunnel for -olcrtc.local-soak-duration and reports
+// throughput periodically. Both writer and reader run concurrently on the
+// same SOCKS connection; with the loopback echo server on the far end
+// each byte gets written, tunneled across, echoed back, and verified.
+func TestLocalThroughputSoak(t *testing.T) {
+	if !*localSoakEnabled {
+		t.Skip("local soak disabled; pass -olcrtc.local-soak to enable")
+	}
+	if *localSoakDuration <= 0 {
+		t.Skip("local soak duration is zero")
+	}
+	if *localSoakChunk <= 0 {
+		t.Fatalf("invalid -olcrtc.local-soak-chunk=%d", *localSoakChunk)
+	}
+
+	// Connection setup itself can be slow (first WebRTC handshake on
+	// some transports), so don't fold it into the duration budget.
+	const setupBudget = 30 * time.Second
+
+	t.Logf("[soak] transport=%s duration=%s chunk=%d verify=%t progress=%s",
+		*localSoakTransport, *localSoakDuration, *localSoakChunk,
+		*localSoakVerify, *localSoakProgress)
+
+	rt := startLocalSoakTunnel(t, *localSoakTransport)
+	echoAddr := startEchoServer(t)
+
+	conn, err := connectViaSOCKSWithin(rt.socksAddr, echoAddr, setupBudget)
+	if err != nil {
+		t.Fatalf("connect via SOCKS: %v", err)
+	}
+	defer func() { _ = conn.Close() }()
+
+	pumpCtx, cancelPump := context.WithTimeout(context.Background(), *localSoakDuration)
+	defer cancelPump()
+
+	stats := runLocalSoakPump(t, pumpCtx, conn, *localSoakChunk, *localSoakVerify, *localSoakProgress)
+
+	if stats.sent == 0 || stats.recv == 0 {
+		t.Fatalf("no traffic moved: sent=%d recv=%d", stats.sent, stats.recv)
+	}
+	if stats.err != nil && !isExpectedShutdownErr(stats.err) {
+		t.Fatalf("pump error: %v", stats.err)
+	}
+
+	t.Logf("[soak] DONE transport=%s elapsed=%s sent=%s recv=%s send=%s/s recv=%s/s",
+		*localSoakTransport,
+		stats.elapsed.Round(time.Second),
+		humanBytes(stats.sent),
+		humanBytes(stats.recv),
+		humanBytes(int64(float64(stats.sent)/stats.elapsed.Seconds())),
+		humanBytes(int64(float64(stats.recv)/stats.elapsed.Seconds())),
+	)
+}
+
+// startLocalSoakTunnel mirrors startTunnel but lets the caller pick the
+// transport (the original is hard-coded to datachannel).
+func startLocalSoakTunnel(t *testing.T, transportName string) *tunnelRuntime {
+	t.Helper()
+
+	carrierName, room := registerMemoryCarrier(t)
+	ctx, cancel := context.WithCancel(context.Background())
+	t.Cleanup(cancel)
+	socksAddr := freeLocalAddr(ctx, t)
+	options := e2eTransportOptions(transportName)
+
+	serverErr := make(chan error, 1)
+	go func() {
+		serverErr <- server.Run(ctx, server.Config{
+			Transport:        transportName,
+			TransportOptions: options,
+			Carrier:          carrierName,
+			RoomURL:          testRoom,
+			KeyHex:           testKeyHex,
+			DNSServer:        localDNSServer,
+		})
+	}()
+	room.waitConnected(t, 1)
+
+	ready := make(chan struct{})
+	clientErr := make(chan error, 1)
+	go func() {
+		clientErr <- client.RunWithReady(ctx, client.Config{
+			Transport:        transportName,
+			TransportOptions: options,
+			Carrier:          carrierName,
+			RoomURL:          testRoom,
+			KeyHex:           testKeyHex,
+			DeviceID:         testClientDeviceID,
+			LocalAddr:        socksAddr,
+			DNSServer:        localDNSServer,
+		}, func() { close(ready) })
+	}()
+	waitForReady(t, ready)
+
+	return &tunnelRuntime{
+		socksAddr: socksAddr,
+		room:      room,
+		cancel:    cancel,
+		serverErr: serverErr,
+		clientErr: clientErr,
+		stopWait:  3 * time.Second,
+	}
+}
+
+type localSoakStats struct {
+	sent, recv int64
+	elapsed    time.Duration
+	err        error
+}
+
+// runLocalSoakPump runs a writer goroutine and a reader goroutine over the
+// same conn until ctx expires, periodically logging progress. Bytes are
+// counted atomically so the progress logger sees a coherent snapshot.
+func runLocalSoakPump(
+	t *testing.T,
+	ctx context.Context,
+	conn net.Conn,
+	chunkSize int,
+	verify bool,
+	progressEvery time.Duration,
+) localSoakStats {
+	t.Helper()
+
+	var sent, recv atomic.Int64
+	start := time.Now()
+
+	progressDone := make(chan struct{})
+	go func() {
+		defer close(progressDone)
+		if progressEvery <= 0 {
+			return
+		}
+		ticker := time.NewTicker(progressEvery)
+		defer ticker.Stop()
+		var lastSent, lastRecv int64
+		lastTime := start
+		for {
+			select {
+			case <-ctx.Done():
+				return
+			case now := <-ticker.C:
+				s, r := sent.Load(), recv.Load()
+				dt := now.Sub(lastTime).Seconds()
+				instSendRate := int64(float64(s-lastSent) / dt)
+				instRecvRate := int64(float64(r-lastRecv) / dt)
+				t.Logf("[soak] elapsed=%s sent=%s recv=%s tx=%s/s rx=%s/s",
+					now.Sub(start).Round(time.Second),
+					humanBytes(s), humanBytes(r),
+					humanBytes(instSendRate), humanBytes(instRecvRate),
+				)
+				lastSent, lastRecv = s, r
+				lastTime = now
+			}
+		}
+	}()
+
+	var (
+		wg      sync.WaitGroup
+		errOnce sync.Once
+		pumpErr error
+	)
+	recordErr := func(err error) {
+		if err == nil {
+			return
+		}
+		errOnce.Do(func() { pumpErr = err })
+	}
+
+	wg.Add(2)
+
+	// Writer: deterministic pattern by absolute byte offset.
+	go func() {
+		defer wg.Done()
+		buf := make([]byte, chunkSize)
+		var off int64
+		for ctx.Err() == nil {
+			fillPattern(buf, off)
+			if _, err := conn.Write(buf); err != nil {
+				if ctx.Err() == nil {
+					recordErr(fmt.Errorf("write at %d: %w", off, err))
+				}
+				return
+			}
+			off += int64(chunkSize)
+			sent.Add(int64(chunkSize))
+		}
+	}()
+
+	// Reader: io.ReadFull echoed bytes, optionally verify against pattern.
+	go func() {
+		defer wg.Done()
+		rdr := bufio.NewReader(conn)
+		echoed := make([]byte, chunkSize)
+		want := make([]byte, chunkSize)
+		var off int64
+		for ctx.Err() == nil {
+			if _, err := io.ReadFull(rdr, echoed); err != nil {
+				if ctx.Err() == nil {
+					recordErr(fmt.Errorf("read at %d: %w", off, err))
+				}
+				return
+			}
+			if verify {
+				fillPattern(want, off)
+				if !bytes.Equal(echoed, want) {
+					recordErr(fmt.Errorf("%w at offset %d", errLocalSoakPayloadMismatch, off))
+					return
+				}
+			}
+			off += int64(chunkSize)
+			recv.Add(int64(chunkSize))
+		}
+	}()
+
+	<-ctx.Done()
+	// Force-close the conn so both pumps unblock from any in-flight I/O.
+	// SetDeadline-in-the-past is the canonical kick.
+	_ = conn.SetDeadline(time.Unix(1, 0))
+	wg.Wait()
+	<-progressDone
+
+	return localSoakStats{
+		sent:    sent.Load(),
+		recv:    recv.Load(),
+		elapsed: time.Since(start),
+		err:     pumpErr,
+	}
+}
+
+// isExpectedShutdownErr filters errors that just mean "we asked the conn
+// to stop" — deadline expirations from our SetDeadline kick, EOF from the
+// peer half-closing, etc.
+func isExpectedShutdownErr(err error) bool {
+	if err == nil {
+		return true
+	}
+	if errors.Is(err, io.EOF) || errors.Is(err, io.ErrUnexpectedEOF) {
+		return true
+	}
+	if errors.Is(err, context.Canceled) || errors.Is(err, context.DeadlineExceeded) {
+		return true
+	}
+	var nerr net.Error
+	if errors.As(err, &nerr) && nerr.Timeout() {
+		return true
+	}
+	if errors.Is(err, net.ErrClosed) {
+		return true
+	}
+	return false
+}
+
+// humanBytes formats a byte count with a binary-unit suffix.
+func humanBytes(n int64) string {
+	const (
+		kib = 1 << 10
+		mib = 1 << 20
+		gib = 1 << 30
+		tib = 1 << 40
+	)
+	switch {
+	case n >= tib:
+		return fmt.Sprintf("%.2f TiB", float64(n)/float64(tib))
+	case n >= gib:
+		return fmt.Sprintf("%.2f GiB", float64(n)/float64(gib))
+	case n >= mib:
+		return fmt.Sprintf("%.2f MiB", float64(n)/float64(mib))
+	case n >= kib:
+		return fmt.Sprintf("%.2f KiB", float64(n)/float64(kib))
+	default:
+		return fmt.Sprintf("%d B", n)
+	}
+}