diff --git a/cmd/proxy/main.go b/cmd/proxy/main.go
index a0b603d..c4bea05 100644
--- a/cmd/proxy/main.go
+++ b/cmd/proxy/main.go
@@ -1880,7 +1880,13 @@ func addFetchErrorForAllPartitions(resp *kmsg.FetchResponse, req *kmsg.FetchRequ
 // encodeProduceRequest serializes a produce request with header into a wire frame.
 func encodeProduceRequest(header *protocol.RequestHeader, req *kmsg.ProduceRequest) []byte {
 	formatter := kmsg.NewRequestFormatter(kmsg.FormatterClientID(clientIDStr(header.ClientID)))
-	return formatter.AppendRequest(nil, req, header.CorrelationID)
+	b := formatter.AppendRequest(nil, req, header.CorrelationID)
+	// AppendRequest emits a 4-byte size prefix, but WriteFrame re-adds one;
+	// strip it here so the wire frame is single-prefixed (header+body only).
+	if len(b) < 4 {
+		return b
+	}
+	return b[4:]
 }
 
 // parseProduceResponse deserializes a produce response from wire bytes.
@@ -1900,7 +1906,13 @@ func parseProduceResponse(data []byte, version int16) (*kmsg.ProduceResponse, er
 // encodeFetchRequest serializes a fetch request with header into a wire frame.
 func encodeFetchRequest(header *protocol.RequestHeader, req *kmsg.FetchRequest) []byte {
 	formatter := kmsg.NewRequestFormatter(kmsg.FormatterClientID(clientIDStr(header.ClientID)))
-	return formatter.AppendRequest(nil, req, header.CorrelationID)
+	b := formatter.AppendRequest(nil, req, header.CorrelationID)
+	// AppendRequest emits a 4-byte size prefix, but WriteFrame re-adds one;
+	// strip it here so the wire frame is single-prefixed (header+body only).
+	if len(b) < 4 {
+		return b
+	}
+	return b[4:]
 }
 
 // parseFetchResponse deserializes a fetch response from wire bytes.
diff --git a/cmd/proxy/main_test.go b/cmd/proxy/main_test.go
index d69ade0..d996375 100644
--- a/cmd/proxy/main_test.go
+++ b/cmd/proxy/main_test.go
@@ -22,6 +22,7 @@ import (
 	"io"
 	"log/slog"
 	"net"
+	"sort"
 	"sync"
 	"testing"
 	"time"
@@ -1180,3 +1181,308 @@ func TestForwardFetchRetriesOnBackendError(t *testing.T) {
 		t.Fatalf("expected >=2 backend attempts (error then retry success), got %d", fb.attempts)
 	}
 }
+
+// assertSinglePrefixFrame drives encoded bytes through the exact wire path
+// forwardToBackend uses: WriteFrame (which prepends a 4-byte big-endian size) ->
+// ReadFrame -> ParseRequest (the broker decode path). It asserts the
+// single-length-prefix invariant explicitly:
+//
+//  1. After WriteFrame, the first 4 bytes equal big-endian(len(rest)), AND
+//  2. rest (the frame payload the broker reads back) equals the encode output
+//     byte-for-byte. Together these prove there is EXACTLY ONE size prefix on the
+//     wire. If the encode funcs kept AppendRequest's own prefix, the wire frame
+//     would be [size][size][header][body]; the outer size would then be 4 larger
+//     than len(encoded) and the round-tripped payload would carry a spurious
+//     leading 4 bytes, so both asserts catch the regression.
+//
+// It returns the parsed header and request for the caller to assert on.
+func assertSinglePrefixFrame(t *testing.T, encoded []byte) (*protocol.RequestHeader, kmsg.Request) {
+	t.Helper()
+
+	var buf bytes.Buffer
+	if err := protocol.WriteFrame(&buf, encoded); err != nil {
+		t.Fatalf("WriteFrame: %v", err)
+	}
+	wire := buf.Bytes()
+	if len(wire) < 4 {
+		t.Fatalf("framed bytes too short: %d", len(wire))
+	}
+
+	// Assert 1: the on-wire size prefix is exactly len(encoded), not len(encoded)+4.
+	gotSize := binary.BigEndian.Uint32(wire[:4])
+	if int(gotSize) != len(encoded) {
+		t.Fatalf("on-wire size prefix=%d, want %d (len of encode output); a value of %d would mean a double prefix",
+			gotSize, len(encoded), len(encoded)+4)
+	}
+	// Assert 2: the framed payload equals the encode output byte-for-byte. This is
+	// the strict single-prefix proof: WriteFrame adds exactly one prefix and the
+	// encode output carries none of its own.
+	rest := wire[4:]
+	if !bytes.Equal(rest, encoded) {
+		t.Fatalf("frame payload (len %d) does not equal encode output (len %d) byte-for-byte; double prefix?",
+			len(rest), len(encoded))
+	}
+
+	frame, err := protocol.ReadFrame(&buf)
+	if err != nil {
+		t.Fatalf("ReadFrame: %v", err)
+	}
+	if !bytes.Equal(frame.Payload, encoded) {
+		t.Fatalf("ReadFrame payload does not match encode output byte-for-byte; double prefix?")
+	}
+	gotHeader, parsedReq, err := protocol.ParseRequest(frame.Payload)
+	if err != nil {
+		t.Fatalf("broker-path ParseRequest failed (double prefix bug?): %v", err)
+	}
+	return gotHeader, parsedReq
+}
+
+// makeFetchRequestTyped builds a fetch v12 request (the last version that carries
+// the topic NAME on the wire; v13+ switches to TopicID UUIDs) with the given
+// topic -> partitions layout, so the name survives a round-trip and makes a
+// meaningful single-prefix assertion.
+func makeFetchRequestTyped(version int16, topics map[string][]int32) *kmsg.FetchRequest {
+	req := kmsg.NewPtrFetchRequest()
+	req.SetVersion(version)
+	req.ReplicaID = -1
+	req.MaxWaitMillis = 500
+	req.MinBytes = 1
+	req.MaxBytes = 1048576
+	req.SessionEpoch = -1
+	// Sort topic names so encode output is deterministic across runs.
+	names := make([]string, 0, len(topics))
+	for name := range topics {
+		names = append(names, name)
+	}
+	sort.Strings(names)
+	for _, name := range names {
+		ft := kmsg.NewFetchRequestTopic()
+		ft.Topic = name
+		for _, p := range topics[name] {
+			fp := kmsg.NewFetchRequestTopicPartition()
+			fp.Partition = p
+			fp.FetchOffset = int64(p) * 10
+			fp.PartitionMaxBytes = 524288
+			ft.Partitions = append(ft.Partitions, fp)
+		}
+		req.Topics = append(req.Topics, ft)
+	}
+	return req
+}
+
+// TestEncodeRequestNoDoubleLengthPrefix is a real round-trip regression for the
+// double-length-prefix bug on the REQUEST re-marshal path:
+// encodeFetchRequest/encodeProduceRequest produce the bytes forwardToBackend
+// writes via protocol.WriteFrame, which prepends a 4-byte size. If the encode
+// funcs also kept AppendRequest's own size prefix, the on-wire frame would be
+// [size][size][header][body]; a broker reading it via protocol.ParseRequest would
+// see the inner size's bytes as apiKey/apiVersion (the observed "decode Produce
+// v###: not enough data"). Every sub-test pushes the encoded bytes through
+// WriteFrame -> ReadFrame -> ParseRequest exactly as the broker would, and asserts
+// a SINGLE length prefix (see assertSinglePrefixFrame) plus that the header and
+// topics survive intact.
+func TestEncodeRequestNoDoubleLengthPrefix(t *testing.T) {
+	clientID := "proxy-test"
+	fetchHeader := func(corr int32) *protocol.RequestHeader {
+		return &protocol.RequestHeader{
+			APIKey:        protocol.APIKeyFetch,
+			APIVersion:    12,
+			CorrelationID: corr,
+			ClientID:      &clientID,
+		}
+	}
+
+	t.Run("fetch_single_partition", func(t *testing.T) {
+		req := makeFetchRequestTyped(12, map[string][]int32{"events.er1_items": {0}})
+		header := fetchHeader(1)
+
+		gotHeader, parsedReq := assertSinglePrefixFrame(t, encodeFetchRequest(header, req))
+
+		if gotHeader.APIKey != protocol.APIKeyFetch || gotHeader.APIVersion != 12 || gotHeader.CorrelationID != 1 {
+			t.Fatalf("header garbled: key=%d ver=%d corr=%d", gotHeader.APIKey, gotHeader.APIVersion, gotHeader.CorrelationID)
+		}
+		fr := parsedReq.(*kmsg.FetchRequest)
+		if len(fr.Topics) != 1 || fr.Topics[0].Topic != "events.er1_items" {
+			t.Fatalf("topics: %+v", fr.Topics)
+		}
+		if len(fr.Topics[0].Partitions) != 1 {
+			t.Fatalf("partitions: want 1 got %d", len(fr.Topics[0].Partitions))
+		}
+	})
+
+	t.Run("fetch_multi_partition", func(t *testing.T) {
+		req := makeFetchRequestTyped(12, map[string][]int32{
+			"events.er1_items":      {0, 1, 2},
+			"citizen.audit.queries": {0, 1, 2},
+		})
+		header := fetchHeader(4242)
+
+		gotHeader, parsedReq := assertSinglePrefixFrame(t, encodeFetchRequest(header, req))
+
+		if gotHeader.APIVersion != 12 {
+			t.Fatalf("apiVersion: want 12 got %d (double prefix garbles version)", gotHeader.APIVersion)
+		}
+		if gotHeader.CorrelationID != 4242 {
+			t.Fatalf("correlationID: want 4242 got %d", gotHeader.CorrelationID)
+		}
+		fr := parsedReq.(*kmsg.FetchRequest)
+		if len(fr.Topics) != 2 {
+			t.Fatalf("topics: want 2 got %d", len(fr.Topics))
+		}
+		for _, ft := range fr.Topics {
+			if len(ft.Partitions) != 3 {
+				t.Fatalf("topic %q partitions: want 3 got %d", ft.Topic, len(ft.Partitions))
+			}
+		}
+	})
+
+	t.Run("fetch_empty_partitions", func(t *testing.T) {
+		// An empty topic / no records still produces a well-formed fetch sub-request
+		// with zero partitions; it must frame singly too. This is the shape a fetch
+		// against an empty topic takes after groupFetchPartitionsByBroker filtering.
+		req := makeFetchRequestTyped(12, map[string][]int32{"events.er1_items": {}})
+		header := fetchHeader(7)
+
+		gotHeader, parsedReq := assertSinglePrefixFrame(t, encodeFetchRequest(header, req))
+
+		if gotHeader.APIVersion != 12 {
+			t.Fatalf("apiVersion: want 12 got %d", gotHeader.APIVersion)
+		}
+		fr := parsedReq.(*kmsg.FetchRequest)
+		if len(fr.Topics) != 1 || len(fr.Topics[0].Partitions) != 0 {
+			t.Fatalf("want 1 topic with 0 partitions, got %+v", fr.Topics)
+		}
+	})
+
+	t.Run("fetch_multi_broker_fanout", func(t *testing.T) {
+		// Reproduce the canUseOriginal == false hot path: groupFetchPartitionsByBroker
+		// splits a fetch into one sub-request per owning broker, each of which is
+		// independently encodeFetchRequest'd and WriteFrame'd. With no router every
+		// partition lands under the round-robin group, so to exercise a genuine
+		// multi-broker split we build the sub-requests the same way the grouper does
+		// (one *kmsg.FetchRequest per broker, settings copied from the parent) and
+		// assert EACH sub-request frames singly and round-trips.
+		full := makeFetchRequestTyped(12, map[string][]int32{
+			"events.er1_items":      {0, 1, 2, 3},
+			"citizen.audit.queries": {0, 1},
+		})
+		// Owner map mimicking three brokers owning disjoint partitions.
+		owner := func(topic string, part int32) string {
+			return []string{"broker-0", "broker-1", "broker-2"}[int(part)%3]
+		}
+		subReqs := splitFetchByOwnerForTest(full, owner)
+		if len(subReqs) != 3 {
+			t.Fatalf("expected 3 broker sub-requests, got %d", len(subReqs))
+		}
+
+		header := fetchHeader(31337)
+		encodings := make([][]byte, 0, len(subReqs))
+		for addr, sub := range subReqs {
+			enc := encodeFetchRequest(header, sub)
+			encodings = append(encodings, enc)
+			gotHeader, parsedReq := assertSinglePrefixFrame(t, enc)
+			if gotHeader.APIVersion != 12 {
+				t.Fatalf("broker %s sub-request apiVersion: want 12 got %d", addr, gotHeader.APIVersion)
+			}
+			fr, ok := parsedReq.(*kmsg.FetchRequest)
+			if !ok {
+				t.Fatalf("broker %s: parsed %T, not *kmsg.FetchRequest", addr, parsedReq)
+			}
+			if len(fr.Topics) == 0 {
+				t.Fatalf("broker %s sub-request has no topics", addr)
+			}
+		}
+
+		// No buffer aliasing on the fan-out hot path: each encodeFetchRequest calls
+		// AppendRequest(nil, ...) (a fresh allocation) and returns b[4:] (a sub-slice
+		// of that fresh buffer), so distinct sub-requests must not share backing
+		// storage. Distinct lengths or distinct backing arrays both prove this; we
+		// assert no two sub-request encodings alias the same backing array.
+		for i := 0; i < len(encodings); i++ {
+			for j := i + 1; j < len(encodings); j++ {
+				if len(encodings[i]) > 0 && len(encodings[j]) > 0 &&
+					&encodings[i][0] == &encodings[j][0] {
+					t.Fatalf("fan-out buffer aliasing: sub-request %d and %d share backing array", i, j)
+				}
+			}
+		}
+	})
+
+	t.Run("produce", func(t *testing.T) {
+		header := &protocol.RequestHeader{
+			APIKey:        protocol.APIKeyProduce,
+			APIVersion:    9,
+			CorrelationID: 99,
+			ClientID:      &clientID,
+		}
+		req := kmsg.NewPtrProduceRequest()
+		req.SetVersion(9)
+		req.TimeoutMillis = 1000
+		for _, topicName := range []string{"events.er1_items", "citizen.audit.queries"} {
+			pt := kmsg.NewProduceRequestTopic()
+			pt.Topic = topicName
+			for p := int32(0); p < 2; p++ {
+				pp := kmsg.NewProduceRequestTopicPartition()
+				pp.Partition = p
+				pp.Records = []byte{0x01, 0x02, 0x03}
+				pt.Partitions = append(pt.Partitions, pp)
+			}
+			req.Topics = append(req.Topics, pt)
+		}
+
+		gotHeader, parsedReq := assertSinglePrefixFrame(t, encodeProduceRequest(header, req))
+
+		if gotHeader.APIKey != protocol.APIKeyProduce {
+			t.Fatalf("apiKey: want %d got %d (double prefix)", protocol.APIKeyProduce, gotHeader.APIKey)
+		}
+		if gotHeader.APIVersion != 9 {
+			t.Fatalf("apiVersion: want 9 got %d (double prefix garbles version)", gotHeader.APIVersion)
+		}
+		pr := parsedReq.(*kmsg.ProduceRequest)
+		if len(pr.Topics) != 2 {
+			t.Fatalf("topics: want 2 got %d", len(pr.Topics))
+		}
+		if pr.Topics[1].Topic != "citizen.audit.queries" {
+			t.Fatalf("topic[1]: want citizen.audit.queries got %q", pr.Topics[1].Topic)
+		}
+	})
+}
+
+// splitFetchByOwnerForTest mirrors how groupFetchPartitionsByBroker builds one
+// *kmsg.FetchRequest per owning broker (settings copied from the parent, topics
+// and partitions distributed by owner). It is owner-map-driven so the test does
+// not need a live etcd-backed PartitionRouter to exercise a multi-broker fan-out.
+func splitFetchByOwnerForTest(req *kmsg.FetchRequest, owner func(topic string, part int32) string) map[string]*kmsg.FetchRequest {
+	groups := make(map[string]*kmsg.FetchRequest)
+	topicIdx := make(map[string]map[string]int)
+	for _, topic := range req.Topics {
+		for _, part := range topic.Partitions {
+			addr := owner(topic.Topic, part.Partition)
+			sub, ok := groups[addr]
+			if !ok {
+				sub = &kmsg.FetchRequest{
+					Version:        req.Version,
+					ReplicaID:      req.ReplicaID,
+					MaxWaitMillis:  req.MaxWaitMillis,
+					MinBytes:       req.MinBytes,
+					MaxBytes:       req.MaxBytes,
+					IsolationLevel: req.IsolationLevel,
+					SessionID:      req.SessionID,
+					SessionEpoch:   req.SessionEpoch,
+				}
+				groups[addr] = sub
+				topicIdx[addr] = make(map[string]int)
+			}
+			idx, ok := topicIdx[addr][topic.Topic]
+			if !ok {
+				idx = len(sub.Topics)
+				st := kmsg.NewFetchRequestTopic()
+				st.Topic = topic.Topic
+				sub.Topics = append(sub.Topics, st)
+				topicIdx[addr][topic.Topic] = idx
+			}
+			sub.Topics[idx].Partitions = append(sub.Topics[idx].Partitions, part)
+		}
+	}
+	return groups
+}