feat: added full support in USB (local) & ethernet (local or remote) of audio for Icom

internal/audio/codec.go

@@ -0,0 +1,57 @@
+package audio
+
+import "fmt"
+
+// Codec converts between the wire payload of a network audio stream (Icom 50003)
+// and OpsLog's internal PCM (16 kHz mono 16-bit little-endian — the format the
+// capture/render engine and the pcmRing use). It exists so the transport code
+// (icomaudio.go) never hard-codes a format: today PCM is an identity passthrough;
+// tomorrow an Opus codec implements the same two methods and drops in unchanged.
+// This mirrors how civTransport abstracts the CAT byte stream from its transport.
+type Codec interface {
+	// Name is a short label for logs/UI.
+	Name() string
+	// Decode turns one received audio payload into internal PCM. The returned
+	// slice is freshly allocated (the caller may retain it).
+	Decode(payload []byte) ([]byte, error)
+	// Encode turns internal PCM into a payload to transmit.
+	Encode(pcm []byte) ([]byte, error)
+}
+
+// pcm16Codec is the uncompressed 16-bit-PCM codec — the Icom "uncompressed"
+// audio mode (rxcodec/txcodec in the conninfo). Icom sends little-endian 16-bit
+// mono samples, which is byte-for-byte OpsLog's internal format, so decode/encode
+// are copies. It is the Phase-4/5 default: zero-dependency, lossless, ideal on a
+// LAN. Opus (for WAN/internet bandwidth) becomes another Codec later.
+//
+// NOTE: rate conversion is deliberately NOT this layer's job. The Icom RX audio
+// is 16 kHz = our internal rate, so RX needs none. The rig's TX side may run at a
+// different rate (the captured conninfo showed 8 kHz) — Phase 5 will resample in
+// the TX path before Encode; keeping the codec rate-agnostic keeps that concern
+// in one place.
+type pcm16Codec struct{}
+
+// NewPCM16Codec returns the uncompressed 16-bit PCM codec.
+func NewPCM16Codec() Codec { return pcm16Codec{} }
+
+func (pcm16Codec) Name() string { return "pcm16" }
+
+func (pcm16Codec) Decode(payload []byte) ([]byte, error) {
+	// Icom PCM payloads are whole 16-bit samples; an odd length means a truncated
+	// packet — trim the stray byte rather than emit a half-sample click.
+	n := len(payload) &^ 1
+	if n != len(payload) {
+		if n == 0 {
+			return nil, fmt.Errorf("pcm16: payload too short (%d bytes)", len(payload))
+		}
+	}
+	out := make([]byte, n)
+	copy(out, payload[:n])
+	return out, nil
+}
+
+func (pcm16Codec) Encode(pcm []byte) ([]byte, error) {
+	out := make([]byte, len(pcm))
+	copy(out, pcm)
+	return out, nil
+}

internal/audio/codec_test.go

@@ -0,0 +1,37 @@
+package audio
+
+import "testing"
+
+func TestPCM16CodecRoundTrip(t *testing.T) {
+	c := NewPCM16Codec()
+	in := []byte{0x01, 0x02, 0x03, 0x04, 0xff, 0x7f}
+	enc, err := c.Encode(in)
+	if err != nil {
+		t.Fatalf("encode: %v", err)
+	}
+	dec, err := c.Decode(enc)
+	if err != nil {
+		t.Fatalf("decode: %v", err)
+	}
+	if string(dec) != string(in) {
+		t.Fatalf("round-trip mismatch: got % X want % X", dec, in)
+	}
+}
+
+func TestPCM16CodecTrimsOddByte(t *testing.T) {
+	c := NewPCM16Codec()
+	dec, err := c.Decode([]byte{0x10, 0x20, 0x30}) // 3 bytes = 1 sample + stray
+	if err != nil {
+		t.Fatalf("decode: %v", err)
+	}
+	if len(dec) != 2 {
+		t.Fatalf("expected the stray byte trimmed to 2, got %d", len(dec))
+	}
+}
+
+func TestPCM16CodecRejectsSingleByte(t *testing.T) {
+	c := NewPCM16Codec()
+	if _, err := c.Decode([]byte{0x10}); err == nil {
+		t.Fatalf("expected an error for a sub-sample payload")
+	}
+}

internal/audio/engine.go

@@ -5,6 +5,7 @@ package audio
 import (
 	"fmt"
 	"runtime"
+	"sync"
 	"time"
 	"unsafe"
 
@@ -269,3 +270,151 @@ func playPCM(deviceID string, pcm []byte, rate, ch, bits int, stop <-chan struct
 		time.Sleep(10 * time.Millisecond)
 	}
 }
+
+// pcmRing is a thread-safe, latency-bounded FIFO of PCM bytes feeding a live
+// render stream. Producers (a USB-codec capture, or a decoded network audio
+// stream) Push freshly-arrived samples; the render loop Pulls. It is the shared
+// hand-off point between "where the audio comes from" (USB device / UDP 50003)
+// and "where it's heard" (any WASAPI output) — so the transport can be swapped
+// without touching the render side, mirroring the civTransport split on the CAT
+// side. On overflow the oldest audio is dropped to keep latency bounded; on
+// underrun Pull simply returns short and the render loop pads with silence.
+type pcmRing struct {
+	mu  sync.Mutex
+	buf []byte
+	max int // hard cap in bytes (drops oldest beyond this → bounded latency)
+}
+
+// newPCMRing makes a ring whose backlog is capped at maxBytes. Size it from the
+// acceptable latency: bytesPerSec (=32000) worth ≈ 1 s.
+func newPCMRing(maxBytes int) *pcmRing {
+	if maxBytes <= 0 {
+		maxBytes = bytesPerSec // 1 s default
+	}
+	return &pcmRing{max: maxBytes}
+}
+
+// Push appends samples, dropping the oldest audio if the backlog would exceed
+// the cap (a slow/absent consumer never makes the producer block or grow without
+// bound). A short glitch beats runaway latency for live monitoring.
+func (r *pcmRing) Push(p []byte) {
+	if len(p) == 0 {
+		return
+	}
+	r.mu.Lock()
+	r.buf = append(r.buf, p...)
+	if len(r.buf) > r.max {
+		drop := len(r.buf) - r.max
+		r.buf = append(r.buf[:0], r.buf[drop:]...)
+	}
+	r.mu.Unlock()
+}
+
+// pull removes and returns up to maxBytes of queued PCM (a private copy), or nil
+// when empty. The render loop pads any shortfall with silence.
+func (r *pcmRing) pull(maxBytes int) []byte {
+	r.mu.Lock()
+	defer r.mu.Unlock()
+	if len(r.buf) == 0 || maxBytes <= 0 {
+		return nil
+	}
+	n := maxBytes
+	if n > len(r.buf) {
+		n = len(r.buf)
+	}
+	out := make([]byte, n)
+	copy(out, r.buf[:n])
+	r.buf = append(r.buf[:0], r.buf[n:]...)
+	return out
+}
+
+// renderStream continuously renders PCM pulled from src to a device until stop
+// closes — the streaming counterpart to playPCM's fixed buffer. On underrun it
+// writes silence rather than glitching, keeping the WASAPI clock steady so live
+// monitor audio flows smoothly even when the source stalls briefly. Runs on a
+// COM-initialised, OS-locked thread.
+func renderStream(deviceID string, rate, ch, bits int, stop <-chan struct{}, src *pcmRing) error {
+	runtime.LockOSThread()
+	defer runtime.UnlockOSThread()
+	if err := coInit(); err != nil {
+		return fmt.Errorf("CoInitialize: %w", err)
+	}
+	defer ole.CoUninitialize()
+
+	dev, err := openDevice(wca.ERender, deviceID)
+	if err != nil {
+		return err
+	}
+	defer dev.Release()
+
+	var ac *wca.IAudioClient
+	if err := dev.Activate(wca.IID_IAudioClient, wca.CLSCTX_ALL, nil, &ac); err != nil {
+		return fmt.Errorf("activate render: %w", err)
+	}
+	defer ac.Release()
+
+	frameBytes := ch * bits / 8
+	if frameBytes <= 0 {
+		return fmt.Errorf("bad audio format")
+	}
+	wfx := &wca.WAVEFORMATEX{
+		WFormatTag: 1, NChannels: uint16(ch), NSamplesPerSec: uint32(rate),
+		NAvgBytesPerSec: uint32(rate * frameBytes), NBlockAlign: uint16(frameBytes),
+		WBitsPerSample: uint16(bits), CbSize: 0,
+	}
+	if err := ac.Initialize(wca.AUDCLNT_SHAREMODE_SHARED, autoConvert,
+		wca.REFERENCE_TIME(bufferDuration100ns), 0, wfx, nil); err != nil {
+		return fmt.Errorf("initialize render: %w", err)
+	}
+	var bufFrames uint32
+	if err := ac.GetBufferSize(&bufFrames); err != nil {
+		return err
+	}
+	var arc *wca.IAudioRenderClient
+	if err := ac.GetService(wca.IID_IAudioRenderClient, &arc); err != nil {
+		return fmt.Errorf("get render service: %w", err)
+	}
+	defer arc.Release()
+
+	// feed fills up to `frames` render frames: as much real audio as the ring
+	// has, the remainder silence (so the buffer stays full and the clock steady).
+	feed := func(frames int) error {
+		if frames <= 0 {
+			return nil
+		}
+		var data *byte
+		if err := arc.GetBuffer(uint32(frames), &data); err != nil {
+			return err
+		}
+		dst := unsafe.Slice(data, frames*frameBytes)
+		got := src.pull(frames * frameBytes)
+		n := copy(dst, got)
+		for i := n; i < len(dst); i++ {
+			dst[i] = 0 // silence-fill the shortfall
+		}
+		arc.ReleaseBuffer(uint32(frames), 0)
+		return nil
+	}
+
+	if err := feed(int(bufFrames)); err != nil { // pre-fill to avoid a start glitch
+		return err
+	}
+	if err := ac.Start(); err != nil {
+		return fmt.Errorf("start render: %w", err)
+	}
+	defer ac.Stop()
+
+	for {
+		select {
+		case <-stop:
+			return nil
+		default:
+		}
+		var padding uint32
+		ac.GetCurrentPadding(&padding)
+		if err := feed(int(bufFrames - padding)); err != nil {
+			return err
+		}
+		time.Sleep(8 * time.Millisecond)
+	}
+}

internal/audio/manager.go

@@ -15,7 +15,10 @@ type Manager struct {
 	recStop  chan struct{}
 	recDone  chan recResult
 	playStop chan struct{}
-	onChange func() // fired on any record/playback state transition
+	monStop  chan struct{} // RX monitor passthrough (capture → render)
+	monRing  *pcmRing      // live audio hand-off, also fed by the network stream
+	txStop   chan struct{} // TX audio passthrough (mic → rig)
+	onChange func()        // fired on any record/playback state transition
 }
 
 type recResult struct {
@@ -135,3 +138,130 @@ func (m *Manager) StopPlayback() {
 		m.notify()
 	}
 }
+
+// ---- RX audio monitor (Phase 2: USB codec passthrough) --------------------
+//
+// StartMonitor pipes live RX audio from inputDev (e.g. the rig's "USB Audio
+// CODEC" capture endpoint) to outputDev (your speakers/headset) through a
+// latency-bounded ring, so you HEAR the radio inside OpsLog. The very same ring
+// is later fed by the network 50003 stream instead of a USB capture — the render
+// half is transport-agnostic. inputDev "" = system default capture.
+func (m *Manager) StartMonitor(inputDev, outputDev string) error {
+	return m.startMonitor(inputDev, outputDev, true)
+}
+
+// StartMonitorSink starts ONLY the render side (no USB capture) so an external
+// producer — the network 50003 stream — can feed decoded RX PCM via
+// PushMonitorAudio. Same output path as StartMonitor, minus the capture goroutine.
+func (m *Manager) StartMonitorSink(outputDev string) error {
+	return m.startMonitor("", outputDev, false)
+}
+
+// startMonitor wires the RX monitor: always a render loop pulling from monRing;
+// when capture is true it also captures inputDev into that ring (USB monitor).
+// When false the ring is fed only by PushMonitorAudio (network audio).
+func (m *Manager) startMonitor(inputDev, outputDev string, capture bool) error {
+	m.mu.Lock()
+	if m.monStop != nil {
+		m.mu.Unlock()
+		return fmt.Errorf("monitor already running")
+	}
+	stop := make(chan struct{})
+	ring := newPCMRing(bytesPerSec / 2) // ~500 ms cap — low latency for live monitor
+	m.monStop, m.monRing = stop, ring
+	m.mu.Unlock()
+
+	if capture {
+		// Producer: capture the rig's USB audio into the ring.
+		go func() {
+			_ = captureStream(inputDev, stop, func(chunk []byte) { ring.Push(chunk) })
+		}()
+	}
+	// Consumer: render the ring to the output device at the internal 16 kHz mono.
+	go func() {
+		_ = renderStream(outputDev, sampleRate, channels, bitsPerSample, stop, ring)
+	}()
+	m.notify()
+	return nil
+}
+
+// StopMonitor stops the RX monitor passthrough.
+func (m *Manager) StopMonitor() {
+	m.mu.Lock()
+	stop := m.monStop
+	m.monStop, m.monRing = nil, nil
+	m.mu.Unlock()
+	if stop != nil {
+		close(stop)
+		m.notify()
+	}
+}
+
+// MonitorActive reports whether the RX monitor passthrough is running.
+func (m *Manager) MonitorActive() bool {
+	m.mu.Lock()
+	defer m.mu.Unlock()
+	return m.monStop != nil
+}
+
+// PushMonitorAudio feeds externally-sourced PCM (16 kHz mono 16-bit) into the
+// active monitor's output — the hook the network 50003 audio stream uses to play
+// decoded RX through the very same output path a USB capture feeds. No-op when no
+// monitor is running. Keeps the unexported ring inside the package.
+func (m *Manager) PushMonitorAudio(pcm []byte) {
+	m.mu.Lock()
+	ring := m.monRing
+	m.mu.Unlock()
+	if ring != nil {
+		ring.Push(pcm)
+	}
+}
+
+// ---- TX audio passthrough (Phase 3: live mic → rig over USB) --------------
+//
+// StartTXAudio pipes your live microphone (micDev) into the rig's audio input
+// (toRadioDev — for a USB-connected rig, its "USB Audio CODEC" render endpoint),
+// so you talk through the PC. It is the mirror of StartMonitor (same ring +
+// capture + render primitives, source/sink swapped). PTT keying is the caller's
+// job (the app layer keys PTT before this and unkeys after) so this stays a pure
+// audio route. The captured 16 kHz mono stream is also the exact shape the future
+// network 50003 TX will encode and send — so Phase 5 reuses this capture side.
+func (m *Manager) StartTXAudio(micDev, toRadioDev string) error {
+	m.mu.Lock()
+	if m.txStop != nil {
+		m.mu.Unlock()
+		return fmt.Errorf("TX audio already running")
+	}
+	stop := make(chan struct{})
+	ring := newPCMRing(bytesPerSec / 4) // ~250 ms — tighter for live TX latency
+	m.txStop = stop
+	m.mu.Unlock()
+
+	go func() {
+		_ = captureStream(micDev, stop, func(chunk []byte) { ring.Push(chunk) })
+	}()
+	go func() {
+		_ = renderStream(toRadioDev, sampleRate, channels, bitsPerSample, stop, ring)
+	}()
+	m.notify()
+	return nil
+}
+
+// StopTXAudio stops the TX mic→rig passthrough.
+func (m *Manager) StopTXAudio() {
+	m.mu.Lock()
+	stop := m.txStop
+	m.txStop = nil
+	m.mu.Unlock()
+	if stop != nil {
+		close(stop)
+		m.notify()
+	}
+}
+
+// TXAudioActive reports whether the TX mic→rig passthrough is running.
+func (m *Manager) TXAudioActive() bool {
+	m.mu.Lock()
+	defer m.mu.Unlock()
+	return m.txStop != nil
+}

internal/cat/civ/civ.go

@@ -313,6 +313,8 @@ func ModelName(addr byte) string {
 		return "IC-7300"
 	case 0x98:
 		return "IC-7610"
+	case 0x7C:
+		return "IC-9100"
 	case 0xA2:
 		return "IC-9700"
 	case 0xA4:

internal/cat/flex.go

@@ -803,24 +803,44 @@ func (f *Flex) ReadState() (RigState, error) {
 // band) never hijacks the main frequency. Returns (-1, nil) when no slice is in
 // use. Caller holds f.mu.
 func (f *Flex) mainSliceLocked() (int, *flexSlice) {
-	best, bestS := 1<<30, (*flexSlice)(nil)
-	for idx, s := range f.slices {
+	// Iterate in ASCENDING index order — NEVER map-iteration order, which Go
+	// randomises. When two slices transiently BOTH report active=1 (e.g. an
+	// external controller like DXHunter activates a slice on another band while
+	// ours still holds active, before SmartSDR sends active=0 to the old one),
+	// map order returned a RANDOM active slice each call → the operating frequency
+	// flip-flopped 40m/20m every poll and the Ultrabeam motors chased it forever.
+	// Deterministic order = the lowest-indexed active slice wins, stably.
+	firstInUse := -1
+	for _, idx := range f.sortedSliceIdxLocked() {
+		s := f.slices[idx]
 		if !s.inUse {
 			continue
 		}
+		if firstInUse < 0 {
+			firstInUse = idx
+		}
 		if s.active {
 			return idx, s
 		}
-		if idx < best {
-			best, bestS = idx, s
-		}
 	}
-	if bestS != nil {
-		return best, bestS
+	if firstInUse >= 0 {
+		return firstInUse, f.slices[firstInUse]
 	}
 	return -1, nil
 }
 
+// sortedSliceIdxLocked returns the slice indices in ascending order so every
+// slice-selection helper is deterministic (map iteration is randomised). Caller
+// holds f.mu.
+func (f *Flex) sortedSliceIdxLocked() []int {
+	idxs := make([]int, 0, len(f.slices))
+	for idx := range f.slices {
+		idxs = append(idxs, idx)
+	}
+	sort.Ints(idxs)
+	return idxs
+}
+
 // activeSliceIndexLocked returns the slice index to send commands to (the main
 // slice, else 0). Caller holds f.mu.
 func (f *Flex) activeSliceIndexLocked() int {
@@ -841,8 +861,8 @@ func sliceLetter(idx int) string {
 // txSliceLocked returns the slice flagged as the transmitter (tx=1), or nil.
 // Caller holds f.mu.
 func (f *Flex) txSliceLocked() *flexSlice {
-	for _, s := range f.slices {
-		if s.inUse && s.tx {
+	for _, idx := range f.sortedSliceIdxLocked() {
+		if s := f.slices[idx]; s.inUse && s.tx {
 			return s
 		}
 	}
@@ -871,7 +891,8 @@ func (f *Flex) operatingLocked() (main, rx, tx *flexSlice) {
 	if main != nil && main != txS && main.freqHz != txS.freqHz && BandFromHz(main.freqHz) == bt {
 		rx = main
 	} else {
-		for _, s := range f.slices {
+		for _, idx := range f.sortedSliceIdxLocked() {
+			s := f.slices[idx]
 			if s.inUse && s != txS && s.freqHz != txS.freqHz && BandFromHz(s.freqHz) == bt {
 				rx = s
 				break
@@ -927,6 +948,15 @@ func (f *Flex) SetFrequency(hz int64) error {
 	f.mu.Lock()
 	idx := f.activeSliceIndexLocked()
 	connected := f.conn != nil
+	// Optimistically update the active slice's cached freq NOW, before the radio
+	// echoes the slice status back. Otherwise ReadState/FlexState keep reporting
+	// the OLD freq for the round-trip: the top display (optimistic liveFreqHz)
+	// jumped to the new band while the slice cache — which the FlexPanel and the
+	// Ultrabeam follow loop read — still showed the old one, so the antenna chased
+	// the stale value. The real echo confirms/corrects this a moment later.
+	if s := f.slices[idx]; s != nil {
+		s.freqHz = hz
+	}
 	f.mu.Unlock()
 	if !connected {
 		return fmt.Errorf("flex: not connected")
@@ -952,6 +982,13 @@ func (f *Flex) SetMode(mode string) error {
 	if fm == "" {
 		return fmt.Errorf("flex: unsupported mode %q", mode)
 	}
+	// Optimistically cache the new mode too (same reasoning as SetFrequency) so the
+	// panel reflects it immediately instead of lagging the radio's echo.
+	f.mu.Lock()
+	if s := f.slices[idx]; s != nil {
+		s.mode = fm
+	}
+	f.mu.Unlock()
 	// "slice s <rx> mode=<m>" — set command per the SmartSDR API.
 	f.send(fmt.Sprintf("slice s %d mode=%s", idx, fm))
 	return nil

internal/cat/icomaudio.go

@@ -0,0 +1,219 @@
+package cat
+
+// icomaudio.go — the NETWORK AUDIO stream (UDP 50003) for the Icom LAN protocol.
+// It is the third stream alongside control (50001) and CI-V (50002): once the
+// control login + conninfo (with rxenable=1) authorize audio, the rig streams RX
+// audio here as data packets. This file dials/handshakes/keeps-alive that socket
+// exactly like the CI-V stream (icomnet.go) — those parts are byte-for-byte the
+// PROVEN transport — and hands each received audio payload to a sink callback
+// (the app decodes it via an audio.Codec and plays it through the RX monitor).
+//
+// Reuses icomnet.go's helpers (icnCtrl, icnHandshake, icnPingReply, icnRecv,
+// icnLocalID, icnLE) and the same seq/retransmit discipline.
+//
+// ⚠️ PAYLOAD OFFSET PENDING ON-RIG VERIFICATION. The stream framing (handshake,
+// ping, idle, retransmit, common 16-byte header) is identical to CI-V and proven.
+// The AUDIO data packet's inner layout — where the PCM starts and the datalen
+// field — is reconstructed from wfview's audio_packet (ident@0x10, datalen@0x12,
+// sendseq@0x14, audio@0x16) but NOT yet confirmed against a real 50003 capture.
+// audioPump logs the first few raw packets (icaDumpFirst) so the offset can be
+// confirmed/corrected on the first on-rig test without a packet capture, the same
+// way the CI-V/scope framing was iterated. Nothing here can destabilize CAT: the
+// audio stream is opt-in and entirely separate from control/CI-V.
+
+import (
+	"net"
+	"sync"
+	"sync/atomic"
+	"time"
+)
+
+// icaAudioOffset is where the PCM payload begins inside an audio data packet
+// (wfview audio_packet: 16-byte common header + ident@0x10 + datalen@0x12 +
+// sendseq@0x14 → audio@0x16). Isolated as a const so a capture-confirmed change
+// is a one-line edit.
+const icaAudioOffset = 0x16
+
+// icaDumpFirst is how many initial audio packets to hex-dump to the debug log for
+// offset verification. After the layout is confirmed on a real rig this can go to
+// 0 (or the const above corrected).
+const icaDumpFirst = 6
+
+// icomAudio is the connected audio stream. RX only for now (Phase 4); TX (Phase
+// 5) will add an encode+send path mirroring icomNet.Write.
+type icomAudio struct {
+	conn         *net.UDPConn
+	aID, aRemote uint32
+
+	sink func([]byte) // receives each raw audio payload (app decodes + plays)
+
+	// Receive-side retransmit (audio is a heavy stream, like the scope): track the
+	// rig's data-packet send seq and ask it to resend gaps, or the rig drops the
+	// session. Same mechanism as icomNet. Owned solely by audioPump → no lock.
+	rxHaveSeq bool
+	rxLastSeq uint16
+	rxMissing map[uint16]int
+
+	dumped int          // packets hex-dumped so far (≤ icaDumpFirst)
+	lastRx atomic.Int64 // UnixNano of last packet (liveness)
+
+	done      chan struct{}
+	closeOnce sync.Once
+}
+
+func (a *icomAudio) markRx() { a.lastRx.Store(time.Now().UnixNano()) }
+
+// Close tears the audio stream down (disconnect a few times; UDP is lossy).
+func (a *icomAudio) Close() {
+	a.closeOnce.Do(func() {
+		close(a.done)
+		for i := 0; i < 3; i++ {
+			_, _ = a.conn.Write(icnCtrl(0x05, 0, a.aID, a.aRemote)) // disconnect
+			time.Sleep(15 * time.Millisecond)
+		}
+		_ = a.conn.Close()
+		debugLog.Printf("icom audio: stream closed")
+	})
+}
+
+// dialIcomAudio opens the audio UDP stream to rig:50003, binding LOCAL :50003
+// (mirroring the civ stream's local :50002). The control conninfo (rxenable=1,
+// audioport=50003) must already have authorized it. sink receives each raw audio
+// payload. cancel aborts a slow dial (Stop/Start).
+func dialIcomAudio(host string, sink func([]byte), cancel <-chan struct{}) (*icomAudio, error) {
+	araddr, err := net.ResolveUDPAddr("udp4", net.JoinHostPort(host, "50003"))
+	if err != nil {
+		return nil, err
+	}
+	conn, err := net.DialUDP("udp4", &net.UDPAddr{Port: 50003}, araddr)
+	if err != nil {
+		debugLog.Printf("icom audio: cannot bind local :50003 (Remote Utility running?): %v", err)
+		return nil, err
+	}
+	aID := icnLocalID(conn)
+	aRemote, err := icnHandshake(conn, aID, cancel)
+	if err != nil {
+		_ = conn.Close()
+		debugLog.Printf("icom audio: handshake FAILED: %v", err)
+		return nil, err
+	}
+	_ = conn.SetReadBuffer(1 << 20)
+	a := &icomAudio{
+		conn: conn, aID: aID, aRemote: aRemote,
+		sink:      sink,
+		rxMissing: make(map[uint16]int),
+		done:      make(chan struct{}),
+	}
+	a.markRx()
+	debugLog.Printf("icom audio: stream up (rig id 0x%08X) — awaiting RX audio", aRemote)
+	go a.audioPump()
+	return a, nil
+}
+
+// audioPump drains the audio socket: replies to pings, sends idle keepalives,
+// requests retransmits for lost packets, and hands each audio payload to sink.
+func (a *icomAudio) audioPump() {
+	buf := make([]byte, 8192)
+	lastIdle := time.Now()
+	lastReq := time.Now()
+	for {
+		select {
+		case <-a.done:
+			return
+		default:
+		}
+		_ = a.conn.SetReadDeadline(time.Now().Add(100 * time.Millisecond))
+		if k, err := a.conn.Read(buf); err == nil && k >= 16 {
+			a.markRx()
+			switch typ := icnLE.Uint16(buf[4:]); {
+			case typ == 0x07: // ping
+				_, _ = a.conn.Write(icnPingReply(buf[:k], a.aID, a.aRemote))
+			case typ == 0x01: // retransmit request from the rig (we send no tracked audio yet)
+			case typ == 0x05: // rig-initiated disconnect
+				debugLog.Printf("icom audio: rig sent DISCONNECT — audio stream dropped by the rig")
+			case typ == 0x00 && k > icaAudioOffset: // audio data packet
+				a.trackRxSeq(icnLE.Uint16(buf[6:]))
+				if a.dumped < icaDumpFirst {
+					a.dumped++
+					debugLog.Printf("icom audio raw #%d: len=%d head=% X", a.dumped, k, buf[:min(icaAudioOffset+8, k)])
+				}
+				if a.sink != nil {
+					payload := append([]byte(nil), buf[icaAudioOffset:k]...)
+					a.sink(payload)
+				}
+			}
+		}
+		if time.Since(lastIdle) > 100*time.Millisecond {
+			_, _ = a.conn.Write(icnCtrl(0x00, 0, a.aID, a.aRemote))
+			lastIdle = time.Now()
+		}
+		if time.Since(lastReq) > 100*time.Millisecond {
+			a.sendRetransmitReq()
+			lastReq = time.Now()
+		}
+	}
+}
+
+// trackRxSeq / sendRetransmitReq mirror icomNet's receive-side retransmit exactly
+// (audio is as loss-sensitive as the scope stream). Duplicated deliberately so
+// the audio stream owns its own seq state with no shared locking.
+func (a *icomAudio) trackRxSeq(seq uint16) {
+	if !a.rxHaveSeq {
+		a.rxHaveSeq = true
+		a.rxLastSeq = seq
+		return
+	}
+	switch d := int16(seq - a.rxLastSeq); {
+	case d == 0:
+	case d < 0:
+		delete(a.rxMissing, seq)
+	case d == 1:
+		a.rxLastSeq = seq
+	case int(d) <= icnMaxMissing:
+		for f := a.rxLastSeq + 1; f != seq; f++ {
+			a.rxMissing[f] = 0
+		}
+		a.rxLastSeq = seq
+	default:
+		a.rxMissing = make(map[uint16]int)
+		a.rxLastSeq = seq
+	}
+}
+
+func (a *icomAudio) sendRetransmitReq() {
+	if len(a.rxMissing) == 0 {
+		return
+	}
+	if len(a.rxMissing) > icnMaxMissing {
+		a.rxMissing = make(map[uint16]int)
+		return
+	}
+	var seqs []uint16
+	for s, cnt := range a.rxMissing {
+		if cnt >= 4 {
+			delete(a.rxMissing, s)
+			continue
+		}
+		a.rxMissing[s] = cnt + 1
+		seqs = append(seqs, s)
+	}
+	switch {
+	case len(seqs) == 0:
+		return
+	case len(seqs) == 1:
+		_, _ = a.conn.Write(icnCtrl(0x01, seqs[0], a.aID, a.aRemote))
+	default:
+		b := make([]byte, 16+4*len(seqs))
+		icnLE.PutUint32(b[0:], uint32(len(b)))
+		icnLE.PutUint16(b[4:], 0x01)
+		icnLE.PutUint32(b[8:], a.aID)
+		icnLE.PutUint32(b[12:], a.aRemote)
+		off := 16
+		for _, s := range seqs {
+			icnLE.PutUint16(b[off:], s)
+			icnLE.PutUint16(b[off+2:], s)
+			off += 4
+		}
+		_, _ = a.conn.Write(b)
+	}
+}

internal/cat/icomnet.go

@@ -36,7 +36,13 @@ var icnBE = binary.BigEndian
 // NewIcomNet builds an (unconnected) Icom backend whose transport is the network
 // stream. host is the rig's IP/hostname; user/pass are the rig's Network User1
 // credentials. Reuses the whole IcomSerial controller — only `open` differs.
-func NewIcomNet(host, user, pass string, civAddr int, digitalDefault string) *IcomSerial {
+//
+// audioSink (optional) enables the network RX audio stream (UDP 50003): when
+// non-nil the conninfo asks the rig to stream audio and each received payload is
+// passed to audioSink (the app decodes it via an audio.Codec and plays it). nil
+// = CI-V only (the proven default). The audio stream is fully separate from CAT,
+// so enabling it can't affect freq/mode/DSP control.
+func NewIcomNet(host, user, pass string, civAddr int, digitalDefault string, audioSink func([]byte)) *IcomSerial {
 	if civAddr <= 0 || civAddr > 0xFF {
 		civAddr = 0x98 // IC-7610
 	}
@@ -57,7 +63,7 @@ func NewIcomNet(host, user, pass string, civAddr int, digitalDefault string) *Ic
 		b.dialMu.Lock()
 		cancel := b.dialCancel
 		b.dialMu.Unlock()
-		return dialIcomNet(host, user, pass, "OpsLog", b.rigAddr, cancel)
+		return dialIcomNet(host, user, pass, "OpsLog", b.rigAddr, cancel, audioSink)
 	}
 	return b
 }
@@ -130,6 +136,10 @@ type icomNet struct {
 	// but link fine" (stay connected) from "link dead" (reconnect). See Alive().
 	lastRx atomic.Int64
 
+	// audio is the optional RX audio stream (UDP 50003). nil when audio is off.
+	// Torn down alongside the CI-V/control streams in Close.
+	audio *icomAudio
+
 	done      chan struct{}
 	closeOnce sync.Once
 }
@@ -232,6 +242,9 @@ var icnTrace = false
 func (n *icomNet) Close() error {
 	n.closeOnce.Do(func() {
 		close(n.done)
+		if n.audio != nil {
+			n.audio.Close()
+		}
 		// Tell the rig we're leaving so it frees its SINGLE control session at
 		// once. If it never gets a disconnect it holds the session for minutes and
 		// refuses every new login — which is why a lost link (or a hard app exit)
@@ -475,8 +488,9 @@ func (n *icomNet) resend(seq uint16) {
 
 // ------------------------- connect -------------------------
 
-func dialIcomNet(host, user, pass, compName string, rigAddr byte, cancel <-chan struct{}) (*icomNet, error) {
-	debugLog.Printf("icom net: connecting to %s (user %q, comp %q, rig addr 0x%02X)", host, user, compName, rigAddr)
+func dialIcomNet(host, user, pass, compName string, rigAddr byte, cancel <-chan struct{}, audioSink func([]byte)) (*icomNet, error) {
+	wantAudio := audioSink != nil
+	debugLog.Printf("icom net: connecting to %s (user %q, comp %q, rig addr 0x%02X, audio=%v)", host, user, compName, rigAddr, wantAudio)
 	// ---- control stream (50001): handshake → login → token → conninfo ----
 	craddr, err := net.ResolveUDPAddr("udp4", net.JoinHostPort(host, "50001"))
 	if err != nil {
@@ -562,7 +576,11 @@ func dialIcomNet(host, user, pass, compName string, rigAddr byte, cancel <-chan
 		rigMAC = make([]byte, 6)
 	}
 
-	_, _ = ctrl.Write(icnConnInfo(cTracked, cInner, tokReq, cID, cRemote, token, user, rigMAC, 50002, 50003))
+	var rxEnable byte
+	if wantAudio {
+		rxEnable = 0x01 // ask the rig to stream RX audio on 50003
+	}
+	_, _ = ctrl.Write(icnConnInfo(cTracked, cInner, tokReq, cID, cRemote, token, user, rigMAC, 50002, 50003, rxEnable))
 	cTracked++
 	cInner++
 	drainEnd := time.Now().Add(500 * time.Millisecond)
@@ -634,6 +652,18 @@ func dialIcomNet(host, user, pass, compName string, rigAddr byte, cancel <-chan
 
 	go n.ctrlPump()
 	go n.civPump()
+
+	// Optional RX audio stream (50003). The rig was told (conninfo rxEnable=1) to
+	// stream audio; open the socket + handshake now. A failure here is NON-fatal:
+	// CAT works without audio, so we log and continue rather than tear down a
+	// perfectly good control/CI-V session.
+	if wantAudio {
+		if a, err := dialIcomAudio(host, audioSink, cancel); err != nil {
+			debugLog.Printf("icom net: audio stream FAILED (CAT unaffected): %v", err)
+		} else {
+			n.audio = a
+		}
+	}
 	return n, nil
 }
 
@@ -772,7 +802,7 @@ func icnTokenRenew(seq, innerSeq, tokReq uint16, sentid, rcvdid, token uint32) [
 	return b
 }
 
-func icnConnInfo(seq, innerSeq, tokReq uint16, sentid, rcvdid, token uint32, user string, rigMAC []byte, civPort, audioPort uint16) []byte {
+func icnConnInfo(seq, innerSeq, tokReq uint16, sentid, rcvdid, token uint32, user string, rigMAC []byte, civPort, audioPort uint16, rxEnable byte) []byte {
 	b := make([]byte, 0x90)
 	icnLE.PutUint32(b[0:], 0x90)
 	icnLE.PutUint16(b[6:], seq)
@@ -788,8 +818,8 @@ func icnConnInfo(seq, innerSeq, tokReq uint16, sentid, rcvdid, token uint32, use
 	copy(b[0x2a:0x30], rigMAC)
 	copy(b[0x40:0x60], []byte("IC-7610"))
 	copy(b[0x60:0x70], icnPasscode(user))
-	b[0x70] = 0x00 // rxenable (audio off — CI-V only)
-	b[0x71] = 0x00 // txenable
+	b[0x70] = rxEnable // rxenable: 1 opens the 50003 RX audio stream, 0 = CI-V only
+	b[0x71] = 0x00     // txenable (Phase 5)
 	b[0x72] = 0x10 // rxcodec
 	b[0x73] = 0x04 // txcodec
 	icnBE.PutUint32(b[0x74:], 16000)