chore: release v0.25.8

internal/cat/icomserial.go

@@ -1,6 +1,7 @@
 package cat
 
 import (
+	"errors"
 	"fmt"
 	"strings"
 	"sync"
@@ -535,11 +536,36 @@ func (b *IcomSerial) SetMode(mode string) error {
 	return nil
 }
 
+// errIcomAckLost is "the reply never came" — as opposed to a reply that said no
+// (NG) or a dead port. Only this one is worth repeating: the command may well
+// have been carried out and only its acknowledgement lost. Sentinel rather than
+// a formatted string so callers can tell the two apart.
+var errIcomAckLost = errors.New("icom: timeout waiting for response")
+
+// SetPTT keys or unkeys the transmitter (CI-V 0x1C 0x00), retrying ONCE when the
+// acknowledgement is lost.
+//
+// A missing FB is not a missing command — the rig acts on the frame as soon as it
+// decodes it, and what expires is our wait for the answer on a bus shared with
+// the rig's own transceive updates. JTDX in "Split Operating: Fake It" moves the
+// dial immediately before every key-down, so the PTT ack queues behind that
+// traffic, and one lost ack was fatal: rigctld answered RPRT -9, JTDX read that
+// as losing rig control and tore the connection down mid-over, reopening it a
+// moment later (an operator's log shows exactly that, twice, a new rigctld client
+// within 300 ms of each failure). The same session over TCI never failed, because
+// TCI carries no CI-V and needs no Fake It.
+//
+// Re-sending is safe: asking for a state the rig is already in changes nothing.
 func (b *IcomSerial) SetPTT(on bool) error {
 	state := byte(0)
 	if on {
 		state = 1
 	}
+	err := b.exec(civ.CmdPTT, civ.SubPTT, state)
+	if err == nil || !errors.Is(err, errIcomAckLost) {
+		return err
+	}
+	applog.Printf("icom: PTT %v — no acknowledgement in %s, sending it once more", on, icomCmdTimeout)
 	return b.exec(civ.CmdPTT, civ.SubPTT, state)
 }
 
@@ -619,7 +645,7 @@ func (b *IcomSerial) recv(timeout time.Duration, match func(civ.Decoded) bool) (
 		case <-cancel:
 			return civ.Decoded{}, fmt.Errorf("icom: interrupted")
 		case <-deadline:
-			return civ.Decoded{}, fmt.Errorf("icom: timeout waiting for response")
+			return civ.Decoded{}, errIcomAckLost
 		}
 	}
 }

internal/cluster/cluster.go

@@ -18,6 +18,7 @@ import (
 	"strconv"
 	"strings"
 	"sync"
+	"sync/atomic"
 	"time"
 
 	"hamlog/internal/applog"
@@ -355,6 +356,18 @@ func (s *session) run() {
 // that gets skipped.
 var idleTick = 30 * time.Second
 
+// promptTick is the read deadline used until the login handshake is finished.
+// Short, because a node's "login:" / "password:" carries no newline and is only
+// visible when the read times out — see the read loop. It costs a few wake-ups
+// during the first seconds of a connection and nothing afterwards.
+const promptTick = 700 * time.Millisecond
+
+// handshakeWindow bounds how long the fast promptTick applies. A node that has
+// a password configured but never asks for one would otherwise keep the loop
+// waking every 700 ms for the life of the connection, for a prompt that is never
+// coming. Any real login exchange is over in a second or two.
+const handshakeWindow = 20 * time.Second
+
 // quietNotice is the silence after which the log says so, once.
 const quietNotice = 10 * time.Minute
 
@@ -398,16 +411,37 @@ func (s *session) runOnce() (time.Time, error) {
 
 	// Login: send on first prompt OR blindly after 1.5s. Many DXSpider
 	// nodes accept the callsign without re-prompting.
-	loginSent := false
+	//
+	// Atomic because the blind-login timer below and the read loop both touch
+	// these. loginSent used to be a plain bool that the timer NEVER SET: the
+	// callsign went out and nothing recorded it, so the password branch — gated on
+	// loginSent — was dead code, and the session could only reach "connected" by
+	// recognising a welcome banner. On a node whose entire greeting is a bare
+	// "login:" and which then demands a password (f5mzn.org:9000), that left the
+	// server stuck at "connecting" for ever while telnet logged in by hand fine.
+	var loginSent, pwdSent atomic.Bool
+	// CompareAndSwap, not a plain store: the timer and the loop can reach these at
+	// the same moment, and the callsign must be written exactly once.
+	sendLogin := func() {
+		if s.login != "" && loginSent.CompareAndSwap(false, true) {
+			_, _ = conn.Write([]byte(s.login + "\r\n"))
+		}
+	}
+	// Sent ONCE per connection. A node that re-prompts is refusing the password,
+	// and answering with the same one again only loops — better to let the
+	// refusal show in the console than to hide it behind a retry.
+	sendPassword := func() {
+		if s.cfg.Password != "" && loginSent.Load() && pwdSent.CompareAndSwap(false, true) {
+			_, _ = conn.Write([]byte(s.cfg.Password + "\r\n"))
+		}
+	}
 	if s.login != "" {
 		go func() {
 			select {
 			case <-s.stopCh:
 				return
 			case <-time.After(1500 * time.Millisecond):
-				if !loginSent {
-					_, _ = conn.Write([]byte(s.login + "\r\n"))
-				}
+				sendLogin()
 			}
 		}()
 	}
@@ -453,7 +487,21 @@ func (s *session) runOnce() (time.Time, error) {
 
 	var connectedAt time.Time
 	var quiet time.Duration // how long the node has said nothing
+	var quietNoticed bool   // the long-silence line is said once
 	var pending string      // a line cut in half by a read deadline
+	markConnected := func() {
+		if s.snapshot().State == StateConnected {
+			return
+		}
+		connectedAt = time.Now()
+		s.mu.Lock()
+		s.status.State = StateConnected
+		s.status.ConnectedAt = connectedAt
+		s.status.Error = ""
+		s.mu.Unlock()
+		s.emitStatus()
+		fireInitCommands()
+	}
 	rd := bufio.NewReader(conn)
 	for {
 		select {
@@ -471,7 +519,21 @@ func (s *session) runOnce() (time.Time, error) {
 		// Only a REAL error ends the session. A dead peer still gets caught:
 		// TCP keepalive probes an idle connection and its failure arrives here
 		// as an error, not as a timeout.
-		_ = conn.SetReadDeadline(time.Now().Add(idleTick))
+		// SHORT deadline until the handshake is done, the long idle tick after.
+		//
+		// A cluster writes its prompts WITHOUT a trailing newline, and ReadString
+		// only returns on one — so a prompt is never a "line" at all, it is whatever
+		// is sitting in the buffer when the read deadline expires. At the ordinary
+		// 30 s tick that made a bare "login:" invisible for half a minute and a
+		// following "password:" invisible for another, which is long enough for the
+		// node to give up on us. Only the handshake needs the fast tick; once logged
+		// in, a long deadline is exactly what we want (see idleTick).
+		tick := idleTick
+		if time.Since(linkUpAt) < handshakeWindow &&
+			(!loginSent.Load() || (s.cfg.Password != "" && !pwdSent.Load())) {
+			tick = promptTick
+		}
+		_ = conn.SetReadDeadline(time.Now().Add(tick))
 		chunk, err := rd.ReadString('\n')
 		if err != nil {
 			var ne net.Error
@@ -480,16 +542,48 @@ func (s *session) runOnce() (time.Time, error) {
 				// Keep it: a spot line straddling the deadline would otherwise lose
 				// its first half and arrive as nonsense, or vanish entirely.
 				pending += chunk
-				quiet += idleTick
+				// …but a newline-less PROMPT is not half a line, it is a question,
+				// and this is the only place it can ever be seen. Answer it, show it
+				// in the console (an operator watching a stuck server has a right to
+				// see what the node actually asked), and drop it so it is not glued
+				// onto the front of the next real line.
+				if p := strings.TrimSpace(pending); p != "" {
+					switch {
+					case !loginSent.Load() && s.login != "" && isLoginPrompt(p):
+						s.emitLine(p, false)
+						pending = ""
+						sendLogin()
+						if s.cfg.Password == "" {
+							markConnected()
+						}
+						continue
+					case !pwdSent.Load() && isPasswordPrompt(p):
+						s.emitLine(p, false)
+						pending = ""
+						if s.cfg.Password == "" {
+							// Nothing to answer with. Shown in the console rather than
+							// swallowed: an unanswered "password:" sitting there IS the
+							// explanation for a server that never finishes connecting,
+							// and it is something the operator can act on.
+							continue
+						}
+						sendPassword()
+						markConnected()
+						continue
+					}
+				}
+				quiet += tick
 				// Said once at the first long silence, so a genuinely mute node is
 				// visible without a line every tick.
-				if quiet == quietNotice {
+				if !quietNoticed && quiet >= quietNotice {
+					quietNoticed = true
 					applog.Printf("cluster[%s] no traffic for %s — still connected", s.cfg.Name, quiet)
 				}
 				continue
 			}
 			return connectedAt, fmt.Errorf("read: %w", err)
 		}
+		quietNoticed = false
 		quiet = 0
 		line := pending + chunk
 		pending = ""
@@ -519,28 +613,23 @@ func (s *session) runOnce() (time.Time, error) {
 			}
 		}
 
-		// Login on explicit prompt.
-		if !loginSent && s.login != "" && isLoginPrompt(line) {
-			_, _ = conn.Write([]byte(s.login + "\r\n"))
-			loginSent = true
+		// Login on explicit prompt — the case where the node DID terminate it with
+		// a newline. The newline-less form is handled on the timeout path above.
+		if !loginSent.Load() && s.login != "" && isLoginPrompt(line) {
+			s.emitLine(line, false)
+			sendLogin()
 			continue
 		}
-		// Password on prompt (rare).
-		if loginSent && s.cfg.Password != "" && isPasswordPrompt(line) {
-			_, _ = conn.Write([]byte(s.cfg.Password + "\r\n"))
+		// Password on prompt.
+		if !pwdSent.Load() && isPasswordPrompt(line) {
+			s.emitLine(line, false)
+			sendPassword()
 			continue
 		}
 
 		// Mark connected once we've sent login OR seen a welcome banner.
-		if s.snapshot().State != StateConnected && (loginSent || isWelcome(line)) {
-			connectedAt = time.Now()
-			s.mu.Lock()
-			s.status.State = StateConnected
-			s.status.ConnectedAt = connectedAt
-			s.status.Error = ""
-			s.mu.Unlock()
-			s.emitStatus()
-			fireInitCommands()
+		if loginSent.Load() || isWelcome(line) {
+			markConnected()
 		}
 
 		// EVERY line goes to the console — spot or not. This is the whole point:
@@ -701,6 +790,38 @@ func parseSpot(line string) (Spot, bool) {
 	}, true
 }
 
+// NewLocalSpot builds the Spot for a DX announcement WE just sent, so it lands
+// in the operator's own list at once.
+//
+// A node does not necessarily broadcast a spot back to the station that sent it:
+// DXSpider suppresses the echo to the originator, and a node-side filter can eat
+// it too. So an operator spotted a station, watched their own spot list stay
+// empty, and concluded the spot had never gone out — when it had. This is the
+// spot the echo would have carried, built from what we sent, deliberately the
+// same shape so the UI's call+band de-dupe folds the two into one row on the
+// nodes that DO echo.
+func NewLocalSpot(srv ServerConfig, spotter, dxCall string, freqKHz float64, comment string) Spot {
+	freqHz := int64(freqKHz*1000 + 0.5)
+	now := time.Now()
+	sp := Spot{
+		SourceID:   srv.ID,
+		SourceName: srv.Name,
+		Spotter:    strings.ToUpper(strings.TrimSpace(spotter)),
+		DXCall:     strings.ToUpper(strings.TrimSpace(dxCall)),
+		FreqKHz:    freqKHz,
+		FreqHz:     freqHz,
+		Band:       bandFromHz(freqHz),
+		Comment:    strings.TrimSpace(comment),
+		TimeUTC:    now.UTC().Format("1504") + "Z",
+		ReceivedAt: now,
+	}
+	// Raw reads like the node's own broadcast — it is what anything showing the
+	// source line expects, and it keeps a local spot legible in the log.
+	sp.Raw = fmt.Sprintf("DX de %s: %9.1f  %-12s %-30s %s",
+		sp.Spotter, sp.FreqKHz, sp.DXCall, sp.Comment, sp.TimeUTC)
+	return sp
+}
+
 func isLoginPrompt(s string) bool {
 	low := strings.ToLower(s)
 	return strings.Contains(low, "login:") ||

internal/cluster/login_test.go

@@ -0,0 +1,105 @@
+package cluster
+
+import (
+	"bufio"
+	"net"
+	"strconv"
+	"strings"
+	"testing"
+	"time"
+)
+
+// A node whose whole greeting is a bare "login:" — no newline, no banner — and
+// which then demands a password must still log in.
+//
+// Reported on f5mzn.org:9000: telnet by hand worked, OpsLog sat at "connecting"
+// for ever. Two faults met there. The prompts carry no newline, and the read
+// loop only ever looked at complete LINES, so neither prompt was seen at all;
+// and the blind 1.5 s login never recorded that it had sent the callsign, which
+// left the password branch — gated on that flag — permanently switched off.
+//
+// The test speaks the node's side literally: "login:" with no newline, then
+// "password:" with no newline, then a spot. It asserts both answers arrive and
+// that the session reaches Connected without any welcome banner to lean on.
+func TestBareLoginAndPasswordPromptsWithoutNewlines(t *testing.T) {
+	ln, err := net.Listen("tcp", "127.0.0.1:0")
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer ln.Close()
+
+	host, portStr, _ := net.SplitHostPort(ln.Addr().String())
+	port, err := strconv.Atoi(portStr)
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	type answers struct{ login, pwd string }
+	got := make(chan answers, 1)
+	go func() {
+		c, err := ln.Accept()
+		if err != nil {
+			return
+		}
+		defer c.Close()
+		rd := bufio.NewReader(c)
+		// No newline, exactly as the node sends it.
+		if _, err := c.Write([]byte("login: ")); err != nil {
+			return
+		}
+		call, err := rd.ReadString('\n')
+		if err != nil {
+			return
+		}
+		if _, err := c.Write([]byte("password: ")); err != nil {
+			return
+		}
+		pwd, err := rd.ReadString('\n')
+		if err != nil {
+			return
+		}
+		got <- answers{strings.TrimSpace(call), strings.TrimSpace(pwd)}
+		// Something to prove the link is live and parsing again afterwards.
+		_, _ = c.Write([]byte("DX de F4BPO:     14074.0  OY1CT        FT8            1234Z\r\n"))
+		time.Sleep(time.Second)
+	}()
+
+	spots := make(chan Spot, 4)
+	s := &session{
+		cfg:      ServerConfig{Name: "pwd node", Host: host, Port: port, Password: "s3cret"},
+		login:    "F4BPO",
+		onSpot:   func(sp Spot) { spots <- sp },
+		onLine:   func(Line) {},
+		onStatus: func() {},
+		stopCh:   make(chan struct{}),
+	}
+	done := make(chan error, 1)
+	go func() { _, err := s.runOnce(); done <- err }()
+	defer func() { close(s.stopCh); <-done }()
+
+	select {
+	case a := <-got:
+		if a.login != "F4BPO" {
+			t.Errorf("callsign sent = %q, want F4BPO", a.login)
+		}
+		if a.pwd != "s3cret" {
+			t.Errorf("password sent = %q, want s3cret — the prompt carried no newline", a.pwd)
+		}
+	case <-time.After(10 * time.Second):
+		t.Fatal("the node's newline-less prompts were never answered")
+	}
+
+	select {
+	case sp := <-spots:
+		if sp.DXCall != "OY1CT" {
+			t.Errorf("spot from the wrong station: %+v", sp)
+		}
+	case <-time.After(5 * time.Second):
+		t.Fatal("no spot after the login — the stream is not being parsed")
+	}
+
+	// Connected without a welcome banner: the handshake alone must be enough.
+	if st := s.snapshot().State; st != StateConnected {
+		t.Errorf("state = %q, want %q — the server would still show as connecting", st, StateConnected)
+	}
+}

internal/extsvc/clublog.go

@@ -117,6 +117,13 @@ func UploadClublogADIF(ctx context.Context, client *http.Client, cfg ServiceConf
 	if api == "" {
 		api = clublogAppAPIKey
 	}
+	// putlogs.php reads the upload as an ADIF *file*, so it needs a header.
+	// Callers that already build a full document (the QSL Manager) pass one;
+	// callers that only have <EOR>-terminated records (the on-close flush) do
+	// not, and a headerless file is rejected. Same rule as the LoTW writer.
+	if !strings.Contains(strings.ToUpper(adifDoc), "<EOH>") {
+		adifDoc = "OpsLog Club Log upload\n<PROGRAMID:6>OpsLog <EOH>\n" + adifDoc
+	}
 
 	var buf bytes.Buffer
 	mw := multipart.NewWriter(&buf)

internal/extsvc/eqsl.go

@@ -267,6 +267,78 @@ func UploadEQSL(ctx context.Context, client *http.Client, user, pswd, qthNick, a
 	return UploadResult{OK: false, Message: reason}, fmt.Errorf("eqsl: upload failed: %s", reason)
 }
 
+// eqslBatchMax is the largest number of records eQSL asks a single upload to
+// carry ("upload only files smaller than about 1000 records at a time", eQSL's
+// own ImportADIF interface notes). Callers chunk to this.
+const eqslBatchMax = 1000
+
+// UploadEQSLBatch pushes MANY ADIF records to eQSL.cc in ONE request.
+//
+// ImportADIF.cfm is a file importer, not a per-QSO endpoint: it takes one *or
+// more* QSOs and answers "Result: X out of Y records added" — the plural in its
+// own reply. So an on-close sweep or a bulk upload is one request, not one per
+// contact. No ADIF header is prepended: the single-record path has always posted
+// bare <EOR> records and eQSL accepts them (per ADIF, a file starting with '<'
+// has no header), and there is no reason to change what is known to work.
+//
+// A PARTIAL result ("97 out of 100") sets Ignored so the caller can say so.
+// eQSL does not identify which records it left out, and in practice they are
+// QSOs it already holds — the same duplicate that UploadEQSL reports as success.
+func UploadEQSLBatch(ctx context.Context, client *http.Client, user, pswd, qthNick string, records []string) (UploadResult, error) {
+	user = strings.ToUpper(strings.TrimSpace(user))
+	if user == "" {
+		return UploadResult{}, fmt.Errorf("eqsl: username (callsign) not set")
+	}
+	if strings.TrimSpace(pswd) == "" {
+		return UploadResult{}, fmt.Errorf("eqsl: password not set")
+	}
+	docs := make([]string, 0, len(records))
+	for _, r := range records {
+		if strings.TrimSpace(r) == "" {
+			continue
+		}
+		docs = append(docs, eqslRecordWithNickname(strings.TrimRight(r, "\r\n"), qthNick))
+	}
+	if len(docs) == 0 {
+		return UploadResult{}, fmt.Errorf("eqsl: empty adif batch")
+	}
+	if len(docs) > eqslBatchMax {
+		return UploadResult{}, fmt.Errorf("eqsl: batch of %d exceeds the %d-record limit", len(docs), eqslBatchMax)
+	}
+
+	body, err := eqslPost(ctx, client, user, pswd, strings.Join(docs, "\n"))
+	if err != nil {
+		return UploadResult{OK: false, Message: body}, err
+	}
+	if reason := authErrEQSL(body); reason != "" {
+		return UploadResult{OK: false, Message: reason}, fmt.Errorf("eqsl: %s", reason)
+	}
+
+	// The counted result is read FIRST here, unlike the single-record path: a
+	// batch reply routinely carries both "Result: 97 out of 100 records added"
+	// and a "Bad record: Duplicate" line for the other three, and matching the
+	// duplicate first would throw away the count that says the rest went in.
+	if m := eqslResultRe.FindStringSubmatch(body); m != nil {
+		added, _ := strconv.Atoi(m[1])
+		total, _ := strconv.Atoi(m[2])
+		if added >= 1 {
+			return UploadResult{OK: true, Message: strings.TrimSpace(m[0]), Ignored: added < total}, nil
+		}
+		// "0 out of N" — nothing added. A re-upload of QSOs eQSL already holds
+		// lands here, and that is not a failure.
+		if strings.Contains(strings.ToLower(body), "duplicate") {
+			return UploadResult{OK: true, Message: "already in logbook"}, nil
+		}
+		reason := eqslReason(body)
+		return UploadResult{OK: false, Message: reason}, fmt.Errorf("eqsl: batch upload failed: %s", reason)
+	}
+	if strings.Contains(strings.ToLower(body), "duplicate") {
+		return UploadResult{OK: true, Message: "already in logbook"}, nil
+	}
+	reason := eqslReason(body)
+	return UploadResult{OK: false, Message: reason}, fmt.Errorf("eqsl: batch upload failed: %s", reason)
+}
+
 // eqslReason trims an eQSL reply to a short human-readable reason: the first
 // "Error:" / "Warning:" / "Bad record:" line if present, else the whole body
 // (capped), else a generic phrase.

internal/extsvc/manager.go

@@ -295,9 +295,10 @@ func (m *Manager) CloseUploadCount() int {
 }
 
 // FlushOnClose uploads every QSO due for an on-close push, scanning the whole
-// logbook (not just this session). Called from the shutdown sequence. QRZ/Club
-// Log go one-by-one (fast HTTP); LoTW is signed and uploaded as a single TQSL
-// batch. Returns the number of QSOs uploaded successfully.
+// logbook (not just this session). Called from the shutdown sequence. QRZ and
+// the rest go one-by-one (fast HTTP, no batch API); LoTW is signed and uploaded
+// as a single TQSL batch, and Club Log goes through its batch endpoint.
+// Returns the number of QSOs uploaded successfully.
 func (m *Manager) FlushOnClose() int {
 	if m.deps.CloseUploadIDs == nil {
 		return 0
@@ -312,41 +313,190 @@ func (m *Manager) FlushOnClose() int {
 		switch svc {
 		case ServiceLoTW:
 			uploaded += m.flushLoTWBatch(ids, cfg.LoTW)
-		case ServiceQRZ:
-			for _, id := range ids {
-				if ok, _ := m.upload(svc, id, cfg.QRZ); ok {
-					uploaded++
-				}
-			}
 		case ServiceClublog:
-			for _, id := range ids {
-				if ok, _ := m.upload(svc, id, cfg.Clublog); ok {
-					uploaded++
-				}
-			}
-		case ServiceHRDLog:
-			for _, id := range ids {
-				if ok, _ := m.upload(svc, id, cfg.HRDLog); ok {
-					uploaded++
-				}
-			}
+			uploaded += m.flushClublogBatch(ids, cfg.Clublog)
 		case ServiceEQSL:
-			for _, id := range ids {
-				if ok, _ := m.upload(svc, id, cfg.EQSL); ok {
-					uploaded++
-				}
-			}
+			uploaded += m.flushEQSLBatch(ids, cfg.EQSL)
+		case ServiceQRZ:
+			uploaded += m.flushOneByOne(svc, ids, cfg.QRZ)
+		case ServiceHRDLog:
+			uploaded += m.flushOneByOne(svc, ids, cfg.HRDLog)
 		case ServiceCloudlog:
-			for _, id := range ids {
-				if ok, _ := m.upload(svc, id, cfg.Cloudlog); ok {
-					uploaded++
-				}
-			}
+			uploaded += m.flushOneByOne(svc, ids, cfg.Cloudlog)
 		}
 	}
 	return uploaded
 }
 
+// uploadPace is the shortest gap between two consecutive single-QSO uploads in
+// an on-close sweep. QRZ, HRDLog and Cloudlog have no batch endpoint — HRDLog's
+// NewEntry.aspx keeps only the first record of a multi-record ADIF — so a sweep
+// of a freshly imported log is unavoidably one request per contact. It does not
+// have to arrive as fast as the link allows, though: that burst is what a
+// service reads as a robot, and what got an operator's IP threatened at Club Log
+// (see flushClublogBatch). The gap costs nothing in practice, since a round trip
+// to any of these already takes longer than it.
+const uploadPace = 200 * time.Millisecond
+
+// flushOneByOne uploads ids one request at a time, paced. For the services that
+// have no batch API; everything else has its own flush<Service>Batch.
+func (m *Manager) flushOneByOne(svc Service, ids []int64, cfg ServiceConfig) int {
+	uploaded := 0
+	for i, id := range ids {
+		if i > 0 {
+			time.Sleep(uploadPace)
+		}
+		if ok, _ := m.upload(svc, id, cfg); ok {
+			uploaded++
+		}
+	}
+	return uploaded
+}
+
+// eqslBatchChunk is how many QSOs go into one ImportADIF.cfm request. eQSL's own
+// limit is ten times this (eqslBatchMax); the smaller chunk keeps one refused
+// record from taking a thousand others down with it, and keeps the form body
+// small enough to be unremarkable.
+const eqslBatchChunk = 100
+
+// flushEQSLBatch uploads the on-close eQSL QSOs through ImportADIF.cfm in
+// batches instead of one request per contact. Same reasoning as
+// flushClublogBatch — eQSL's import endpoint has always taken a whole file, so
+// the one-at-a-time loop was making hundreds of requests it never needed to.
+func (m *Manager) flushEQSLBatch(ids []int64, cfg ServiceConfig) int {
+	uploaded := 0
+	var records []string
+	var kept []int64
+
+	send := func() {
+		if len(records) == 0 {
+			return
+		}
+		// nil client: UploadEQSLBatch then builds one with a 30 s timeout rather
+		// than reusing the 20 s budget of a single realtime QSO.
+		res, err := UploadEQSLBatch(context.Background(), nil, cfg.Username, cfg.Password, cfg.QTHNickname, records)
+		if err != nil || !res.OK {
+			if err == nil {
+				err = errFromResult(res)
+			}
+			m.logf("extsvc: eqsl batch upload (%d QSOs) failed: %v", len(kept), err)
+			if m.deps.NotifyError != nil {
+				m.deps.NotifyError(ServiceEQSL, 0, err)
+			}
+		} else {
+			// res.Ignored means eQSL took the file but left records out. Say the
+			// count out loud: the whole chunk is still marked sent (eQSL never
+			// says WHICH it dropped, and in practice they are QSOs it already
+			// had), so the log line is the only trace of the shortfall.
+			if res.Ignored {
+				m.logf("extsvc: eqsl batch upload PARTIAL (%d QSOs sent) %s", len(kept), res.Message)
+			} else {
+				m.logf("extsvc: eqsl batch upload OK (%d QSOs) %s", len(kept), res.Message)
+			}
+			if m.deps.MarkUploaded != nil {
+				for _, id := range kept {
+					m.deps.MarkUploaded(ServiceEQSL, id, res.LogID)
+				}
+			}
+			uploaded += len(kept)
+		}
+		records = records[:0]
+		kept = kept[:0]
+	}
+
+	for _, id := range ids {
+		if m.deps.ShouldUpload != nil && !m.deps.ShouldUpload(ServiceEQSL, id) {
+			continue
+		}
+		// eQSL keeps the QSO's own station call; the account is identified by the
+		// credentials and the optional QTH nickname — as in upload().
+		rec, ok := m.deps.BuildADIF(id, "")
+		if !ok {
+			continue
+		}
+		records = append(records, rec)
+		kept = append(kept, id)
+		if len(records) >= eqslBatchChunk {
+			send()
+		}
+	}
+	send()
+	return uploaded
+}
+
+// clublogBatchChunk is how many QSOs go into one putlogs.php request. Club Log
+// dedupes server-side, so chunking is not about correctness — it keeps a single
+// malformed record from failing a whole ten-thousand-QSO document, and matches
+// what the QSL Manager's bulk upload already uses.
+const clublogBatchChunk = 100
+
+// flushClublogBatch uploads the on-close Club Log QSOs through the BATCH
+// endpoint (putlogs.php) rather than one realtime.php call each.
+//
+// It used to walk the ids and call UploadClublog per QSO. On-close upload sweeps
+// the WHOLE logbook, so importing an ADIF — or simply switching Club Log on over
+// an existing log — turned one app close into hundreds of realtime.php posts.
+// That endpoint is reserved for an operator logging contacts as they work them,
+// and Club Log blocks the IP of anything that batches through it: an OpsLog user
+// was flagged by G7VJR for 185 QSOs in four minutes, which is this loop, not a
+// pile-up. Batch upload is the mechanism Club Log provides for exactly this.
+func (m *Manager) flushClublogBatch(ids []int64, cfg ServiceConfig) int {
+	uploaded := 0
+	var records []string
+	var kept []int64
+
+	send := func() {
+		if len(records) == 0 {
+			return
+		}
+		// nil client on purpose: UploadClublogADIF then builds one with a 120 s
+		// timeout. m.deps.Client is the 20 s budget of a single realtime QSO,
+		// which a hundred-QSO document on a slow link would blow through.
+		res, err := UploadClublogADIF(context.Background(), nil, cfg, strings.Join(records, "\n"))
+		if err != nil || !res.OK {
+			if err == nil {
+				err = errFromResult(res)
+			}
+			m.logf("extsvc: clublog batch upload (%d QSOs) failed: %v", len(kept), err)
+			if m.deps.NotifyError != nil {
+				m.deps.NotifyError(ServiceClublog, 0, err)
+			}
+		} else {
+			m.logf("extsvc: clublog batch upload OK (%d QSOs) %s", len(kept), res.Message)
+			if m.deps.MarkUploaded != nil {
+				for _, id := range kept {
+					m.deps.MarkUploaded(ServiceClublog, id, res.LogID)
+				}
+			}
+			uploaded += len(kept)
+		}
+		records = records[:0]
+		kept = kept[:0]
+	}
+
+	for _, id := range ids {
+		// Skip QSOs not eligible (already sent). The wrong-logbook guard that
+		// upload() applies per QSO is not repeated here: closeUploadIDs has
+		// already filtered the sweep down to this logbook's callsign.
+		if m.deps.ShouldUpload != nil && !m.deps.ShouldUpload(ServiceClublog, id) {
+			continue
+		}
+		// Club Log takes the logbook callsign as its own form field, so the ADIF
+		// keeps the QSO's own station call (no override) — as in upload().
+		rec, ok := m.deps.BuildADIF(id, "")
+		if !ok {
+			continue
+		}
+		records = append(records, rec)
+		kept = append(kept, id)
+		if len(records) >= clublogBatchChunk {
+			send()
+		}
+	}
+	send()
+	return uploaded
+}
+
 // flushLoTWBatch signs+uploads all queued LoTW QSOs in one TQSL run, then
 // stamps each as uploaded on success.
 func (m *Manager) flushLoTWBatch(ids []int64, cfg ServiceConfig) int {