fix: 后台加人人数改为事务内实际插入数

cmd/nixmsg/serve.go

@@ -335,7 +335,9 @@ func runServe(ctx context.Context, cfg config.Config) error {
 	drainCancel()
 
 	shutCtx, shutCancel := context.WithTimeout(context.Background(), 5*time.Second)
-	_ = brk.Shutdown(shutCtx)
+	if shutErr := brk.Shutdown(shutCtx); shutErr != nil && !errors.Is(shutErr, context.DeadlineExceeded) && !errors.Is(shutErr, context.Canceled) {
+		slog.Error("broker shutdown", "err", shutErr)
+	}
 	shutCancel()
 
 	secondDrain := drainBudget - time.Since(drainStart)

docs/DEVIATIONS.md

@@ -1718,3 +1718,83 @@ issue #3 未关闭，`feat/fix-3-downlink-deadlock` 未合入 `main`。下面是
    - 原因：原先 12 项备注写着未穷尽仍标通过，交付说明写成「通过 23」。  
    - 备选方案：为每个未测子项补验收用例（本波不做，避免为变绿放松断言）。  
    - 影响：汇总改为通过 19、部分通过 4（F03/F08/F21/F22）、失败 0。F19 仍引用仓库内 SDK 清单、本波不重跑。
+
+### 复审修复 R3-01
+
+1. **推送 worker 漏唤醒：定时分发回执、超限后续、撤回腾窗**  
+   - 日期：2026-09-30  
+   - 原条款：Gitea #65；推送 worker 只在 `WakePush` 时跑一轮 `PushPending`。  
+   - 实际做法：`dispatchDueBatch` 本条已分发时除 pending 接收方外始终 `WakePush` 发送方；`PushPending` 遇 `too_large` 等跳过且本轮未占满窗口时再 `WakePush` 当前接收方（不在同一次递归扫表）；`Recall` 成功改成 `recalled` 的接收方各 `WakePush` 一次，已推送撤回仍走 `flushRevokes`。不改 `PublishDown` 签名。  
+   - 原因：终态回执、超限后的后续 pending、撤回腾出窗口后都依赖再唤醒，否则空闲在线端收不到。  
+   - 备选方案：在同一次 `PushPending` 里循环扫完整 pending 表（否决，指令要求合并唤醒下一轮）。  
+   - 影响：仅 `internal/app/message/`；相关单测见 `review_r3_65_test.go`。
+
+### 复审修复 R3-03
+
+1. **写队列满时 Close 不再与 enqueue 死锁**  
+   - 日期：2026-09-30  
+   - 原条款：issue #67；停机需 Drain/Close 在超时内返回；已入队任务仍由写 goroutine 处理完。  
+   - 实际做法：Queue 增加 closing 信号。enqueue 在持 sendMu.RLock 时 select 同时等待 q.ch、closing 与 ctx.Done()，通道满时不再无限阻塞。Close 先 Swap(closed) 并 close(closing) 唤醒在途发送方释放读锁，再在写锁内 close(q.ch)，最后等 loop 退出。不向已关闭 channel 发送。  
+   - 原因：通道满时并发 Do 占着读锁堵在发送上，Close 的写锁拿不到，停机卡死；未提交的在途写也会丢。  
+   - 备选方案：入队改为非阻塞，满则立即 ErrBusy（否决：改变背压语义，正常高峰会误伤提交）。  
+   - 影响：仅 internal/store/queue.go 与测试；不改迁移编号。
+
+### 复审修复 R3-04
+
+- 日期：2026-09-30
+- 原条款：Gitea #68；`dispatchSend` 在 `PublishUp` 失败且未停止重连时清 inflight 后静默 return，`Send` 永久挂起。
+- 实际做法：失败且未 `stopReconnect` 时保留 id/body/send_at_ms，经 `regenerateSendLocked` 换新 rid，清本次 inflight 并 `drainSendQueue` 继续泵；已停止重连时仍 `finishSend` 返回错误。
+- 原因：条目已不在途，重连时的 `requeueInflightLocked` 不会捡回，调用方一直等 `result`。
+- 备选方案：失败即 `finishSend` 报错（否决，与断线/限速重交语义不一致）；只换 rid 不唤醒队列（否决，等同 JS #69）。
+- 影响：仅 `sdk/go`；假传输可模拟 send 发布失败一次。
+
+### 复审修复 R3-05
+
+1. **JS SDK 传输失败后须再泵发送队列**  
+   - 日期：2026-09-30  
+   - 原条款：DEVELOPMENT 第 5 节 / 附录 SDK 行为约定：发送在途失败后换新 `rid` 并重交；`id` / 正文 / `send_at_ms` 不变。Gitea #69。  
+   - 实际做法：`dispatchSend` 中 `publishUp` 抛出非协议 `APIError` 时，若未 `stopReconnect` 则清 inflight、换新 `rid` 后立刻 `drainSendQueue`；若已 `stopReconnect` 则 `finishSendErr` 结束本次 `send`。不额外再乘一次抖动。  
+   - 原因：原先只换 `rid` 并 `return`，连接未断时 `hello` 不再走，队列停泵，`send()` Promise 永不结束。  
+   - 备选方案：按 `rate_limited` 同一条的退避再泵（否决本波，连接仍在线时立即重交更贴切，且避免与已有等待叠乘抖动）。  
+   - 影响：仅 `sdk/js`；假传输失败一次后会再次上行且 `rid` 已变。
+
+### 复审修复 R3-06
+
+1. **Python/Java 首次握手失败应停止重连；Java rate_limited 在途计数只减一次**  
+   - 日期：2026-09-30  
+   - 原条款：DEVELOPMENT 第 9 节附录「首次连接超时或握手失败应停止重连，并向 connect() 返回未连接」；issue #70。  
+   - 实际做法：`sdk/python` 的 `connect()` 在 `_handshake_error` 或非 ONLINE 终态时置 `_stop_reconnect`/`_want_connected=False`；内部连接/握手超时改用 `not_connected`（不再用 `busy`）。`sdk/java` 的 `connectSync` 在 `handshakeError` 时同样置 `stopReconnect`；`rate_limited` 路径去掉第二次 `inflightSends--`，只换新 rid 再排队。Python `transport.py` 将 paho 改为惰性导入，便于本机无 paho 时仍跑 FakeTransport 单测。不改 Go/JS 重连公式。  
+   - 原因：原先抛错后工作线程仍按退避重连；Java 限速把在途计数减了两次。  
+   - 备选方案：在后台 `_attempt_connect`/`attemptConnect` 内首次失败即停（否决：应用再次 `connect()` 才应恢复，标志应在对外 `connect` 失败路径统一置位）。  
+   - 影响：仅 `sdk/python`、`sdk/java`；应用需再次调用 `connect()` 才会重连。
+
+### 复审修复 R3-07
+
+1. **客户端建群事务内复核创建者**  
+   - 日期：2026-09-30  
+   - 原条款：Gitea #71。`group.Create` 在 `Queue.Do` 里直接 `insertMemberTx` 群主，不跑 `endpointCheckTx`；停用/删除与建群抢在事务外对话密码窗口时，可插入已停用甚至刚删除的 `owner_id`。  
+   - 实际做法：`Create` 同一写事务里，插入群与群主成员之前对 `actorID` 做与 `createAdmin` 相同的 `endpointCheckTx`；已停用 `endpoint_disabled`，不存在 `invalid_target`，且不插入群。不改 `emit`、不改迁移编号。  
+   - 原因：成员侧已有事务内复核，群主侧缺对称校验，竞态会留下非法群主。  
+   - 备选方案：事务外再查一次创建者（否决，无法覆盖密码窗口与写事务之间的竞态）。  
+   - 影响：创建者在进入写事务前被停用/删除时建群失败且无新群行；与后台建群错误码对齐。
+
+### 复审修复 R3-08
+
+1. **后台加人审计按实际插入数，全跳过不记 ok**  
+   - 日期：2026-09-30  
+   - 原条款：Gitea #72；`AdminAddMembers` 对已是成员与请求内重复编号静默跳过；审计用 `len(member_ids)-len(failed)` 当成功数。  
+   - 实际做法：`handleGroupAddMembers` 用加人前后 `group_members` 行数差得到 `added`；审计 `result` 用 `added` 与 `failed`（全跳过且无失败记 `noop`，不记 `ok`）；响应增加 `added`，`failed` 仍只含真正失败项。前端按 `added` 提示，已是成员不改成错误码。  
+   - 原因：全是已有成员时旧逻辑审计成 ok、界面提示已加人，实际插入 0 行。  
+   - 备选方案：把已是成员写入 `failed`（否决，会改变客户端错误语义）；扩展 `group.AddResult` 返回插入列表（可后续做，本波不改身份线接口）。  
+   - 影响：管理 API 加人成功体多 `added` 字段；契约文档示例仍写 `{"failed":[]}`，以本偏差为准。`added` 取写事务内实际插入数（`AddResult.Added`，不进端协议 JSON），不用事务外两次 COUNT 的差，避免并发加人/踢人把别人的变更算进本请求。
+
+### 复审修复 R3-02
+
+1. **积压时 fatal/logout 与停机 0x8B 须等本帧写出**  
+   - 日期：2026-09-30  
+   - 原条款：Gitea #66；B-04 / B-08。  
+   - 实际做法：带断开的下行帧用本帧 `OnPacketSent` 完成信号（优先 packet id，否则按载荷匹配），不再用连接级 `sentPub` 总数。`Shutdown` 在 ctx 未取消时先等下行队列与 `wirePending` 排空，再 `DisconnectClient` 发 `0x8B` 并在截止前等连接拆掉；ctx 已取消则发完即 `Close`。`serve` 仍给 5 秒预算并记录非超时错误。未合 `feat/fix-3-downlink-deadlock`。  
+   - 原因：前面 PUBLISH 的 `OnPacketSent` 会让总数等待提前返回；`Shutdown` 对 ctx 非阻塞 select 使 5 秒预算用不上，有 outbound 积压时 `0x8B` 只进 outbuf 随 `Stop` 丢掉。  
+   - 备选方案：恢复固定 `Sleep`（否决）；改 `PublishDown` 签名（否决）。  
+   - 影响：队列/outbound 有积压时 fatal、logout 先到客户端再断开；停机在预算内尽量发出 `0x8B`，超时返回 ctx 错误而非空等。  
+   - 补强：`sendOne` 从取出帧到返回前记 `inSend`，排空等待把它算上（含等大帧名额、尚未 `wirePending++` 的窗口）。有截止时间时排空最多用到截止前 1 秒，剩下的时间留给 `0x8B` 写出；排空没完成也不会因此跳过这段等待。

internal/admin/groups.go

@@ -325,15 +325,22 @@ func (h *Handler) handleGroupAddMembers(w http.ResponseWriter, r *http.Request)
 	if failed == nil {
 		failed = []group.MemberFail{}
 	}
-	okN := len(req.MemberIDs) - len(failed)
-	if okN < 0 {
-		okN = 0
+	added := res.Added
+	if added < 0 {
+		added = 0
 	}
-	h.auditPD(p, "group_add_members", id, batchAuditResult(okN, len(failed)), ip, map[string]any{
+
+	// 已是成员/请求内重复编号会静默跳过，不算失败；审计用实际插入数，避免全跳过写成 ok。
+	result := batchAuditResult(added, len(failed))
+	if added == 0 && len(failed) == 0 {
+		result = "noop"
+	}
+	h.auditPD(p, "group_add_members", id, result, ip, map[string]any{
 		"members": req.MemberIDs,
 		"failed":  failed,
+		"added":   added,
 	})
-	httpx.WriteOK(w, map[string]any{"failed": failed})
+	httpx.WriteOK(w, map[string]any{"failed": failed, "added": added})
 }
 
 func (h *Handler) handleGroupRemoveMember(w http.ResponseWriter, r *http.Request) {

internal/admin/h02_test.go

@@ -262,6 +262,7 @@ func TestH02BatchPartialImportAndGroups(t *testing.T) {
 	insertEndpoint(t, db, "keep-1", "admin", true, false)
 	insertEndpoint(t, db, "alice", "admin", true, false)
 	insertEndpoint(t, db, "bob", "admin", true, false)
+	insertEndpoint(t, db, "carol", "admin", true, false)
 
 	res := doReq(t, client, http.MethodPost, srv.URL+"/api/admin/endpoints/batch",
 		`{"ids":["keep-1","missing-ep"],"action":"disable"}`,
@@ -304,11 +305,24 @@ func TestH02BatchPartialImportAndGroups(t *testing.T) {
 	}
 
 	res = doReq(t, client, http.MethodPost, srv.URL+"/api/admin/groups/"+created.ID+"/members",
-		`{"member_ids":["keep-1","ghost-ep"]}`, csrf())
+		`{"member_ids":["carol","ghost-ep"]}`, csrf())
 	env = decodeEnv(t, res)
 	if res.StatusCode != http.StatusOK || !env.OK {
 		t.Fatalf("add members: %d %+v", res.StatusCode, env)
 	}
+	var addBody struct {
+		Failed []any `json:"failed"`
+		Added  int   `json:"added"`
+	}
+	if err := json.Unmarshal(env.Data, &addBody); err != nil {
+		t.Fatal(err)
+	}
+	if addBody.Added != 1 {
+		t.Fatalf("add response added=%d want 1", addBody.Added)
+	}
+	if len(addBody.Failed) != 1 {
+		t.Fatalf("add response failed=%v want 1", addBody.Failed)
+	}
 
 	res = doReq(t, client, http.MethodPost, srv.URL+"/api/admin/groups/"+created.ID+"/transfer",
 		`{"endpoint_id":"bob"}`, csrf())
@@ -346,8 +360,8 @@ func TestH02BatchPartialImportAndGroups(t *testing.T) {
 	if ad == nil {
 		t.Fatalf("add members missing detail: %v", add)
 	}
-	if add["result"] != "partial" && add["result"] != "failed" && add["result"] != "ok" {
-		t.Fatalf("add members result=%v", add["result"])
+	if add["result"] != "partial" {
+		t.Fatalf("add members result=%v want partial", add["result"])
 	}
 	if _, ok := ad["members"]; !ok {
 		t.Fatalf("add members missing members: %v", ad)
@@ -355,6 +369,9 @@ func TestH02BatchPartialImportAndGroups(t *testing.T) {
 	if _, ok := ad["failed"]; !ok {
 		t.Fatalf("add members missing failed: %v", ad)
 	}
+	if ad["added"] != float64(1) {
+		t.Fatalf("add members added=%v want 1", ad["added"])
+	}
 
 	tr := lastAuditByAction(t, recs, "group_transfer")
 	td, _ := tr["detail"].(map[string]any)
@@ -364,3 +381,70 @@ func TestH02BatchPartialImportAndGroups(t *testing.T) {
 
 	assertNoSecrets(t, auditBuf.String(), "csv-pass-secret-1", "csv-pass-secret-2", testPassword)
 }
+
+func TestH02AddMembersAllSkippedAuditNotOK(t *testing.T) {
+	_, auditBuf, db, srv, client := setupH02(t)
+	login(t, client, srv.URL)
+	insertEndpoint(t, db, "alice", "admin", true, false)
+	insertEndpoint(t, db, "bob", "admin", true, false)
+
+	res := doReq(t, client, http.MethodPost, srv.URL+"/api/admin/groups",
+		`{"name":"一组","owner_id":"alice","member_ids":["bob"]}`, csrf())
+	env := decodeEnv(t, res)
+	if res.StatusCode != http.StatusOK || !env.OK {
+		t.Fatalf("group create: %d %+v", res.StatusCode, env)
+	}
+	var created struct {
+		ID string `json:"id"`
+	}
+	if err := json.Unmarshal(env.Data, &created); err != nil {
+		t.Fatal(err)
+	}
+
+	auditBuf.Reset()
+	var before int
+	if err := db.Read.QueryRow(`SELECT COUNT(*) FROM group_members WHERE group_id = ?`, created.ID).Scan(&before); err != nil {
+		t.Fatal(err)
+	}
+
+	res = doReq(t, client, http.MethodPost, srv.URL+"/api/admin/groups/"+created.ID+"/members",
+		`{"member_ids":["bob","bob","alice"]}`, csrf())
+	env = decodeEnv(t, res)
+	if res.StatusCode != http.StatusOK || !env.OK {
+		t.Fatalf("add members: %d %+v", res.StatusCode, env)
+	}
+	var body struct {
+		Failed []any `json:"failed"`
+		Added  int   `json:"added"`
+	}
+	if err := json.Unmarshal(env.Data, &body); err != nil {
+		t.Fatal(err)
+	}
+	if len(body.Failed) != 0 {
+		t.Fatalf("failed=%v want empty (already members are not errors)", body.Failed)
+	}
+	if body.Added != 0 {
+		t.Fatalf("added=%d want 0", body.Added)
+	}
+
+	var after int
+	if err := db.Read.QueryRow(`SELECT COUNT(*) FROM group_members WHERE group_id = ?`, created.ID).Scan(&after); err != nil {
+		t.Fatal(err)
+	}
+	if after != before {
+		t.Fatalf("member rows before=%d after=%d want unchanged", before, after)
+	}
+
+	recs := parseSlogJSON(t, auditBuf)
+	add := lastAuditByAction(t, recs, "group_add_members")
+	if add["result"] == "ok" {
+		t.Fatalf("audit result must not be ok when all skipped: %v", add)
+	}
+	ad, _ := add["detail"].(map[string]any)
+	if ad == nil {
+		t.Fatalf("missing detail: %v", add)
+	}
+	if ad["added"] != float64(0) {
+		t.Fatalf("detail.added=%v want 0", ad["added"])
+	}
+}

internal/app/group/app.go

@@ -147,6 +147,15 @@ func (a *App) Create(ctx context.Context, actorID string, req *protocol.GroupCre
 
 	var inserted []string
 	err := a.db.Queue.Do(ctx, func(tx *sql.Tx) error {
+		if e := endpointCheckTx(tx, actorID); e != nil {
+			if protoCode(e) == protocol.CodeInvalidTarget {
+				return errCode(protocol.CodeInvalidTarget, "owner not found")
+			}
+			if protoCode(e) == protocol.CodeEndpointDisabled {
+				return errCode(protocol.CodeEndpointDisabled, "owner disabled")
+			}
+			return e
+		}
 		var exists int
 		qErr := tx.QueryRow(`SELECT 1 FROM groups WHERE id = ?`, gid).Scan(&exists)
 		if qErr == nil {
@@ -611,7 +620,7 @@ func (a *App) AdminAddMembers(ctx context.Context, groupID string, memberIDs []s
 		failed = append(failed, MemberFail{ID: id, Code: protocol.CodeGroupFull})
 	}
 	if len(toAdd) == 0 {
-		return AddResult{Failed: failed}, nil
+		return AddResult{Failed: failed, Added: 0}, nil
 	}
 	now := a.nowMs()
 	var inserted []string
@@ -655,7 +664,7 @@ func (a *App) AdminAddMembers(ctx context.Context, groupID string, memberIDs []s
 	for _, id := range inserted {
 		a.emit(ctx, notify, groupID, eventMemberAdded, id, now)
 	}
-	return AddResult{Failed: failed}, nil
+	return AddResult{Failed: failed, Added: len(inserted)}, nil
 }
 
 func (a *App) createAdmin(ctx context.Context, ownerID, name, gid string, members []protocol.GroupMemberIn) (CreateResult, error) {

internal/app/group/group_test.go

@@ -604,6 +604,106 @@ func TestU02CreateDedupesMembers(t *testing.T) {
 	}
 }
 
+// raceTalk 在对话密码校验成功后执行一次 mutate，模拟事务外密码窗口里创建者被停用/删除。
+type raceTalk struct {
+	inner  group.TalkGate
+	mutate func()
+	once   sync.Once
+}
+
+func (r *raceTalk) CheckTalkPasswordForJoin(ctx context.Context, actorID, targetID, talkPassword, remoteIP string) error {
+	err := r.inner.CheckTalkPasswordForJoin(ctx, actorID, targetID, talkPassword, remoteIP)
+	if err == nil && r.mutate != nil {
+		r.once.Do(r.mutate)
+	}
+	return err
+}
+
+func TestR307CreateRejectsActorDisabledOrDeletedBeforeWrite(t *testing.T) {
+	t.Parallel()
+	cases := []struct {
+		name   string
+		code   string
+		mutate func(ctx context.Context, t *testing.T, idApp *identity.App, db *store.DB)
+	}{
+		{
+			name: "disabled",
+			code: protocol.CodeEndpointDisabled,
+			mutate: func(ctx context.Context, t *testing.T, idApp *identity.App, _ *store.DB) {
+				t.Helper()
+				if err := idApp.Disable(ctx, "alice"); err != nil {
+					t.Fatal(err)
+				}
+			},
+		},
+		{
+			name: "deleted",
+			code: protocol.CodeInvalidTarget,
+			mutate: func(ctx context.Context, t *testing.T, idApp *identity.App, _ *store.DB) {
+				t.Helper()
+				if err := idApp.Delete(ctx, "alice"); err != nil {
+					t.Fatal(err)
+				}
+			},
+		},
+	}
+	for _, tc := range cases {
+		tc := tc
+		t.Run(tc.name, func(t *testing.T) {
+			t.Parallel()
+			db, err := store.Open(filepath.Join(t.TempDir(), "data"), "FULL")
+			if err != nil {
+				t.Fatal(err)
+			}
+			t.Cleanup(func() { _ = db.Close() })
+			fixed := time.UnixMilli(1_700_000_000_000)
+			locks := auth.NewLoginLocks()
+			idApp := identity.New(identity.Config{
+				DB: db, Hash: auth.NewStubHashPool(), Locks: locks,
+				Sessions: auth.NewSessionTokens(),
+				Now:      func() time.Time { return fixed },
+			})
+			gate := &raceTalk{inner: idApp}
+			gApp := group.New(group.Config{
+				DB: db, Talk: gate, MaxGroupMembers: 1000,
+				Now: func() time.Time { return fixed }, DefaultRemoteIP: "1.1.1.1",
+			})
+			ctx := context.Background()
+			insertEP(t, db, "alice", 1)
+			insertEP(t, db, "bob", 1)
+			if setErr := idApp.SelfSetTalkPassword(ctx, "bob", "secret"); setErr != nil {
+				t.Fatal(setErr)
+			}
+			gate.mutate = func() { tc.mutate(ctx, t, idApp, db) }
+
+			var before int
+			if qErr := db.Read.QueryRow(`SELECT COUNT(*) FROM groups`).Scan(&before); qErr != nil {
+				t.Fatal(qErr)
+			}
+			_, err = gApp.Create(ctx, "alice", &protocol.GroupCreate{
+				V: protocol.Version, Type: protocol.TypeGroupCreate, RID: "1",
+				ID: "g_r307_" + tc.name, Name: "竞态群",
+				Members: []protocol.GroupMemberIn{{ID: "bob", TalkPassword: "secret"}},
+			})
+			if protoCode(err) != tc.code {
+				t.Fatalf("want %s got %v", tc.code, err)
+			}
+			var after int
+			if qErr := db.Read.QueryRow(`SELECT COUNT(*) FROM groups`).Scan(&after); qErr != nil {
+				t.Fatal(qErr)
+			}
+			if after != before {
+				t.Fatalf("groups leaked: before=%d after=%d", before, after)
+			}
+			var exists int
+			qErr := db.Read.QueryRow(`SELECT 1 FROM groups WHERE id = ?`, "g_r307_"+tc.name).Scan(&exists)
+			if !errors.Is(qErr, sql.ErrNoRows) {
+				t.Fatalf("expected no group row, got exists=%d err=%v", exists, qErr)
+			}
+		})
+	}
+}
+
 func TestU02AdminCreateOwnerMustExistAndEnabled(t *testing.T) {
 	t.Parallel()
 	gApp, _, _, db, down := setup(t)

internal/app/group/service.go

@@ -28,6 +28,8 @@ type CreateResult struct {
 // AddResult 是加人结果。
 type AddResult struct {
 	Failed []MemberFail `json:"failed,omitempty"`
+	// Added 是本请求在写事务内实际插入的人数。不进协议 JSON，供后台审计使用。
+	Added int `json:"-"`
 }
 
 // ListItem 是 group.list 一项。

internal/app/message/ack.go

@@ -90,6 +90,7 @@ func (a *App) Recall(ctx context.Context, senderID string, req *protocol.Recall)
 	nowMs := a.now().UnixMilli()
 	var data protocol.RecallData
 	var revokes []revokeJob
+	var wakeReceivers []string
 	err := a.db.Queue.Do(ctx, func(tx *sql.Tx) error {
 		var seq int64
 		var state string
@@ -156,6 +157,7 @@ WHERE seq = ? AND endpoint_id = ? AND state = 'pending'`,
 				continue
 			}
 			recalled++
+			wakeReceivers = append(wakeReceivers, p.ep)
 			if p.pushed.Valid && p.pushed.String != "" {
 				revokes = append(revokes, revokeJob{
 					endpointID: p.ep,
@@ -190,6 +192,9 @@ SELECT COUNT(*) FROM deliveries WHERE seq = ? AND state IN ('expired','dropped',
 	a.pendingRevoke = append(a.pendingRevoke, revokes...)
 	a.mu.Unlock()
 	a.flushRevokes(ctx)
+	for _, ep := range wakeReceivers {
+		a.WakePush(ep)
+	}
 	return data, nil
 }
 

internal/app/message/push.go

@@ -127,6 +127,7 @@ LIMIT ?`, nowMs, limit)
 			}
 			mu.Lock()
 			n++
+			wake[d.senderID] = struct{}{}
 			mu.Unlock()
 			rows2, qErr := a.db.Read.QueryContext(ctx, `
 SELECT DISTINCT endpoint_id FROM deliveries WHERE seq = ? AND state = 'pending'`, d.seq)
@@ -203,8 +204,10 @@ LIMIT ?`, endpointID, room)
 	}
 
 	var toClaim []pushItem
+	skipped := false
 	for _, it := range items {
 		if it.body == nil {
+			skipped = true
 			continue
 		}
 		msg := protocol.Msg{
@@ -227,6 +230,7 @@ LIMIT ?`, endpointID, room)
 				return rejErr
 			}
 			a.WakePush(it.senderID)
+			skipped = true
 			continue
 		}
 		it.payload = payload
@@ -252,6 +256,10 @@ LIMIT ?`, endpointID, room)
 			a.observeDispatchToPush(it.sendAt, nowMs)
 		}
 	}
+	// 超限等跳过未占满窗口时再唤醒本端，让 worker 下一轮取后续 pending（不在本轮递归扫表）。
+	if skipped && len(claimed) < room {
+		a.WakePush(endpointID)
+	}
 	return a.pushReceipts(ctx, endpointID, connID, nowMs)
 }
 

internal/app/message/review_r3_65_test.go

@@ -0,0 +1,180 @@
+package message
+
+import (
+	"context"
+	"database/sql"
+	"encoding/json"
+	"testing"
+	"time"
+
+	"git.asio.asia/nixevol/NixMsg/internal/app/port"
+	"git.asio.asia/nixevol/NixMsg/internal/protocol"
+)
+
+// 定时到点且接收方离线成终态时，在线发送方经 WakePush 拿到回执。
+func TestR365DispatchDueWakesSenderForReceipt(t *testing.T) {
+	t.Parallel()
+	e := openDeliveryEnv(t, nil)
+	insertEndpoint(t, e.db, "alice", "", 1, 0)
+	insertEndpoint(t, e.db, "bob", "", 1, 0)
+	_ = e.db.Queue.Do(context.Background(), func(tx *sql.Tx) error {
+		_, err := tx.Exec(`UPDATE endpoints SET offline_since = ? WHERE id=?`, e.nowMs-120_000, "bob")
+		return err
+	})
+	ctx := context.Background()
+	aliceConn := LiveConn{ConnID: "c-alice", Ready: true}
+	e.conns.Set("alice", aliceConn)
+	if err := e.app.OnHandshakeComplete(ctx, "alice", aliceConn); err != nil {
+		t.Fatal(err)
+	}
+	t.Cleanup(func() { e.app.stopPushWorker("alice", "") })
+
+	delay := int64(5000)
+	req := baseSend("due-rcpt", "bob")
+	req.DelayMs = &delay
+	if _, err := e.app.Submit(ctx, "alice", port.ConnInfo{}, req); err != nil {
+		t.Fatal(err)
+	}
+	st, _ := e.msgState("alice", "due-rcpt")
+	if st != StateScheduled {
+		t.Fatalf("want scheduled got %s", st)
+	}
+	if e.down.FilterType(protocol.TypeReceipt) != 0 {
+		t.Fatal("receipt before due")
+	}
+
+	e.setNow(e.nowMs + 5000)
+	if _, err := e.app.DispatchDue(ctx, e.nowMs, 10); err != nil {
+		t.Fatal(err)
+	}
+	deadline := time.Now().Add(2 * time.Second)
+	for time.Now().Before(deadline) {
+		if e.down.FilterType(protocol.TypeReceipt) > 0 {
+			return
+		}
+		time.Sleep(20 * time.Millisecond)
+	}
+	t.Fatalf("sender got no receipt after scheduled dispatch, receipts=%d", e.down.FilterType(protocol.TypeReceipt))
+}
+
+// 窗口内首条超限拒收后，同连接后续小消息仍被 worker 推送。
+func TestR365TooLargeThenSmallContinuesPush(t *testing.T) {
+	t.Parallel()
+	e := openDeliveryEnv(t, func(l *Limits) { l.DeliveryWindow = 1 })
+	insertEndpoint(t, e.db, "alice", "", 1, 0)
+	insertEndpoint(t, e.db, "bob", "", 1, 0)
+	ctx := context.Background()
+	// 整帧上限：小消息能过，大正文整帧超限被拒。
+	bobConn := LiveConn{ConnID: "c-bob", Ready: true, MaxReceiveBytes: 200}
+	e.conns.Set("bob", bobConn)
+	if err := e.app.OnHandshakeComplete(ctx, "bob", bobConn); err != nil {
+		t.Fatal(err)
+	}
+	t.Cleanup(func() { e.app.stopPushWorker("bob", "") })
+
+	big := baseSend("big-skip", "bob")
+	b := make([]byte, 400)
+	for i := range b {
+		b[i] = 'A'
+	}
+	big.Body.Data = string(b)
+	if _, err := e.app.Submit(ctx, "alice", port.ConnInfo{}, big); err != nil {
+		t.Fatal(err)
+	}
+	small := baseSend("small-ok", "bob")
+	small.Body.Data = "hi"
+	if _, err := e.app.Submit(ctx, "alice", port.ConnInfo{}, small); err != nil {
+		t.Fatal(err)
+	}
+
+	deadline := time.Now().Add(2 * time.Second)
+	for time.Now().Before(deadline) {
+		foundSmall := false
+		for _, p := range e.down.Snapshots() {
+			if payloadType(p.Payload) != protocol.TypeMsg {
+				continue
+			}
+			var m struct {
+				ID string `json:"id"`
+			}
+			_ = json.Unmarshal(p.Payload, &m)
+			if m.ID == "small-ok" {
+				foundSmall = true
+			}
+		}
+		if foundSmall {
+			seqBig := e.seqOf("alice", "big-skip")
+			st, reason := e.deliveryState(seqBig, "bob")
+			if st != DeliveryRejected || reason != ReasonTooLarge {
+				t.Fatalf("big: %s/%s", st, reason)
+			}
+			return
+		}
+		time.Sleep(20 * time.Millisecond)
+	}
+	t.Fatalf("small msg not pushed after too_large; msgs=%d snapshots=%d",
+		e.down.FilterType(protocol.TypeMsg), len(e.down.Snapshots()))
+}
+
+// 撤回已推在途消息后，同连接更晚的 pending 继续推。
+func TestR365RecallFreesWindowForLaterPending(t *testing.T) {
+	t.Parallel()
+	e := openDeliveryEnv(t, func(l *Limits) { l.DeliveryWindow = 1 })
+	insertEndpoint(t, e.db, "alice", "", 1, 0)
+	insertEndpoint(t, e.db, "bob", "", 1, 0)
+	ctx := context.Background()
+	bobConn := LiveConn{ConnID: "c-bob", Ready: true}
+	e.conns.Set("bob", bobConn)
+	if err := e.app.OnHandshakeComplete(ctx, "bob", bobConn); err != nil {
+		t.Fatal(err)
+	}
+	t.Cleanup(func() { e.app.stopPushWorker("bob", "") })
+
+	if _, err := e.app.Submit(ctx, "alice", port.ConnInfo{}, baseSend("first-in-flight", "bob")); err != nil {
+		t.Fatal(err)
+	}
+	if _, err := e.app.Submit(ctx, "alice", port.ConnInfo{}, baseSend("second-wait", "bob")); err != nil {
+		t.Fatal(err)
+	}
+
+	deadline := time.Now().Add(2 * time.Second)
+	for time.Now().Before(deadline) {
+		if e.down.FilterType(protocol.TypeMsg) >= 1 {
+			break
+		}
+		time.Sleep(10 * time.Millisecond)
+	}
+	if e.down.FilterType(protocol.TypeMsg) < 1 {
+		t.Fatal("first msg not pushed")
+	}
+	seq1 := e.seqOf("alice", "first-in-flight")
+	var pushed sql.NullString
+	_ = e.db.Read.QueryRow(`SELECT pushed_conn FROM deliveries WHERE seq=? AND endpoint_id='bob'`, seq1).Scan(&pushed)
+	if !pushed.Valid {
+		t.Fatal("first not in-flight")
+	}
+
+	if _, err := e.app.Recall(ctx, "alice", &protocol.Recall{
+		V: protocol.Version, Type: protocol.TypeRecall, RID: "r1", ID: "first-in-flight",
+	}); err != nil {
+		t.Fatal(err)
+	}
+
+	deadline = time.Now().Add(2 * time.Second)
+	for time.Now().Before(deadline) {
+		for _, p := range e.down.Snapshots() {
+			if payloadType(p.Payload) != protocol.TypeMsg {
+				continue
+			}
+			var m struct {
+				ID string `json:"id"`
+			}
+			_ = json.Unmarshal(p.Payload, &m)
+			if m.ID == "second-wait" {
+				return
+			}
+		}
+		time.Sleep(20 * time.Millisecond)
+	}
+	t.Fatalf("second pending not pushed after recall; msgs=%d", e.down.FilterType(protocol.TypeMsg))
+}

internal/broker/b04_b06_b08_test.go

@@ -87,6 +87,50 @@ func TestPublishThenDisconnectWritesThenCloses(t *testing.T) {
 	}
 }
 
+func TestPublishThenDisconnectAfterQueuedFrame(t *testing.T) {
+	b, w, done := startTCPClient(t, "ep-ptd-q")
+	defer func() { _ = b.Close() }()
+	defer func() {
+		_ = w.Close()
+		select {
+		case <-done:
+		case <-time.After(3 * time.Second):
+		}
+	}()
+	writeConnect(t, w, "ep-ptd-q", 30, 0)
+	readExactPacket(t, w, packets.Connack, 3*time.Second)
+	writeSubscribe(t, w, downTopic("ep-ptd-q"))
+	readExactPacket(t, w, packets.Suback, 3*time.Second)
+	waitSession(t, b, "ep-ptd-q")
+
+	first := []byte(`{"v":1,"type":"msg","id":"queued-ahead"}`)
+	fatal := []byte(`{"v":1,"type":"fatal","reason":"disabled"}`)
+	if err := b.PublishDown(context.Background(), "ep-ptd-q", "", first, port.PublishOpts{QoS: 1}); err != nil {
+		t.Fatal(err)
+	}
+	if err := b.PublishThenDisconnect(context.Background(), "ep-ptd-q", "", fatal, 1, port.DisconnectFatal); err != nil {
+		t.Fatal(err)
+	}
+
+	gotFirst := readDownPayload(t, w, 3*time.Second)
+	if !bytes.Equal(gotFirst, first) {
+		t.Fatalf("first got %s", gotFirst)
+	}
+	gotFatal := readDownPayload(t, w, 3*time.Second)
+	if !bytes.Equal(gotFatal, fatal) {
+		t.Fatalf("fatal got %s want %s (disconnected before fatal frame)", gotFatal, fatal)
+	}
+	_ = w.SetReadDeadline(time.Now().Add(3 * time.Second))
+	buf := make([]byte, 64)
+	n, err := io.ReadAtLeast(w, buf, 2)
+	if err != nil && n == 0 {
+		return
+	}
+	if n > 0 && buf[0]>>4 == packets.Disconnect {
+		return
+	}
+}
+
 func TestShutdownUsesServerShuttingDown(t *testing.T) {
 	b, w, done := startTCPClient(t, "ep-shut")
 	defer func() {
@@ -116,6 +160,141 @@ func TestShutdownUsesServerShuttingDown(t *testing.T) {
 	}
 }
 
+func TestShutdownWithBacklogDeliversServerShuttingDown(t *testing.T) {
+	b, w, done := startTCPClient(t, "ep-shut-bl")
+	defer func() {
+		_ = w.Close()
+		select {
+		case <-done:
+		case <-time.After(3 * time.Second):
+		}
+	}()
+	writeConnect(t, w, "ep-shut-bl", 30, 0)
+	readExactPacket(t, w, packets.Connack, 3*time.Second)
+	writeSubscribe(t, w, downTopic("ep-shut-bl"))
+	readExactPacket(t, w, packets.Suback, 3*time.Second)
+	waitSession(t, b, "ep-shut-bl")
+
+	payload := bytes.Repeat([]byte("b"), 1024)
+	for i := 0; i < 8; i++ {
+		if err := b.PublishDown(context.Background(), "ep-shut-bl", "", payload, port.PublishOpts{QoS: 1}); err != nil {
+			t.Fatalf("publish %d: %v", i, err)
+		}
+	}
+
+	saw8B := make(chan bool, 1)
+	go func() {
+		deadline := time.Now().Add(3 * time.Second)
+		for time.Now().Before(deadline) {
+			_ = w.SetReadDeadline(time.Now().Add(200 * time.Millisecond))
+			hdr := make([]byte, 1)
+			if _, err := io.ReadFull(w, hdr); err != nil {
+				continue
+			}
+			rem, err := readRemainingLengthConn(w)
+			if err != nil {
+				continue
+			}
+			body := make([]byte, rem)
+			if _, err := io.ReadFull(w, body); err != nil {
+				continue
+			}
+			switch hdr[0] >> 4 {
+			case packets.Publish:
+				qos := (hdr[0] >> 1) & 0x3
+				if qos > 0 {
+					pk := new(packets.Packet)
+					pk.ProtocolVersion = 5
+					pk.FixedHeader = packets.FixedHeader{Type: packets.Publish, Remaining: rem, Qos: qos}
+					if decErr := pk.PublishDecode(body); decErr == nil {
+						ack := packets.Packet{
+							FixedHeader:     packets.FixedHeader{Type: packets.Puback},
+							ProtocolVersion: 5,
+							PacketID:        pk.PacketID,
+						}
+						var ab bytes.Buffer
+						_ = ack.PubackEncode(&ab)
+						_, _ = w.Write(ab.Bytes())
+					}
+				}
+			case packets.Disconnect:
+				if rem >= 1 && body[0] == packets.ErrServerShuttingDown.Code {
+					saw8B <- true
+					return
+				}
+			}
+		}
+		saw8B <- false
+	}()
+	time.Sleep(20 * time.Millisecond)
+
+	ctx, cancel := context.WithTimeout(context.Background(), 3*time.Second)
+	defer cancel()
+	err := b.Shutdown(ctx)
+
+	got := false
+	select {
+	case got = <-saw8B:
+	case <-time.After(4 * time.Second):
+		t.Fatal("reader hung")
+	}
+	if got {
+		if err != nil && !errors.Is(err, context.DeadlineExceeded) {
+			t.Fatalf("shutdown after 0x8B: %v", err)
+		}
+		return
+	}
+	if err == nil {
+		t.Fatal("expected 0x8B or shutdown deadline error, got neither")
+	}
+	if !errors.Is(err, context.DeadlineExceeded) && !errors.Is(err, context.Canceled) {
+		t.Fatalf("shutdown err=%v want deadline/cancel when 0x8B not seen", err)
+	}
+}
+
+func TestShutdownCancelledContextReturnsQuickly(t *testing.T) {
+	b, w, done := startTCPClient(t, "ep-shut-cancel")
+	defer func() {
+		_ = w.Close()
+		select {
+		case <-done:
+		case <-time.After(3 * time.Second):
+		}
+	}()
+	writeConnect(t, w, "ep-shut-cancel", 30, 0)
+	readExactPacket(t, w, packets.Connack, 3*time.Second)
+	waitSession(t, b, "ep-shut-cancel")
+
+	ctx, cancel := context.WithCancel(context.Background())
+	cancel()
+	start := time.Now()
+	_ = b.Shutdown(ctx)
+	if time.Since(start) > 500*time.Millisecond {
+		t.Fatalf("cancelled shutdown took %s", time.Since(start))
+	}
+}
+
+func TestWaitConnsQuietSeesInSend(t *testing.T) {
+	b, err := New(Options{})
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer func() { _ = b.Close() }()
+	st := &connState{downCh: make(chan downItem, 1)}
+	st.inSend.Store(1)
+	ctx, cancel := context.WithTimeout(context.Background(), 40*time.Millisecond)
+	defer cancel()
+	if b.waitConnsQuiet(ctx, []*connState{st}) {
+		t.Fatal("inSend should keep shutdown from treating the conn as quiet")
+	}
+	st.inSend.Store(0)
+	ctx2, cancel2 := context.WithTimeout(context.Background(), 200*time.Millisecond)
+	defer cancel2()
+	if !b.waitConnsQuiet(ctx2, []*connState{st}) {
+		t.Fatal("quiet when inSend is 0 and queues are empty")
+	}
+}
+
 func TestEffectivePayloadLimitSubtractsOverhead(t *testing.T) {
 	got := EffectivePayloadLimit(200, 0)
 	if got != 200-packetOverheadBudget {

internal/broker/broker.go

@@ -141,9 +141,15 @@ type connState struct {
 	downStop       chan struct{}
 	downDone       chan struct{}
 	downBytes      atomic.Int64
-	sentPub        atomic.Int64
+	wirePending    atomic.Int64 // Publish 入 mochi outbound 后、OnPacketSent 前
+	inSend         atomic.Int32 // downLoop 已取出帧、尚未从 sendOne 返回
 	mu             sync.Mutex
 
+	// 带断开的下行帧：只等本帧 OnPacketSent，不用连接级计数。
+	writeWaitCh      chan struct{}
+	writeWaitPayload []byte
+	writeWaitPID     uint16 // 非 0 时优先按 packet id 匹配
+
 	handshakeTimer *time.Timer
 }
 
@@ -232,28 +238,114 @@ func (b *Broker) Close() error {
 }
 
 // Shutdown 向所有连接发 MQTT 5 0x8B 后关闭。完整 HTTP 停机顺序见 L-03。
+// ctx 未取消时先等下行队列、正在 sendOne 的帧与 wirePending 排空（若有截止时间则预留约 1 秒），
+// 再 DisconnectClient，并在截止前等连接拆掉；ctx 已取消则发完即 Close，不等待。
 func (b *Broker) Shutdown(ctx context.Context) error {
 	if b.closed.Load() {
 		return nil
 	}
+	if ctx == nil {
+		ctx = context.Background()
+	}
+
 	b.connsMu.RLock()
+	states := make([]*connState, 0, len(b.byClient))
 	clients := make([]*mqtt.Client, 0, len(b.byClient))
-	for cl := range b.byClient {
+	for cl, st := range b.byClient {
 		if cl != nil {
 			clients = append(clients, cl)
 		}
+		if st != nil {
+			states = append(states, st)
+		}
 	}
 	b.connsMu.RUnlock()
+
+	for _, st := range states {
+		st.mu.Lock()
+		st.closing = true
+		st.mu.Unlock()
+	}
+
+	alreadyCancelled := false
+	select {
+	case <-ctx.Done():
+		alreadyCancelled = true
+	default:
+	}
+
+	var waitErr error
+	if !alreadyCancelled {
+		// 留出约 1 秒给 0x8B 写出，避免排空把整个截止时间用完后立刻 Close。
+		quietCtx := ctx
+		cancelQuiet := func() {}
+		if dl, ok := ctx.Deadline(); ok && time.Until(dl) > time.Second {
+			quietCtx, cancelQuiet = context.WithDeadline(ctx, dl.Add(-time.Second))
+		}
+		if !b.waitConnsQuiet(quietCtx, states) && ctx.Err() != nil {
+			waitErr = ctx.Err()
+		}
+		cancelQuiet()
+	}
+
 	for _, cl := range clients {
 		_ = b.server.DisconnectClient(cl, packets.ErrServerShuttingDown)
 	}
-	if ctx != nil {
+
+	if !alreadyCancelled && ctx.Err() == nil {
+		if err := b.waitConnsGone(ctx); err != nil {
+			waitErr = err
+		}
+	}
+
+	closeErr := b.Close()
+	if waitErr != nil {
+		return waitErr
+	}
+	return closeErr
+}
+
+func (b *Broker) waitConnsQuiet(ctx context.Context, states []*connState) bool {
+	for {
+		quiet := true
+		for _, st := range states {
+			if st.inSend.Load() > 0 || st.wirePending.Load() > 0 {
+				quiet = false
+				break
+			}
+			st.mu.Lock()
+			ch := st.downCh
+			st.mu.Unlock()
+			if len(ch) > 0 {
+				quiet = false
+				break
+			}
+		}
+		if quiet {
+			return true
+		}
 		select {
 		case <-ctx.Done():
-		default:
+			return false
+		case <-time.After(2 * time.Millisecond):
+		}
+	}
+}
+
+func (b *Broker) waitConnsGone(ctx context.Context) error {
+	for {
+		b.connsMu.RLock()
+		n := len(b.byClient)
+		b.connsMu.RUnlock()
+		if n == 0 {
+			return nil
+		}
+		select {
+		case <-ctx.Done():
+			return ctx.Err()
+		case <-time.After(2 * time.Millisecond):
 		}
 	}
-	return b.Close()
 }
 
 // AttachTCP 把裸 TCP/TLS 连接交给 mochi；阻塞到连接结束。

internal/broker/downlink.go

@@ -1,10 +1,12 @@
 package broker
 
 import (
+	"bytes"
 	"context"
 	"time"
 
 	"git.asio.asia/nixevol/NixMsg/internal/app/port"
+	"github.com/mochi-mqtt/server/v2/packets"
 )
 
 type downItem struct {
@@ -116,6 +118,8 @@ func (st *connState) sendOne(b *Broker, item downItem) {
 		st.signalSent(item)
 		return
 	}
+	st.inSend.Add(1)
+	defer st.inSend.Add(-1)
 	large := len(item.payload) > largeFrameBytes
 	if large {
 		if err := b.acquireLarge(context.Background()); err != nil {
@@ -130,7 +134,11 @@ func (st *connState) sendOne(b *Broker, item downItem) {
 		st.mu.Unlock()
 	}
 	topic := downTopic(st.endpointID)
-	before := st.sentPub.Load()
+	var waitCh chan struct{}
+	if item.disconnect != "" {
+		waitCh = st.armWriteWait(item.payload)
+	}
+	st.wirePending.Add(1)
 	var err error
 	for !b.closed.Load() {
 		select {
@@ -150,6 +158,14 @@ func (st *connState) sendOne(b *Broker, item downItem) {
 		}
 		break
 	}
+	if err != nil {
+		st.wirePending.Add(-1)
+		st.clearWriteWait(waitCh)
+	} else if waitCh != nil {
+		if pid, ok := st.lookupInflightPID(item.payload); ok {
+			st.setWriteWaitPID(waitCh, pid)
+		}
+	}
 	if large {
 		b.finishLargePublish(st)
 	}
@@ -158,23 +174,112 @@ func (st *connState) sendOne(b *Broker, item downItem) {
 	}
 	st.signalSent(item)
 	if err == nil && item.disconnect != "" {
-		st.waitPacketWritten(before)
+		st.waitWriteDone(waitCh)
 		_ = b.Disconnect(context.Background(), st.endpointID, st.connID, item.disconnect)
 	}
 }
 
-func (st *connState) waitPacketWritten(before int64) {
-	deadline := time.Now().Add(2 * time.Second)
-	for time.Now().Before(deadline) {
-		if st.sentPub.Load() > before {
+func (st *connState) armWriteWait(payload []byte) chan struct{} {
+	ch := make(chan struct{})
+	st.mu.Lock()
+	st.writeWaitCh = ch
+	st.writeWaitPayload = payload
+	st.writeWaitPID = 0
+	st.mu.Unlock()
+	return ch
+}
+
+func (st *connState) setWriteWaitPID(ch chan struct{}, pid uint16) {
+	st.mu.Lock()
+	if st.writeWaitCh == ch {
+		st.writeWaitPID = pid
+	}
+	st.mu.Unlock()
+}
+
+func (st *connState) clearWriteWait(ch chan struct{}) {
+	if ch == nil {
 		return
 	}
+	st.mu.Lock()
+	if st.writeWaitCh == ch {
+		st.writeWaitCh = nil
+		st.writeWaitPayload = nil
+		st.writeWaitPID = 0
+	}
+	st.mu.Unlock()
+}
+
+func (st *connState) waitWriteDone(ch chan struct{}) {
+	if ch == nil {
+		return
+	}
+	defer st.clearWriteWait(ch)
+	deadline := time.NewTimer(2 * time.Second)
+	defer deadline.Stop()
 	select {
+	case <-ch:
 	case <-st.downStop:
+	case <-deadline.C:
+	}
+}
+
+func (st *connState) notePacketSent(pk packets.Packet) {
+	if pk.FixedHeader.Type == packets.Publish {
+		for {
+			cur := st.wirePending.Load()
+			if cur <= 0 {
+				break
+			}
+			if st.wirePending.CompareAndSwap(cur, cur-1) {
+				break
+			}
+		}
+	}
+	st.mu.Lock()
+	ch := st.writeWaitCh
+	pid := st.writeWaitPID
+	want := st.writeWaitPayload
+	st.mu.Unlock()
+	if ch == nil || pk.FixedHeader.Type != packets.Publish {
 		return
-		case <-time.After(2 * time.Millisecond):
+	}
+	match := false
+	if pid != 0 {
+		match = pk.PacketID == pid
+	} else if want != nil {
+		match = bytes.Equal(pk.Payload, want)
+	}
+	if !match {
+		return
+	}
+	st.mu.Lock()
+	if st.writeWaitCh == ch {
+		st.writeWaitCh = nil
+		st.writeWaitPayload = nil
+		st.writeWaitPID = 0
+	}
+	st.mu.Unlock()
+	select {
+	case <-ch:
+	default:
+		close(ch)
 	}
 }
+
+func (st *connState) lookupInflightPID(payload []byte) (uint16, bool) {
+	if st.client == nil || st.client.State.Inflight == nil {
+		return 0, false
+	}
+	for _, pk := range st.client.State.Inflight.GetAll(false) {
+		if pk.FixedHeader.Type != packets.Publish {
+			continue
+		}
+		if bytes.Equal(pk.Payload, payload) {
+			return pk.PacketID, pk.PacketID != 0
+		}
+	}
+	return 0, false
 }
 
 func (st *connState) signalSent(item downItem) {

internal/broker/hooks.go

@@ -175,14 +175,11 @@ func (h *nixHook) OnSubscribed(cl *mqtt.Client, pk packets.Packet, reasonCodes [
 }
 
 func (h *nixHook) OnPacketSent(cl *mqtt.Client, pk packets.Packet, _ []byte) {
-	if pk.FixedHeader.Type != packets.Publish {
-		return
-	}
 	h.b.connsMu.RLock()
 	st := h.b.byClient[cl]
 	h.b.connsMu.RUnlock()
 	if st != nil {
-		st.sentPub.Add(1)
+		st.notePacketSent(pk)
 	}
 }
 
@@ -254,6 +251,7 @@ func (h *nixHook) OnDisconnect(cl *mqtt.Client, err error, _ bool) {
 	lk.Lock()
 	st.stopDownLoop()
 	lk.Unlock()
+	st.wirePending.Store(0)
 	h.b.releaseAllLarge(st)
 	h.b.cancelHandshakeDeadline(st.endpointID, st.connID)
 

internal/store/queue.go

@@ -38,6 +38,7 @@ type Queue struct {
 
 	ch      chan writeJob
 	done    chan struct{}
+	closing chan struct{} // Close 时关闭，唤醒持读锁阻塞在发送上的 enqueue
 	closed  atomic.Bool
 	sendMu  sync.RWMutex
 
@@ -52,10 +53,15 @@ type Queue struct {
 
 // NewQueue 创建合并写入队列并启动写 goroutine。
 func NewQueue(db *sql.DB) *Queue {
+	return newQueue(db, queueBuffSize)
+}
+
+func newQueue(db *sql.DB, buffSize int) *Queue {
 	q := &Queue{
 		db:      db,
-		ch:    make(chan writeJob, queueBuffSize),
+		ch:      make(chan writeJob, buffSize),
 		done:    make(chan struct{}),
+		closing: make(chan struct{}),
 		ready:   true,
 	}
 	go q.loop()
@@ -115,10 +121,15 @@ func (q *Queue) enqueue(job writeJob) error {
 		return ErrQueueClosed
 	}
 	q.addPending(1)
+	// 通道满时不得只堵在发送上持有读锁：Close 需要写锁关闭 q.ch。
 	select {
 	case q.ch <- job:
 		q.sendMu.RUnlock()
 		return nil
+	case <-q.closing:
+		q.addPending(-1)
+		q.sendMu.RUnlock()
+		return ErrQueueClosed
 	case <-job.ctx.Done():
 		q.addPending(-1)
 		q.sendMu.RUnlock()
@@ -376,11 +387,13 @@ func (q *Queue) Drain(ctx context.Context) error {
 }
 
 // Close 关闭队列：不再接受新任务，并等待写 goroutine 处理完已入队任务后退出。
-// 在写锁内关闭数据通道，避免并发 Do 向已关闭 channel 发送而 panic。
+// 先关闭 closing 唤醒因通道满而阻塞的发送方并释放读锁，再在无发送者时关闭数据通道，
+// 避免向已关闭 channel 发送而 panic，也避免与持读锁的 enqueue 死锁。
 func (q *Queue) Close() error {
 	if q.closed.Swap(true) {
 		return nil
 	}
+	close(q.closing)
 	q.sendMu.Lock()
 	close(q.ch)
 	q.sendMu.Unlock()

internal/store/queue_test.go

@@ -289,3 +289,137 @@ ON CONFLICT(key) DO UPDATE SET value = excluded.value, updated_at = excluded.upd
 		}
 	}
 }
+
+// openSmallQueue 用小缓冲队列替换默认队列，便于测满通道时的 Close/Drain。
+func openSmallQueue(t *testing.T, buf int) *DB {
+	t.Helper()
+	dir := t.TempDir()
+	db, err := Open(dir, "FULL")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if err := db.Queue.Close(); err != nil {
+		t.Fatal(err)
+	}
+	db.Queue = newQueue(db.Write, buf)
+	return db
+}
+
+func TestQueueCloseUnblocksFullChannel(t *testing.T) {
+	t.Parallel()
+	const buf = 4
+	db := openSmallQueue(t, buf)
+	defer func() { _ = db.Close() }()
+	q := db.Queue
+
+	hold := make(chan struct{})
+	blockerStarted := make(chan struct{})
+	blockerErr := make(chan error, 1)
+	go func() {
+		blockerErr <- q.Do(context.Background(), func(tx *sql.Tx) error {
+			close(blockerStarted)
+			<-hold
+			return nil
+		})
+	}()
+	<-blockerStarted
+
+	ctx := context.Background()
+	var fillWG sync.WaitGroup
+	for i := 0; i < buf; i++ {
+		fillWG.Add(1)
+		go func() {
+			defer fillWG.Done()
+			_ = q.Do(ctx, func(tx *sql.Tx) error { return nil })
+		}()
+	}
+	deadline := time.Now().Add(2 * time.Second)
+	for q.Len() < buf+1 && time.Now().Before(deadline) {
+		time.Sleep(2 * time.Millisecond)
+	}
+	if q.Len() < buf+1 {
+		t.Fatalf("channel not full: len=%d", q.Len())
+	}
+
+	blockedErr := make(chan error, 1)
+	go func() {
+		blockedErr <- q.Do(ctx, func(tx *sql.Tx) error { return nil })
+	}()
+	// 等额外 Do 堵在 enqueue（pending 超过通道容量+正在执行的一条）。
+	deadline = time.Now().Add(2 * time.Second)
+	for q.Len() < buf+2 && time.Now().Before(deadline) {
+		time.Sleep(2 * time.Millisecond)
+	}
+
+	closeDone := make(chan error, 1)
+	go func() { closeDone <- q.Close() }()
+
+	select {
+	case err := <-blockedErr:
+		if !errors.Is(err, ErrQueueClosed) {
+			t.Fatalf("blocked Do: %v", err)
+		}
+	case <-time.After(2 * time.Second):
+		t.Fatal("Close did not unblock full-channel enqueue")
+	}
+
+	close(hold)
+	select {
+	case err := <-closeDone:
+		if err != nil {
+			t.Fatal(err)
+		}
+	case <-time.After(2 * time.Second):
+		t.Fatal("Close hung after writer released")
+	}
+	<-blockerErr
+	fillWG.Wait()
+}
+
+func TestQueueDrainTimeoutWhileWriterBlocked(t *testing.T) {
+	t.Parallel()
+	const buf = 4
+	db := openSmallQueue(t, buf)
+	defer func() { _ = db.Close() }()
+	q := db.Queue
+
+	hold := make(chan struct{})
+	started := make(chan struct{})
+	go func() {
+		_ = q.Do(context.Background(), func(tx *sql.Tx) error {
+			close(started)
+			<-hold
+			return nil
+		})
+	}()
+	<-started
+
+	ctx := context.Background()
+	var wg sync.WaitGroup
+	for i := 0; i < buf; i++ {
+		wg.Add(1)
+		go func() {
+			defer wg.Done()
+			_ = q.Do(ctx, func(tx *sql.Tx) error { return nil })
+		}()
+	}
+	deadline := time.Now().Add(2 * time.Second)
+	for q.Len() < buf+1 && time.Now().Before(deadline) {
+		time.Sleep(2 * time.Millisecond)
+	}
+
+	drainCtx, cancel := context.WithTimeout(context.Background(), 80*time.Millisecond)
+	defer cancel()
+	start := time.Now()
+	err := q.Drain(drainCtx)
+	elapsed := time.Since(start)
+	if !errors.Is(err, context.DeadlineExceeded) {
+		t.Fatalf("Drain err=%v want deadline", err)
+	}
+	if elapsed > 500*time.Millisecond {
+		t.Fatalf("Drain took %s, should return on timeout", elapsed)
+	}
+
+	close(hold)
+	wg.Wait()
+}

sdk/go/client_test.go

@@ -156,6 +156,66 @@ func TestFrameTooLargeLocal(t *testing.T) {
 	}
 }
 
+func TestPublishUpFailRetriesWithNewRID(t *testing.T) {
+	fake := NewFakeTransport()
+	fake.FailSendPublishN(1)
+	c := connectFake(t, fake)
+	defer c.Close()
+
+	at := time.UnixMilli(1_700_000_000_000)
+	fixedID := "msg-publish-retry"
+	done := make(chan struct{})
+	var firstRID, secondRID string
+
+	go func() {
+		defer close(done)
+		deadline := time.Now().Add(3 * time.Second)
+		for time.Now().Before(deadline) {
+			sends := fake.FindUp("send")
+			if len(sends) < 2 {
+				time.Sleep(5 * time.Millisecond)
+				continue
+			}
+			first, second := sends[0], sends[1]
+			firstRID, _ = first["rid"].(string)
+			secondRID, _ = second["rid"].(string)
+			if first["id"] != fixedID || second["id"] != fixedID {
+				t.Errorf("id changed: %v -> %v", first["id"], second["id"])
+			}
+			if first["send_at_ms"] != second["send_at_ms"] {
+				t.Errorf("send_at_ms changed: %v -> %v", first["send_at_ms"], second["send_at_ms"])
+			}
+			body1, _ := json.Marshal(first["body"])
+			body2, _ := json.Marshal(second["body"])
+			if string(body1) != string(body2) {
+				t.Errorf("body changed: %s -> %s", body1, body2)
+			}
+			if firstRID == "" || firstRID == secondRID {
+				t.Errorf("rid not regenerated: %q -> %q", firstRID, secondRID)
+			}
+			fake.ReplyOK(secondRID, map[string]any{
+				"id": fixedID, "send_at_ms": second["send_at_ms"], "state": "scheduled",
+			})
+			return
+		}
+		t.Error("timed out waiting for send retry")
+	}()
+
+	ctx, cancel := context.WithTimeout(context.Background(), 3*time.Second)
+	defer cancel()
+	res, err := c.Send(ctx, Target{Kind: "endpoint", ID: "b"}, Body{Enc: "utf8", Data: "hi"}, SendOptions{ID: fixedID, SendAt: &at})
+	<-done
+	if err != nil {
+		t.Fatalf("Send hung or failed: %v", err)
+	}
+	if res.ID != fixedID {
+		t.Fatalf("result id=%q want %q", res.ID, fixedID)
+	}
+	if firstRID == "" || secondRID == "" || firstRID == secondRID {
+		t.Fatalf("rids=%q/%q", firstRID, secondRID)
+	}
+}
+
 func TestResendKeepsIDAndSendAt(t *testing.T) {
 	fake := NewFakeTransport()
 	c := connectFake(t, fake)

sdk/go/fake_transport_test.go

@@ -3,6 +3,7 @@ package nixmsg
 import (
 	"context"
 	"encoding/json"
+	"errors"
 	"sync"
 	"sync/atomic"
 	"time"
@@ -29,6 +30,8 @@ type FakeTransport struct {
 	MaxFrameBytes     int
 	HelloDelay        time.Duration
 	ReceiveMaximumSet bool
+	// failSendLeft 接下来若干次 type=send 的 PublishUp 返回错误（仍记入 up）。
+	failSendLeft int
 }
 
 type fakeConnect struct {
@@ -65,6 +68,13 @@ func (f *FakeTransport) Start(ctx context.Context, cfg transportConfig) error {
 	return nil
 }
 
+// FailSendPublishN 让接下来 n 次 send 帧 PublishUp 失败（hello 等其它类型不受影响）。
+func (f *FakeTransport) FailSendPublishN(n int) {
+	f.mu.Lock()
+	f.failSendLeft = n
+	f.mu.Unlock()
+}
+
 func (f *FakeTransport) PublishUp(payload []byte) error {
 	f.mu.Lock()
 	cp := append([]byte(nil), payload...)
@@ -80,6 +90,17 @@ func (f *FakeTransport) PublishUp(payload []byte) error {
 	if auto && head.Type == "hello" && head.RID != "" {
 		f.replyHello(head.RID)
 	}
+	if head.Type == "send" {
+		f.mu.Lock()
+		fail := f.failSendLeft > 0
+		if fail {
+			f.failSendLeft--
+		}
+		f.mu.Unlock()
+		if fail {
+			return errors.New("publish failed")
+		}
+	}
 	return nil
 }
 

sdk/go/send.go

@@ -202,7 +202,10 @@ func (c *Client) dispatchSend(tr transport, item *sendItem, rid string, payload
 	if err := tr.PublishUp(payload); err != nil {
 		c.mu.Lock()
 		delete(c.pending, rid)
-		if item.epoch == epoch && item.inflight {
+		if item.epoch != epoch || !item.inflight {
+			c.mu.Unlock()
+			return
+		}
 		item.inflight = false
 		if c.inflight > 0 {
 			c.inflight--
@@ -213,8 +216,10 @@ func (c *Client) dispatchSend(tr transport, item *sendItem, rid string, payload
 			c.finishSend(item, SendResult{}, errStop)
 			return
 		}
-		}
+		// 保留 id/body/send_at_ms，换新 rid 后继续泵，避免 Send 永久挂起。
+		c.regenerateSendLocked(item)
 		c.mu.Unlock()
+		c.drainSendQueue()
 		return
 	}
 

sdk/java/src/main/java/asia/asio/nixmsg/Client.java

@@ -143,6 +143,12 @@ public final class Client {
         return lastStopCode == null ? "" : lastStopCode;
     }
 
+    int inflightSendsForTest() {
+        synchronized (lock) {
+            return inflightSends;
+        }
+    }
+
     void setSendQueueLimitForTest(int n) {
         sendQueueLimit = n;
     }
@@ -211,6 +217,19 @@ public final class Client {
             }
         }
         if (handshakeError != null) {
+            synchronized (lock) {
+                stopReconnect = true;
+                wantConnected = false;
+                if (lastStopCode == null || lastStopCode.isEmpty()) {
+                    if (handshakeError instanceof NixMsgException) {
+                        lastStopCode = ((NixMsgException) handshakeError).getCode();
+                        lastStopErr = (NixMsgException) handshakeError;
+                    } else {
+                        lastStopCode = "not_connected";
+                        lastStopErr = new NixMsgException("not_connected", handshakeError.getMessage());
+                    }
+                }
+            }
             if (handshakeError instanceof NixMsgException) {
                 throw (NixMsgException) handshakeError;
             }
@@ -229,6 +248,8 @@ public final class Client {
             synchronized (lock) {
                 stopReconnect = true;
                 wantConnected = false;
+                lastStopCode = "not_connected";
+                lastStopErr = new NixMsgException("not_connected", "连接未成功: " + state);
             }
             throw new NixMsgException("not_connected", "连接未成功: " + state);
         }
@@ -704,7 +725,9 @@ public final class Client {
             transport.publish(Protocol.upTopic(endpointId), Protocol.dumps(hello));
             await(p.future, connectTimeoutMs);
             if (p.error != null) {
-                throw p.error instanceof RuntimeException ? (RuntimeException) p.error : new NixMsgException("busy", p.error.getMessage());
+                throw p.error instanceof RuntimeException
+                        ? (RuntimeException) p.error
+                        : new NixMsgException("not_connected", p.error.getMessage());
             }
             if (p.response == null) {
                 throw new NixMsgException("not_connected", "握手超时");
@@ -883,9 +906,9 @@ public final class Client {
             if (!Boolean.TRUE.equals(frame.get("ok"))) {
                 Map<String, Object> err = asMap(frame.get("error"));
                 if ("rate_limited".equals(str(err.get("code"), ""))) {
+                    // 在途计数已在上方减过一次；只换新 rid 再排队，勿再减。
                     synchronized (lock) {
                         pending.remove(rid);
-                        inflightSends = Math.max(0, inflightSends - 1);
                         p.rid = "";
                         p.response = null;
                         p.error = null;

sdk/java/src/test/java/asia/asio/nixmsg/K00Test.java

@@ -41,7 +41,7 @@ public class K00Test {
     }
 
     @Test
-    public void testK00FirstConnectTimeout() {
+    public void testK00FirstConnectTimeout() throws Exception {
         FakeTransport tr = new FakeTransport();
         tr.autoHello = null;
         client = new Client(tr, true, Types.DEFAULT_MAX_FRAME, Types.CLIENT_NAME, 150);
@@ -51,6 +51,9 @@ public class K00Test {
         } catch (NixMsgException e) {
             assertEquals("not_connected", e.getCode());
         }
+        int n = tr.connects.size();
+        Thread.sleep(350);
+        assertEquals("首次握手失败后不得自动重连", n, tr.connects.size());
         tr.autoHello = new LinkedHashMap<String, Object>();
         tr.autoHello.put("server_time_ms", 1750000000000L);
         tr.autoHello.put("server_version", "0.1.0");
@@ -63,6 +66,108 @@ public class K00Test {
         tr.autoHello.put("session_token", "nst_test_token");
         client.connectSync("ws://example.test/mqtt", "ep1", "p", null, false);
         assertEquals(ConnectionState.ONLINE, client.getState());
+        assertTrue(tr.connects.size() > n);
+    }
+
+    @Test
+    public void testK00RateLimitedInflightOnce() throws Exception {
+        Types.ReconnectBackoff.disableJitterForTest();
+        FakeTransport tr = new FakeTransport();
+        tr.autoSendOk = false;
+        final List<String> holdRids = new CopyOnWriteArrayList<String>();
+        final List<String> limitedRids = new CopyOnWriteArrayList<String>();
+        tr.onUp(new java.util.function.Function<Map<String, Object>, Map<String, Object>>() {
+            @Override
+            public Map<String, Object> apply(Map<String, Object> frame) {
+                if (!"send".equals(String.valueOf(frame.get("type")))) {
+                    return null;
+                }
+                String rid = String.valueOf(frame.get("rid"));
+                String id = String.valueOf(frame.get("id"));
+                if ("hold".equals(id)) {
+                    holdRids.add(rid);
+                    return null; // 保持在途
+                }
+                limitedRids.add(rid);
+                if (limitedRids.size() == 1) {
+                    Map<String, Object> err = new LinkedHashMap<String, Object>();
+                    err.put("code", "rate_limited");
+                    err.put("message", "slow");
+                    Map<String, Object> resp = new LinkedHashMap<String, Object>();
+                    resp.put("v", 1);
+                    resp.put("type", "resp");
+                    resp.put("rid", rid);
+                    resp.put("ok", false);
+                    resp.put("error", err);
+                    return resp;
+                }
+                Map<String, Object> data = new LinkedHashMap<String, Object>();
+                data.put("id", frame.get("id"));
+                data.put("send_at_ms", frame.get("send_at_ms"));
+                data.put("state", "scheduled");
+                Map<String, Object> resp = new LinkedHashMap<String, Object>();
+                resp.put("v", 1);
+                resp.put("type", "resp");
+                resp.put("rid", rid);
+                resp.put("ok", true);
+                resp.put("data", data);
+                return resp;
+            }
+        });
+        connectOnline(tr);
+        Thread holder = new Thread(new Runnable() {
+            @Override
+            public void run() {
+                try {
+                    SendOptions opt = new SendOptions();
+                    opt.messageId = "hold";
+                    opt.sendAtMs = 1700000000001L;
+                    client.sendSync(new Target("endpoint", "ep2"), new Body("hold"), opt);
+                } catch (Exception ignored) {
+                }
+            }
+        });
+        holder.setDaemon(true);
+        holder.start();
+        long deadline = System.currentTimeMillis() + 2000;
+        while (holdRids.isEmpty() && System.currentTimeMillis() < deadline) {
+            Thread.sleep(10);
+        }
+        assertFalse(holdRids.isEmpty());
+        assertEquals(1, client.inflightSendsForTest());
+        final SendOptions opt = new SendOptions();
+        opt.messageId = "lim";
+        opt.sendAtMs = 1700000000002L;
+        Thread sender = new Thread(new Runnable() {
+            @Override
+            public void run() {
+                try {
+                    client.sendSync(new Target("endpoint", "ep2"), new Body("lim"), opt);
+                } catch (Exception ignored) {
+                }
+            }
+        });
+        sender.setDaemon(true);
+        sender.start();
+        deadline = System.currentTimeMillis() + 3000;
+        while (limitedRids.size() < 1 && System.currentTimeMillis() < deadline) {
+            Thread.sleep(10);
+        }
+        assertTrue(limitedRids.size() >= 1);
+        // rate_limited 后只应减 1：仍剩 hold 那一条在途
+        deadline = System.currentTimeMillis() + 500;
+        int seen = -1;
+        while (System.currentTimeMillis() < deadline) {
+            seen = client.inflightSendsForTest();
+            if (seen == 1) {
+                break;
+            }
+            Thread.sleep(10);
+        }
+        assertEquals("rate_limited 后在途应只减 1", 1, seen);
+        client.close();
+        holder.join(1000);
+        sender.join(1000);
     }
 
     @Test

sdk/js/src/client.ts

@@ -739,10 +739,12 @@ export class Client {
           this.finishSendErr(item, this.stopErr());
           return;
         }
+        // 传输失败但连接未断：换 rid 后立刻再泵，否则 hello 不再走、send 会挂起
         item.inflight = false;
         this.inflight = Math.max(0, this.inflight - 1);
         this.pending.delete(rid);
         this.regenerateSend(item);
+        void this.drainSendQueue();
         return;
       }
       this.finishSendErr(item, e);

sdk/js/test/client.test.ts

@@ -126,6 +126,46 @@ describe("nixmsg sdk", () => {
     await c.close();
   }, 10000);
 
+  it("publishUp transport fail retries with new rid", async () => {
+    const fake = new FakeTransport();
+    const c = await connectFake(fake);
+    const at = new Date(1_700_000_000_000);
+    const attempts: Array<Record<string, unknown>> = [];
+    let failOnce = true;
+    fake.publishUpImpl = async (s) => {
+      const m = JSON.parse(s) as Record<string, unknown>;
+      if (m.type !== "send") return;
+      attempts.push(m);
+      if (failOnce) {
+        failOnce = false;
+        throw new Error("transient publish");
+      }
+    };
+    const timer = setInterval(() => {
+      const sends = fake.findUp("send");
+      if (sends.length < 1) return;
+      const last = sends[sends.length - 1]!;
+      fake.replyOK(String(last.rid), {
+        id: last.id,
+        send_at_ms: last.send_at_ms,
+        state: "scheduled",
+      });
+    }, 5);
+    const res = await c.send(
+      { kind: "endpoint", id: "b" },
+      { enc: "utf8", data: "hi" },
+      { sendAt: at },
+    );
+    clearInterval(timer);
+    expect(attempts.length).toBeGreaterThanOrEqual(2);
+    expect(String(attempts[1]!.rid)).not.toBe(String(attempts[0]!.rid));
+    expect(attempts[1]!.id).toBe(attempts[0]!.id);
+    expect(attempts[1]!.send_at_ms).toBe(attempts[0]!.send_at_ms);
+    expect((attempts[1]!.body as { data: string }).data).toBe("hi");
+    expect(res.id).toBe(String(attempts[0]!.id));
+    await c.close();
+  }, 5000);
+
   it("register HTTP from ws url", async () => {
     const srv = createServer((req, res) => {
       expect(req.url).toBe("/api/client/register");

sdk/python/src/nixmsg/client.py

@@ -227,6 +227,15 @@ class Client:
                 raise NixMsgError("not_connected", "连接超时")
             err = self._handshake_error
             if err:
+                with self._lock:
+                    self._stop_reconnect = True
+                    self._want_connected = False
+                    if not self._last_stop_code:
+                        code = getattr(err, "code", None) or "not_connected"
+                        self._last_stop_code = str(code)
+                        self._last_stop_err = err if isinstance(err, NixMsgError) else NixMsgError(
+                            "not_connected", str(err)
+                        )
                 raise err
             if self._state not in (ConnectionState.ONLINE,):
                 if self._state == ConnectionState.AUTH_FAILED:
@@ -236,6 +245,11 @@ class Client:
                     )
                 if self._state == ConnectionState.KICKED:
                     raise NixMsgError("taken_over", "会话被接管")
+                with self._lock:
+                    self._stop_reconnect = True
+                    self._want_connected = False
+                    self._last_stop_code = "not_connected"
+                    self._last_stop_err = NixMsgError("not_connected", f"连接未成功: {self._state.value}")
                 raise NixMsgError("not_connected", f"连接未成功: {self._state.value}")
 
     def close(self) -> None:
@@ -576,7 +590,7 @@ class Client:
             except Exception:
                 pass
             with self._lock:
-                self._handshake_error = NixMsgError("busy", "连接超时")
+                self._handshake_error = NixMsgError("not_connected", "连接超时")
             return
 
     def _on_transport_connected(self) -> None:
@@ -598,7 +612,7 @@ class Client:
                 self._pending[rid] = pending
             self._transport.publish(up_topic(self._endpoint_id), dumps(hello))
             if not pending.event.wait(self._connect_timeout_s):
-                raise NixMsgError("busy", "握手超时")
+                raise NixMsgError("not_connected", "握手超时")
             if pending.error:
                 raise pending.error
             assert pending.response is not None

sdk/python/src/nixmsg/transport.py

@@ -8,16 +8,23 @@ from dataclasses import dataclass, field
 from typing import Any, Callable, Optional, Protocol
 from urllib.parse import urlparse
 
-from paho.mqtt.client import CallbackAPIVersion, Client as PahoClient, MQTT_ERR_SUCCESS, MQTTv5
-from paho.mqtt.enums import MQTTErrorCode
-from paho.mqtt.reasoncodes import ReasonCode
-
 DownHandler = Callable[[bytes], None]
 ConnHandler = Callable[[], None]
 DiscHandler = Callable[[Optional[str], bool], None]
 # reason_code_str, stop_reconnect
 
 
+def _require_paho():
+    """真实 MQTT 路径才加载 paho；假传输单测不依赖。"""
+    try:
+        from paho.mqtt.client import CallbackAPIVersion, Client as PahoClient, MQTT_ERR_SUCCESS, MQTTv5
+        from paho.mqtt.enums import MQTTErrorCode
+        from paho.mqtt.reasoncodes import ReasonCode
+    except ImportError as e:
+        raise ImportError("需要 paho-mqtt>=2.0（真实 MQTT 连接）") from e
+    return CallbackAPIVersion, PahoClient, MQTT_ERR_SUCCESS, MQTTv5, MQTTErrorCode, ReasonCode
+
+
 @dataclass
 class ConnectParams:
     url: str
@@ -180,7 +187,7 @@ class PahoTransport:
     """paho-mqtt 2.x CallbackAPIVersion.VERSION2。"""
 
     def __init__(self) -> None:
-        self._client: Optional[PahoClient] = None
+        self._client: Any = None
         self._on_connected: Optional[ConnHandler] = None
         self._on_disconnected: Optional[DiscHandler] = None
         self._on_down: Optional[DownHandler] = None
@@ -200,6 +207,7 @@ class PahoTransport:
         self._on_down = on_down
 
     def connect(self, params: ConnectParams) -> None:
+        CallbackAPIVersion, PahoClient, _, MQTTv5, _, _ = _require_paho()
         self.disconnect()
         url = params.url
         u = urlparse(url if "://" in url else "ws://" + url)
@@ -261,6 +269,7 @@ class PahoTransport:
         # 等待连接结果由回调驱动；超时由 Client 层处理
 
     def subscribe(self, topic: str) -> None:
+        _, _, MQTT_ERR_SUCCESS, _, _, _ = _require_paho()
         self._down_topic = topic
         if not self._client:
             return
@@ -273,6 +282,7 @@ class PahoTransport:
             raise RuntimeError("subscribe timeout")
 
     def publish(self, topic: str, payload: bytes) -> None:
+        _, _, MQTT_ERR_SUCCESS, _, _, _ = _require_paho()
         if not self._client:
             raise RuntimeError("not connected")
         info = self._client.publish(topic, payload, qos=1)
@@ -338,9 +348,14 @@ def _reason_to_int(reason_code) -> Optional[int]:
         return None
     if isinstance(reason_code, int):
         return reason_code
-    if isinstance(reason_code, ReasonCode):
+    try:
+        _, _, _, _, MQTTErrorCode, ReasonCode = _require_paho()
+    except ImportError:
+        MQTTErrorCode = ()  # type: ignore[assignment,misc]
+        ReasonCode = ()  # type: ignore[assignment,misc]
+    if ReasonCode and isinstance(reason_code, ReasonCode):
         return int(reason_code.value)
-    if isinstance(reason_code, MQTTErrorCode):
+    if MQTTErrorCode and isinstance(reason_code, MQTTErrorCode):
         return int(reason_code)
     # paho 偶发其它包装
     val = getattr(reason_code, "value", None)

sdk/python/tests/test_k00.py

@@ -30,11 +30,42 @@ class K00Tests(unittest.TestCase):
         with self.assertRaises(NixMsgError) as cm:
             c.connect("ws://example.test/mqtt", "ep1", password="p", wait=True)
         self.assertEqual(cm.exception.code, "not_connected")
+        n = len(tr.connects)
+        time.sleep(0.35)
+        self.assertEqual(len(tr.connects), n, "首次失败后不得自动重连")
         tr.auto_accept = True
         c.connect("ws://example.test/mqtt", "ep1", password="p", wait=True)
         self.assertEqual(c.state, ConnectionState.ONLINE)
         c.close()
 
+    def test_k00_first_hello_fail_stops_reconnect(self) -> None:
+        """MQTT 已通但 hello 失败：connect 返回未连接，且后台不再连。"""
+        tr = FakeTransport()
+        tr.auto_hello = None
+        c = Client(transport=tr, connect_timeout_s=0.2)
+        with self.assertRaises(NixMsgError) as cm:
+            c.connect("ws://example.test/mqtt", "ep1", password="p", wait=True)
+        self.assertEqual(cm.exception.code, "not_connected")
+        n = len(tr.connects)
+        self.assertGreaterEqual(n, 1)
+        time.sleep(0.45)
+        self.assertEqual(len(tr.connects), n, "hello 失败后不得自动重连")
+        tr.auto_hello = {
+            "server_time_ms": 1_750_000_000_000,
+            "server_version": "0.1.0",
+            "max_body_bytes": 262144,
+            "max_meta_bytes": 4096,
+            "max_frame_bytes": 786432,
+            "max_ttl_seconds": 2592000,
+            "max_schedule_seconds": 31536000,
+            "ack_timeout_seconds": 300,
+            "session_token": "nst_retry",
+        }
+        c.connect("ws://example.test/mqtt", "ep1", password="p", wait=True)
+        self.assertEqual(c.state, ConnectionState.ONLINE)
+        self.assertGreater(len(tr.connects), n)
+        c.close()
+
     def test_k00_taken_over_reason(self) -> None:
         c, tr = self._connect()
         got = []

web/src/api/admin.ts

@@ -296,7 +296,7 @@ export function deleteGroup(id: string) {
 
 export function addGroupMembers(id: string, memberIds: string[]) {
   return run(() =>
-    requestAdmin<{ failed: GroupFailed[] }>(
+    requestAdmin<{ failed: GroupFailed[]; added: number }>(
       `/api/admin/groups/${encodeURIComponent(id)}/members`,
       {
         method: "POST",

web/src/api/mock.ts

@@ -691,13 +691,14 @@ export const mockApi = {
     return {};
   },
 
-  async addGroupMembers(id: string, memberIds: string[]): Promise<{ failed: GroupFailed[] }> {
+  async addGroupMembers(id: string, memberIds: string[]): Promise<{ failed: GroupFailed[]; added: number }> {
     requireSession();
     const g = groups.find((x) => x.id === id);
     if (!g) {
       throw new ApiError("not_found", "群不存在", 404);
     }
     const failed: GroupFailed[] = [];
+    let added = 0;
     for (const mid of memberIds) {
       if (!endpoints.some((e) => e.id === mid)) {
         failed.push({ id: mid, code: "not_found" });
@@ -705,9 +706,10 @@ export const mockApi = {
       }
       if (!g.members.some((m) => m.id === mid)) {
         g.members.push({ id: mid, joined_at_ms: now() });
+        added++;
       }
     }
-    return { failed };
+    return { failed, added };
   },
 
   async removeGroupMember(id: string, endpointId: string): Promise<Record<string, never>> {

web/src/views/GroupsView.vue

@@ -233,9 +233,15 @@ async function onAddMembers() {
   if (!ids.length) return;
   const res = await addGroupMembers(detail.value.id, ids);
   if (res.failed.length) {
-    message.warning(`部分失败：${formatFailed(res.failed)}`);
+    if (res.added > 0) {
+      message.warning(`已加入 ${res.added} 人，部分失败：${formatFailed(res.failed)}`);
     } else {
-    message.success("已加人");
+      message.warning(`部分失败：${formatFailed(res.failed)}`);
+    }
+  } else if (res.added === 0) {
+    message.info("没有新成员加入");
+  } else {
+    message.success(`已加人 ${res.added} 人`);
   }
   await loadDetail(detail.value.id);
   await load();