fix: PublishUp 失败时换 rid 并继续泵发送队列

docs/DEVIATIONS.md

@@ -1719,12 +1719,11 @@ issue #3 未关闭，`feat/fix-3-downlink-deadlock` 未合入 `main`。下面是
    - 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 R3-05
+### 复审修复 R3-04
 
-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。
+- 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。

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

@@ -202,19 +202,24 @@ 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 {
-			item.inflight = false
-			if c.inflight > 0 {
-				c.inflight--
-			}
-			if c.stopReconnect {
-				errStop := c.stopErrLocked()
-				c.mu.Unlock()
-				c.finishSend(item, SendResult{}, errStop)
-				return
-			}
+		if item.epoch != epoch || !item.inflight {
+			c.mu.Unlock()
+			return
 		}
+		item.inflight = false
+		if c.inflight > 0 {
+			c.inflight--
+		}
+		if c.stopReconnect {
+			errStop := c.stopErrLocked()
+			c.mu.Unlock()
+			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/js/src/client.ts

@@ -739,12 +739,10 @@ 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

@@ -126,46 +126,6 @@ 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");