fix: 修复 dedicated bulk attach 竞态并优化 short write 补写路径

- 客户端 dedicated attach 回复改为精确读取单帧，避免 attach reply 与后续 NBR1 数据粘连后被误解析
  - 服务端 accepted attach 改为先 detach transport，再直接回 attach reply，随后立即切入 dedicated bulk read loop
  - transport 读循环在 stop 或 transport ownership 失效后不再继续上推已读数据，避免 handoff 后首包被旧 reader 吃掉
  - dedicated bulk record 写路径改为 full-write，消除 short write 导致的 invalid bulk fast payload
  - 优化 vectored write 补写策略：先尝试一次 writev，未写完时直接顺序补完剩余 buffers，减少重复 WriteTo 开销
  - 放宽 vectored write 能力识别，支持通过 UnwrapConn/WriteBuffers 命中 fast path
  - 修复 dedicated batch 排队路径 payload 复用问题，改为深拷贝 queued items
  - 补齐 dedicated attach、short write、payload clone、transport stop/handoff 等回归测试

client_conn_transport.go

@@ -69,6 +69,12 @@ func (c *LogicalConn) readFromTUTransportConnWithBuffer(conn net.Conn, data []by
 }
 
 func (c *LogicalConn) handleTUTransportReadResultWithSession(stopCtx context.Context, conn net.Conn, generation uint64, num int, data []byte, err error) bool {
+	if transportReadShouldStop(stopCtx) || !c.ownsTransportRead(conn, generation) {
+		if c.shouldCloseTransportOnStop(conn) {
+			_ = conn.Close()
+		}
+		return false
+	}
 	if err == os.ErrDeadlineExceeded {
 		if num != 0 {
 			c.pushServerOwnedTransportMessage(data[:num], conn, generation)
@@ -95,6 +101,30 @@ func (c *LogicalConn) handleTUTransportReadResultWithSession(stopCtx context.Con
 	return true
 }
 
+func transportReadShouldStop(stopCtx context.Context) bool {
+	select {
+	case <-sessionStopChan(stopCtx):
+		return true
+	default:
+		return false
+	}
+}
+
+func (c *LogicalConn) ownsTransportRead(conn net.Conn, generation uint64) bool {
+	if c == nil {
+		return false
+	}
+	rt := c.clientConnSessionRuntimeSnapshot()
+	if rt == nil || !rt.transportAttached || rt.transportGeneration != generation {
+		return false
+	}
+	current := rt.tuConn
+	if rt.transport != nil && rt.transport.connSnapshot() != nil {
+		current = rt.transport.connSnapshot()
+	}
+	return current == conn
+}
+
 func (c *LogicalConn) pushServerOwnedTransportMessage(data []byte, conn net.Conn, generation uint64) {
 	if c == nil || len(data) == 0 {
 		return
@@ -163,6 +193,12 @@ func (c *ClientConn) handleTUTransportReadResultWithSession(stopCtx context.Cont
 	if logical := c.LogicalConn(); logical != nil {
 		return logical.handleTUTransportReadResultWithSession(stopCtx, conn, generation, num, data, err)
 	}
+	if transportReadShouldStop(stopCtx) || !c.ownsTransportRead(conn, generation) {
+		if c.shouldCloseClientConnTransportOnStop(conn) {
+			_ = conn.Close()
+		}
+		return false
+	}
 	if err == os.ErrDeadlineExceeded {
 		if num != 0 {
 			c.pushServerOwnedTransportMessage(data[:num], conn, generation)
@@ -189,6 +225,21 @@ func (c *ClientConn) handleTUTransportReadResultWithSession(stopCtx context.Cont
 	return true
 }
 
+func (c *ClientConn) ownsTransportRead(conn net.Conn, generation uint64) bool {
+	if c == nil {
+		return false
+	}
+	rt := c.clientConnSessionRuntimeSnapshot()
+	if rt == nil || !rt.transportAttached || rt.transportGeneration != generation {
+		return false
+	}
+	current := rt.tuConn
+	if rt.transport != nil && rt.transport.connSnapshot() != nil {
+		current = rt.transport.connSnapshot()
+	}
+	return current == conn
+}
+
 func (c *ClientConn) pushServerOwnedTransportMessage(data []byte, conn net.Conn, generation uint64) {
 	if logical := c.LogicalConn(); logical != nil {
 		logical.pushServerOwnedTransportMessage(data, conn, generation)