fix(openai): /v1/messages端点补齐Codex用量快照提取与错误透传规则

对齐/v1/responses的Forward方法，修复两处不一致：
- 成功响应时从响应头提取OAuth账号的Codex使用量数据
- 非failover错误场景下应用管理员配置的错误透传规则

backend/internal/handler/openai_gateway_handler.go

@@ -502,6 +502,11 @@ func (h *OpenAIGatewayHandler) Messages(c *gin.Context) {
 
 	setOpsRequestContext(c, reqModel, reqStream, body)
 
+	// 绑定错误透传服务，允许 service 层在非 failover 错误场景复用规则。
+	if h.errorPassthroughService != nil {
+		service.BindErrorPassthroughService(c, h.errorPassthroughService)
+	}
+
 	subscription, _ := middleware2.GetSubscriptionFromContext(c)
 
 	service.SetOpsLatencyMs(c, service.OpsAuthLatencyMsKey, time.Since(requestStart).Milliseconds())

backend/internal/service/openai_gateway_messages.go

@@ -143,14 +143,26 @@ func (s *OpenAIGatewayService) ForwardAsAnthropic(
 			return nil, &UpstreamFailoverError{StatusCode: resp.StatusCode, ResponseBody: respBody}
 		}
 		// Non-failover error: return Anthropic-formatted error to client
-		return s.handleAnthropicErrorResponse(resp, c)
+		return s.handleAnthropicErrorResponse(resp, c, account)
 	}
 
 	// 9. Handle normal response
+	var result *OpenAIForwardResult
+	var handleErr error
 	if isStream {
-		return s.handleAnthropicStreamingResponse(resp, c, originalModel, mappedModel, startTime)
+		result, handleErr = s.handleAnthropicStreamingResponse(resp, c, originalModel, mappedModel, startTime)
+	} else {
+		result, handleErr = s.handleAnthropicNonStreamingResponse(resp, c, originalModel, mappedModel, startTime)
 	}
-	return s.handleAnthropicNonStreamingResponse(resp, c, originalModel, mappedModel, startTime)
+
+	// Extract and save Codex usage snapshot from response headers (for OAuth accounts)
+	if handleErr == nil && account.Type == AccountTypeOAuth {
+		if snapshot := ParseCodexRateLimitHeaders(resp.Header); snapshot != nil {
+			s.updateCodexUsageSnapshot(ctx, account.ID, snapshot)
+		}
+	}
+
+	return result, handleErr
 }
 
 // handleAnthropicErrorResponse reads an upstream error and returns it in
@@ -158,6 +170,7 @@ func (s *OpenAIGatewayService) ForwardAsAnthropic(
 func (s *OpenAIGatewayService) handleAnthropicErrorResponse(
 	resp *http.Response,
 	c *gin.Context,
+	account *Account,
 ) (*OpenAIForwardResult, error) {
 	body, _ := io.ReadAll(io.LimitReader(resp.Body, 2<<20))
 
@@ -178,6 +191,21 @@ func (s *OpenAIGatewayService) handleAnthropicErrorResponse(
 	}
 	setOpsUpstreamError(c, resp.StatusCode, upstreamMsg, upstreamDetail)
 
+	// Apply error passthrough rules (matches handleErrorResponse pattern in openai_gateway_service.go)
+	if status, errType, errMsg, matched := applyErrorPassthroughRule(
+		c, account.Platform, resp.StatusCode, body,
+		http.StatusBadGateway, "api_error", "Upstream request failed",
+	); matched {
+		writeAnthropicError(c, status, errType, errMsg)
+		if upstreamMsg == "" {
+			upstreamMsg = errMsg
+		}
+		if upstreamMsg == "" {
+			return nil, fmt.Errorf("upstream error: %d (passthrough rule matched)", resp.StatusCode)
+		}
+		return nil, fmt.Errorf("upstream error: %d (passthrough rule matched) message=%s", resp.StatusCode, upstreamMsg)
+	}
+
 	errType := "api_error"
 	switch {
 	case resp.StatusCode == 400: