feat(gateway): harden failover and payload handling

Retry pre-response transport failures across candidates with an explicit stop policy, and propagate end-to-end timing into usage records and UI diagnostics.

Remove legacy body, import, cookie, PII, and tunnel replay caps while preserving optional operator-configured gateway limits.

frontend/src/features/usage/components/HorizontalRequestTimeline.vue

@@ -1023,25 +1023,26 @@ const conversionBoundaryIndex = computed(() => {
   return idx
 })
 
-// 计算链路总耗时（使用成功候选的 latency_ms 字段）
-// 优先使用 latency_ms，因为它与 Usage.response_time_ms 使用相同的时间基准
-// 避免 finished_at - started_at 带来的额外延迟（数据库操作时间）
+// The trace aggregate includes every attempted candidate, including failed
+// failover attempts. Per-candidate latency remains provider-scoped.
 const totalTraceLatency = computed(() => {
   if (!rawTimeline.value || rawTimeline.value.length === 0) return 0
 
-  // 查找成功的候选，使用其 latency_ms
-  const successCandidate = rawTimeline.value.find(c => c.status === 'success')
-  if (successCandidate?.latency_ms != null) {
-    return successCandidate.latency_ms
+  const aggregateLatency = trace.value?.total_latency_ms
+  if (typeof aggregateLatency === 'number' && Number.isFinite(aggregateLatency) && aggregateLatency > 0) {
+    return aggregateLatency
   }
 
-  // 如果没有成功的候选，查找失败但有 latency_ms 的候选
-  const failedWithLatency = rawTimeline.value.find(c => c.status === 'failed' && c.latency_ms != null)
-  if (failedWithLatency?.latency_ms != null) {
-    return failedWithLatency.latency_ms
+  const attemptedLatency = rawTimeline.value.reduce((sum, candidate) => {
+    const latency = normalizeLatencyMs(candidate.latency_ms)
+    return sum + (latency ?? 0)
+  }, 0)
+  if (attemptedLatency > 0) {
+    return attemptedLatency
   }
 
-  // 回退：使用 finished_at - started_at 计算
+  // Historical transport failures may not have latency_ms. Recover the wall
+  // clock span from candidate timestamps for those records.
   let earliestStart: number | null = null
   let latestEnd: number | null = null
 

frontend/src/features/usage/components/RequestDetailDrawer.vue

@@ -175,7 +175,7 @@
                       <span>
                         <span class="text-muted-foreground">耗时</span>
                         <span class="ml-1 font-bold">
-                          {{ formatDurationMs(detail.first_byte_time_ms) }} / {{ formatDurationMs(detail.response_time_ms) }}
+                          {{ formatDurationMs(detail.end_to_end_first_byte_time_ms ?? detail.first_byte_time_ms) }} / {{ formatDurationMs(detail.end_to_end_time_ms ?? detail.response_time_ms) }}
                         </span>
                       </span>
                       <span class="text-muted-foreground">|</span>
@@ -197,7 +197,7 @@
                       <span class="whitespace-nowrap">
                         <span class="text-muted-foreground">耗时</span>
                         <span class="ml-1 font-bold">
-                          {{ formatDurationMs(detail.first_byte_time_ms) }} / {{ formatDurationMs(detail.response_time_ms) }}
+                          {{ formatDurationMs(detail.end_to_end_first_byte_time_ms ?? detail.first_byte_time_ms) }} / {{ formatDurationMs(detail.end_to_end_time_ms ?? detail.response_time_ms) }}
                         </span>
                       </span>
                       <span class="text-muted-foreground">|</span>

frontend/src/features/usage/components/UsageRecordsTable.vue

@@ -358,7 +358,7 @@
               <span>{{ formatOutputRate(getRecordDisplayOutputRate(record)) }}</span>
             </template>
             <span
-              v-else-if="record.response_time_ms != null || record.first_byte_time_ms != null"
+              v-else-if="hasRecordDisplayLatency(record)"
               class="ml-1"
             >{{ formatRecordLatencyPair(record) }} / {{ formatOutputRate(getRecordDisplayOutputRate(record)) }}</span>
             <span
@@ -611,7 +611,8 @@
             v-if="isColumnVisible('performance')"
             class="h-12 font-semibold w-[9%] text-right"
           >
-            <div class="flex flex-col items-end text-xs gap-0.5">
+            <div class="flex flex-col items-end text-[11px] leading-3">
+              <span class="whitespace-nowrap">端到端</span>
               <span class="whitespace-nowrap">首字/总耗时</span>
               <span class="text-muted-foreground font-normal">输出速度</span>
             </div>
@@ -925,7 +926,7 @@
             </div>
             <!-- 已完成状态：首字 + 总耗时 -->
             <div
-              v-else-if="record.response_time_ms != null || record.first_byte_time_ms != null"
+              v-else-if="hasRecordDisplayLatency(record)"
               class="flex flex-col items-end text-xs gap-0.5"
               :title="getRecordPerformanceTitle(record)"
             >
@@ -1445,8 +1446,10 @@ function getRecordCacheTokensTitle(record: UsageRecord): string {
 }
 
 function formatRecordLatencyPair(record: UsageRecord): string {
-  const firstByte = formatRecordDurationSeconds(record.first_byte_time_ms)
-  const total = formatRecordDurationSeconds(record.response_time_ms)
+  const firstByte = formatRecordDurationSeconds(
+    record.end_to_end_first_byte_time_ms ?? record.first_byte_time_ms,
+  )
+  const total = formatRecordDurationSeconds(record.end_to_end_time_ms ?? record.response_time_ms)
   return `${firstByte} / ${total}`
 }
 
@@ -1455,6 +1458,13 @@ function formatRecordDurationSeconds(ms: number | null | undefined): string {
   return `${(ms / 1000).toFixed(2)}s`
 }
 
+function hasRecordDisplayLatency(record: UsageRecord): boolean {
+  return record.end_to_end_time_ms != null
+    || record.end_to_end_first_byte_time_ms != null
+    || record.response_time_ms != null
+    || record.first_byte_time_ms != null
+}
+
 function getRecordDisplayOutputRate(record: UsageRecord): number | null {
   return getDisplayOutputRate({
     output_tokens: record.output_tokens,
@@ -1468,8 +1478,10 @@ function getRecordDisplayOutputRate(record: UsageRecord): number | null {
 function getRecordPerformanceTitle(record: UsageRecord): string {
   const outputRate = getRecordDisplayOutputRate(record)
   return [
-    `首字: ${formatRecordDurationSeconds(record.first_byte_time_ms)}`,
-    `总耗时: ${formatRecordDurationSeconds(record.response_time_ms)}`,
+    `端到端首字: ${formatRecordDurationSeconds(record.end_to_end_first_byte_time_ms ?? record.first_byte_time_ms)}`,
+    `端到端总耗时: ${formatRecordDurationSeconds(record.end_to_end_time_ms ?? record.response_time_ms)}`,
+    `成功候选首字: ${formatRecordDurationSeconds(record.first_byte_time_ms)}`,
+    `成功候选耗时: ${formatRecordDurationSeconds(record.response_time_ms)}`,
     `生成耗时: ${formatRecordDurationSeconds(getGenerationTimeMs(record))}`,
     `输出速度: ${formatOutputRateTokensPerSecond(outputRate)}`,
   ].join('\n')

frontend/src/features/usage/components/__tests__/HorizontalRequestTimeline.spec.ts

@@ -181,6 +181,46 @@ afterEach(() => {
 })
 
 describe('HorizontalRequestTimeline', () => {
+  it('uses the trace aggregate latency instead of the successful candidate latency', async () => {
+    const trace = buildTrace([
+      buildCandidate({
+        id: 'cand-transport-timeout',
+        provider_id: 'provider-timeout',
+        provider_name: 'Provider Timeout',
+        key_id: 'key-timeout',
+        key_name: 'Timeout Key',
+        candidate_index: 0,
+        status: 'failed',
+        latency_ms: 10_000,
+        started_at: '2026-05-06T12:00:00.000Z',
+        finished_at: '2026-05-06T12:00:10.000Z',
+      }),
+      buildCandidate({
+        id: 'cand-success-after-failover',
+        provider_id: 'provider-success',
+        provider_name: 'Provider Success',
+        key_id: 'key-success',
+        key_name: 'Success Key',
+        candidate_index: 1,
+        status: 'success',
+        latency_ms: 626,
+        started_at: '2026-05-06T12:00:10.000Z',
+        finished_at: '2026-05-06T12:00:10.626Z',
+      }),
+    ])
+    trace.total_latency_ms = 10_626
+
+    const root = mountTimeline(trace)
+    await nextTick()
+
+    const heading = [...root.querySelectorAll('h4')]
+      .find(element => element.textContent?.trim() === '请求链路追踪')
+    const overview = heading?.parentElement?.parentElement
+    const displayedLatency = overview?.lastElementChild?.textContent?.trim()
+    expect(displayedLatency).toBe('10.63s')
+    expect(displayedLatency).not.toBe('626ms')
+  })
+
   it('keeps attempted keys visible for ordinary provider groups that are not selected', async () => {
     const trace = buildTrace([
       buildCandidate({

frontend/src/features/usage/components/__tests__/RequestDetailDrawer.pricing.spec.ts

@@ -113,6 +113,48 @@ function buildFastTierDetail(): RequestDetail {
 }
 
 describe('RequestDetailDrawer settlement pricing', () => {
+  it('shows end-to-end latency while keeping output TPS scoped to candidate timing', async () => {
+    apiMocks.getRequestDetail.mockResolvedValue({
+      ...buildEmbeddingDetail(),
+      tokens: { input: 100, output: 50, total: 150 },
+      input_tokens: 100,
+      output_tokens: 50,
+      total_tokens: 150,
+      is_stream: true,
+      upstream_is_stream: true,
+      response_time_ms: 626,
+      first_byte_time_ms: 100,
+      end_to_end_time_ms: 10_626,
+      end_to_end_first_byte_time_ms: 10_120,
+    } satisfies RequestDetail)
+
+    let isOpen!: Ref<boolean>
+    const Host = defineComponent({
+      setup() {
+        isOpen = ref(false)
+        return () => h(RequestDetailDrawer, {
+          isOpen: isOpen.value,
+          requestId: 'usage-embedding-1',
+        })
+      },
+    })
+
+    const root = document.createElement('div')
+    document.body.appendChild(root)
+    const app = createApp(Host)
+    app.mount(root)
+    mountedApps.push({ app, root })
+
+    isOpen.value = true
+    await nextTick()
+
+    await vi.waitFor(() => {
+      expect(document.body.textContent).toContain('10.12s / 10.63s')
+      expect(document.body.textContent).toContain('95.1tps')
+      expect(document.body.textContent).not.toContain('98.8tps')
+    })
+  })
+
   it('renders an input-only embedding tier without treating the missing output price as zero', async () => {
     apiMocks.getRequestDetail.mockResolvedValue(buildEmbeddingDetail())
 

frontend/src/features/usage/components/__tests__/UsageRecordsTable.spec.ts

@@ -195,8 +195,10 @@ describe('UsageRecordsTable', () => {
     const titles = [...root.querySelectorAll<HTMLElement>('[title]')]
       .map((element) => element.getAttribute('title'))
     expect(titles).toContain([
-      '首字: 0.50s',
-      '总耗时: 1.00s',
+      '端到端首字: 0.50s',
+      '端到端总耗时: 1.00s',
+      '成功候选首字: 0.50s',
+      '成功候选耗时: 1.00s',
       '生成耗时: 0.50s',
       '输出速度: 100 tokens/s',
     ].join('\n'))
@@ -204,6 +206,55 @@ describe('UsageRecordsTable', () => {
     expect(titles.join('\n')).not.toContain('首字后生成耗时')
   })
 
+  it('shows end-to-end latency while keeping output TPS scoped to the successful candidate', () => {
+    const root = mountUsageRecordsTable([buildRecord({
+      output_tokens: 50,
+      response_time_ms: 626,
+      first_byte_time_ms: 100,
+      end_to_end_time_ms: 10_626,
+      end_to_end_first_byte_time_ms: 10_120,
+    })])
+
+    const performanceCell = root.querySelector('table tbody tr td:last-child') as HTMLElement
+    expect(performanceCell.textContent).toContain('10.12s / 10.63s')
+    expect(performanceCell.textContent).toContain('95.1 tps')
+    expect(performanceCell.textContent).not.toContain('98.8 tps')
+
+    const titles = [...root.querySelectorAll<HTMLElement>('[title]')]
+      .map((element) => element.getAttribute('title'))
+    expect(titles).toContain([
+      '端到端首字: 10.12s',
+      '端到端总耗时: 10.63s',
+      '成功候选首字: 0.10s',
+      '成功候选耗时: 0.63s',
+      '生成耗时: 0.53s',
+      '输出速度: 95.1 tokens/s',
+    ].join('\n'))
+  })
+
+  it('shows end-to-end latency when candidate timing fields are unavailable', () => {
+    const root = mountUsageRecordsTable([buildRecord({
+      response_time_ms: null,
+      first_byte_time_ms: null,
+      end_to_end_time_ms: 10_626,
+      end_to_end_first_byte_time_ms: 10_120,
+    })])
+
+    const performanceCell = root.querySelector('table tbody tr td:last-child') as HTMLElement
+    expect(performanceCell.textContent).toContain('10.12s / 10.63s')
+
+    const titles = [...root.querySelectorAll<HTMLElement>('[title]')]
+      .map((element) => element.getAttribute('title'))
+    expect(titles).toContain([
+      '端到端首字: 10.12s',
+      '端到端总耗时: 10.63s',
+      '成功候选首字: -',
+      '成功候选耗时: -',
+      '生成耗时: -',
+      '输出速度: -',
+    ].join('\n'))
+  })
+
   it('shows an output speed placeholder when the rate is unavailable', () => {
     const root = mountUsageRecordsTable([buildRecord({
       output_tokens: 0,
@@ -218,8 +269,10 @@ describe('UsageRecordsTable', () => {
 
     const titles = [...root.querySelectorAll<HTMLElement>('[title]')].map((element) => element.title)
     expect(titles).toContain([
-      '首字: 0.50s',
-      '总耗时: 1.00s',
+      '端到端首字: 0.50s',
+      '端到端总耗时: 1.00s',
+      '成功候选首字: 0.50s',
+      '成功候选耗时: 1.00s',
       '生成耗时: 0.50s',
       '输出速度: -',
     ].join('\n'))

frontend/src/features/usage/types.ts

@@ -117,6 +117,8 @@ export interface UsageRecord {
   actual_cost?: number
   response_time_ms?: number | null
   first_byte_time_ms?: number | null  // 首字时间 (TTFB)
+  end_to_end_time_ms?: number | null  // 客户端从请求进入网关到完成的总耗时
+  end_to_end_first_byte_time_ms?: number | null  // 客户端从请求进入网关到首字节的耗时
   is_stream: boolean
   upstream_is_stream?: boolean
   client_requested_stream?: boolean