fix(gemini): prompt 被拦截只报缓存命中时 input_tokens 不再为负

- Vertex 拦截 prompt（promptFeedback.blockReason，如 PROHIBITED_CONTENT）时 usageMetadata
  可能只带 cachedContentTokenCount、缺 promptTokenCount，promptTokenCount − cachedContentTokenCount
  得到负的 input_tokens 与负费用
- extractGeminiUsage 与 antigravity 非流式 / 流式（累计用量、message_start）四处换算改为差值下限 0，
  缓存命中仍照上游所报计入 cache_read_input_tokens
- 补充表测试、原生流式拦截用例与 antigravity 两条路径用例

backend/internal/pkg/antigravity/response_transformer.go

@@ -277,11 +277,12 @@ func (p *NonStreamingProcessor) buildResponse(geminiResp *GeminiResponse, respon
 	}
 
 	// 注意：Gemini 的 promptTokenCount 包含 cachedContentTokenCount，
-	// 但 Claude 的 input_tokens 不包含 cache_read_input_tokens，需要减去
+	// 但 Claude 的 input_tokens 不包含 cache_read_input_tokens，需要减去。
+	// prompt 被拦截时上游可能只报 cachedContentTokenCount，差值下限为 0。
 	usage := ClaudeUsage{}
 	if geminiResp.UsageMetadata != nil {
 		cached := geminiResp.UsageMetadata.CachedContentTokenCount
-		usage.InputTokens = geminiResp.UsageMetadata.PromptTokenCount - cached
+		usage.InputTokens = max(geminiResp.UsageMetadata.PromptTokenCount-cached, 0)
 		usage.OutputTokens = geminiResp.UsageMetadata.CandidatesTokenCount + geminiResp.UsageMetadata.ThoughtsTokenCount
 		usage.CacheReadInputTokens = cached
 		usage.ImageOutputTokens = geminiResp.UsageMetadata.ImageOutputTokens()

backend/internal/pkg/antigravity/stream_transformer.go

@@ -109,10 +109,11 @@ func (p *StreamingProcessor) ProcessLine(line string) []byte {
 
 	// 更新 usage
 	// 注意：Gemini 的 promptTokenCount 包含 cachedContentTokenCount，
-	// 但 Claude 的 input_tokens 不包含 cache_read_input_tokens，需要减去
+	// 但 Claude 的 input_tokens 不包含 cache_read_input_tokens，需要减去。
+	// prompt 被拦截时上游可能只报 cachedContentTokenCount，差值下限为 0。
 	if geminiResp.UsageMetadata != nil {
 		cached := geminiResp.UsageMetadata.CachedContentTokenCount
-		p.inputTokens = geminiResp.UsageMetadata.PromptTokenCount - cached
+		p.inputTokens = max(geminiResp.UsageMetadata.PromptTokenCount-cached, 0)
 		p.outputTokens = geminiResp.UsageMetadata.CandidatesTokenCount + geminiResp.UsageMetadata.ThoughtsTokenCount
 		p.cacheReadTokens = cached
 		p.imageOutputTokens = geminiResp.UsageMetadata.ImageOutputTokens()
@@ -190,7 +191,7 @@ func (p *StreamingProcessor) emitMessageStart(v1Resp *V1InternalResponse) []byte
 	usage := ClaudeUsage{}
 	if v1Resp.Response.UsageMetadata != nil {
 		cached := v1Resp.Response.UsageMetadata.CachedContentTokenCount
-		usage.InputTokens = v1Resp.Response.UsageMetadata.PromptTokenCount - cached
+		usage.InputTokens = max(v1Resp.Response.UsageMetadata.PromptTokenCount-cached, 0)
 		usage.OutputTokens = v1Resp.Response.UsageMetadata.CandidatesTokenCount + v1Resp.Response.UsageMetadata.ThoughtsTokenCount
 		usage.CacheReadInputTokens = cached
 		usage.ImageOutputTokens = v1Resp.Response.UsageMetadata.ImageOutputTokens()

backend/internal/pkg/antigravity/usage_contract_test.go

@@ -38,3 +38,29 @@ func TestGeminiUsageMapping_NoCacheCreationTokens(t *testing.T) {
 		require.Zero(t, usage.CacheCreationInputTokens)
 	})
 }
+
+// usageMetadata 缺 promptTokenCount、只有 cachedContentTokenCount 时，input_tokens 下限为 0，
+// 缓存命中仍计入 cache_read_input_tokens。
+func TestGeminiUsageMapping_MissingPromptTokenCountKeepsInputNonNegative(t *testing.T) {
+	const geminiBody = `{"candidates":[{"content":{"parts":[{"text":"hi"}],"role":"model"},"finishReason":"STOP"}],` +
+		`"usageMetadata":{"cachedContentTokenCount":30}}`
+
+	t.Run("non-stream", func(t *testing.T) {
+		_, usage, err := TransformGeminiToClaude([]byte(geminiBody), "gemini-3.1-pro-preview")
+		require.NoError(t, err)
+		require.NotNil(t, usage)
+		require.Zero(t, usage.InputTokens)
+		require.Equal(t, 30, usage.CacheReadInputTokens)
+	})
+
+	t.Run("stream", func(t *testing.T) {
+		p := NewStreamingProcessor("gemini-3.1-pro-preview")
+		out := p.ProcessLine(`data: {"response":` + geminiBody + `}`)
+		require.True(t, strings.Contains(string(out), `"message_start"`))
+		require.NotContains(t, string(out), `"input_tokens":-`, "message_start 的 input_tokens 不应为负")
+		_, usage := p.Finish()
+		require.NotNil(t, usage)
+		require.Zero(t, usage.InputTokens)
+		require.Equal(t, 30, usage.CacheReadInputTokens)
+	})
+}

backend/internal/service/gemini_messages_compat_service.go

@@ -3036,9 +3036,10 @@ func extractGeminiUsage(data []byte) *ClaudeUsage {
 	}
 
 	// 注意：Gemini 的 promptTokenCount 包含 cachedContentTokenCount，
-	// 但 Claude 的 input_tokens 不包含 cache_read_input_tokens，需要减去
+	// 但 Claude 的 input_tokens 不包含 cache_read_input_tokens，需要减去。
+	// prompt 被拦截时上游可能只报 cachedContentTokenCount，差值下限为 0。
 	return &ClaudeUsage{
-		InputTokens:          prompt - cached,
+		InputTokens:          max(prompt-cached, 0),
 		OutputTokens:         cand + thoughts,
 		CacheReadInputTokens: cached,
 		ImageOutputTokens:    imageTokens,

backend/internal/service/gemini_messages_compat_service_test.go

@@ -947,6 +947,16 @@ func TestExtractGeminiUsage(t *testing.T) {
 				CacheReadInputTokens: 0,
 			},
 		},
+		{
+			name:    "prompt 被拦截只报 cachedContentTokenCount — input 下限为 0",
+			input:   `{"promptFeedback":{"blockReason":"PROHIBITED_CONTENT"},"usageMetadata":{"cachedContentTokenCount":12980}}`,
+			wantNil: false,
+			wantUsage: &ClaudeUsage{
+				InputTokens:          0,
+				OutputTokens:         0,
+				CacheReadInputTokens: 12980,
+			},
+		},
 		{
 			name:    "无 usageMetadata",
 			input:   `{"candidates":[]}`,

backend/internal/service/gemini_native_response_signal_test.go

@@ -180,6 +180,29 @@ func TestGeminiForwardNative_StreamPromptBlockedMarksContentPolicy(t *testing.T)
 	require.Empty(t, upstreamErrorEventsFromContext(t, c))
 }
 
+// Vertex 拦截 prompt 时 usageMetadata 可能只带 cachedContentTokenCount、缺 promptTokenCount：
+// input 下限为 0，缓存命中照上游所报计入 cache_read。
+func TestGeminiForwardNative_StreamPromptBlockedWithOnlyCachedTokensKeepsInputNonNegative(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+	body := `data: {"promptFeedback":{"blockReason":"PROHIBITED_CONTENT"},"usageMetadata":{"cachedContentTokenCount":12980,"totalTokenCount":0,"trafficType":"ON_DEMAND"},"modelVersion":"gemini-3.5-flash-lite"}
+
+`
+	svc := newGeminiSignalService("text/event-stream", body)
+	c, rec := newGeminiNativeTestContext(t)
+
+	result, err := svc.ForwardNative(context.Background(), c, geminiSignalTestAccount(),
+		"gemini-3.5-flash-lite", "streamGenerateContent", true, geminiSignalTestRequest())
+	require.NoError(t, err)
+	require.Equal(t, body, rec.Body.String())
+	require.Equal(t, 0, result.Usage.InputTokens)
+	require.Equal(t, 12980, result.Usage.CacheReadInputTokens)
+	require.Equal(t, 0, result.Usage.OutputTokens)
+
+	streamErrs := GetOpsStreamErrors(c)
+	require.Len(t, streamErrs, 1)
+	require.Equal(t, "PROHIBITED_CONTENT", streamErrs[0].Code)
+}
+
 func TestGeminiForwardNative_EmptyStreamMarksUpstreamFailure(t *testing.T) {
 	gin.SetMode(gin.TestMode)
 	svc := newGeminiSignalService("text/event-stream", "")