fix(gemini): stop cooling Vertex 429s until PST midnight, honor RetryInfo

Vertex AI reports 429s with a standard google.rpc.RetryInfo detail
(retryDelay like "39s"), but ParseGeminiRateLimitResetTime only knew the
AI Studio quotaResetDelay metadata, so parsing always failed for Vertex
bodies. The fallback then lumped service_account accounts in with
API-key/AI-Studio ones and rate limited them until the next PST
midnight, so a transient per-minute throttle could take a pay-as-you-go
account (and with it the whole group) offline for up to a day.

- Parse google.rpc.RetryInfo retryDelay from 429 bodies, capped at 15
  minutes to guard against absurd upstream values.
- When no reset time can be parsed, cool service_account (Vertex)
  accounts down for 60s instead of until PST midnight: pay-as-you-go
  quotas are short-window RPM/TPM with no daily reset.

Fallback behavior for Code Assist, Google One and API-key accounts is
unchanged, covered by new regression tests.

backend/internal/service/gemini_error_policy_test.go

@@ -606,6 +606,77 @@ func TestHandleGeminiUpstreamError_PoolMode429(t *testing.T) {
 	}
 }
 
+// Vertex（service_account）429 响应带 google.rpc.RetryInfo 时，应按 retryDelay 冷却，
+// 而不是回退到 PST 午夜。
+func TestHandleGeminiUpstreamError_VertexRetryInfoUsesParsedDelay(t *testing.T) {
+	repo := &rateLimit429AccountRepoStub{}
+	svc := &GeminiMessagesCompatService{accountRepo: repo}
+
+	account := &Account{
+		ID:       601,
+		Platform: PlatformGemini,
+		Type:     AccountTypeServiceAccount,
+		Credentials: map[string]any{
+			"project_id": "my-vertex-project",
+			"location":   "global",
+		},
+	}
+	body := []byte(`{"error":{"code":429,"message":"Resource exhausted. Please try again later.","status":"RESOURCE_EXHAUSTED","details":[{"@type":"type.googleapis.com/google.rpc.RetryInfo","retryDelay":"39s"}]}}`)
+
+	before := time.Now()
+	svc.handleGeminiUpstreamError(context.Background(), account, http.StatusTooManyRequests, http.Header{}, body)
+
+	require.Equal(t, 1, repo.rateLimitCalls)
+	require.Equal(t, int64(601), repo.lastRateLimitID)
+	require.WithinDuration(t, before.Add(39*time.Second), repo.lastRateLimitReset, 2*time.Second)
+}
+
+// Vertex（service_account）429 响应不带任何可解析的重置时间时，走短冷却兜底，
+// 不再冷却到 PST 午夜（按量付费没有每日配额，长冷却会导致整组不可用）。
+func TestHandleGeminiUpstreamError_VertexFallbackUsesShortCooldown(t *testing.T) {
+	repo := &rateLimit429AccountRepoStub{}
+	svc := &GeminiMessagesCompatService{accountRepo: repo}
+
+	account := &Account{
+		ID:       602,
+		Platform: PlatformGemini,
+		Type:     AccountTypeServiceAccount,
+		Credentials: map[string]any{
+			"project_id": "my-vertex-project",
+		},
+	}
+	body := []byte(`{"error":{"code":429,"message":"Resource exhausted. Please try again later.","status":"RESOURCE_EXHAUSTED"}}`)
+
+	before := time.Now()
+	svc.handleGeminiUpstreamError(context.Background(), account, http.StatusTooManyRequests, http.Header{}, body)
+
+	require.Equal(t, 1, repo.rateLimitCalls)
+	require.Equal(t, int64(602), repo.lastRateLimitID)
+	require.WithinDuration(t, before.Add(geminiVertexFallbackCooldown), repo.lastRateLimitReset, 2*time.Second)
+	// 显式守护：不应被冷却超过 15 分钟（更不应到次日 PST 午夜）
+	require.True(t, repo.lastRateLimitReset.Before(before.Add(15*time.Minute)))
+}
+
+// API Key（AI Studio）无法解析重置时间时仍冷却到 PST 午夜，行为保持不变。
+func TestHandleGeminiUpstreamError_APIKeyFallbackStillPSTMidnight(t *testing.T) {
+	repo := &rateLimit429AccountRepoStub{}
+	svc := &GeminiMessagesCompatService{accountRepo: repo}
+
+	account := &Account{
+		ID:       603,
+		Platform: PlatformGemini,
+		Type:     AccountTypeAPIKey,
+	}
+	body := []byte(`{"error":{"code":429,"message":"rate limit"}}`)
+
+	svc.handleGeminiUpstreamError(context.Background(), account, http.StatusTooManyRequests, http.Header{}, body)
+
+	require.Equal(t, 1, repo.rateLimitCalls)
+	require.Equal(t, int64(603), repo.lastRateLimitID)
+	expected := time.Unix(*nextGeminiDailyResetUnix(), 0)
+	require.WithinDuration(t, expected, repo.lastRateLimitReset, 5*time.Second)
+}
+
 type geminiErrorPolicyRepo struct {
 	mockAccountRepoForGemini
 	setErrorCalls            int

backend/internal/service/gemini_messages_compat_service.go

@@ -40,6 +40,16 @@ const (
 	geminiRetryMaxDelay  = 16 * time.Second
 )
 
+const (
+	// google.rpc.RetryInfo 的标准 @type（Vertex AI 429 响应携带，retryDelay 形如 "39s"）
+	geminiRetryInfoTypeURL = "type.googleapis.com/google.rpc.RetryInfo"
+	// RetryInfo.retryDelay 的上限，防止异常值导致账号长时间不可用
+	geminiRetryInfoMaxDelay = 15 * time.Minute
+	// Vertex（service_account，按量付费）429 无法解析重置时间时的兜底冷却：
+	// 按量付费限速是短窗口 RPM/TPM，没有每日配额，不适用 PST 午夜重置
+	geminiVertexFallbackCooldown = time.Minute
+)
+
 // Gemini tool calling now requires `thoughtSignature` in parts that include `functionCall`.
 // Many clients don't send it; we inject a known dummy signature to satisfy the validator.
 // Ref: https://ai.google.dev/gemini-api/docs/thought-signatures
@@ -3140,6 +3150,10 @@ func (s *GeminiMessagesCompatService) handleGeminiUpstreamError(ctx context.Cont
 			} else {
 				logger.LegacyPrintf("service.gemini_messages_compat", "[Gemini 429] Account %d (Google One OAuth, tier=%s, project=%s) rate limited, cooldown=%v", account.ID, tierID, projectID, time.Until(ra).Truncate(time.Second))
 			}
+		} else if account.IsVertexServiceAccount() {
+			// Vertex 按量付费：短窗限速，无每日配额，短冷却后即可恢复
+			ra = time.Now().Add(geminiVertexFallbackCooldown)
+			logger.LegacyPrintf("service.gemini_messages_compat", "[Gemini 429] Account %d (Vertex service account, project=%s) rate limited, cooldown=%v", account.ID, projectID, time.Until(ra).Truncate(time.Second))
 		} else {
 			// API Key / AI Studio OAuth: PST 午夜
 			if ts := nextGeminiDailyResetUnix(); ts != nil {
@@ -3172,19 +3186,29 @@ func ParseGeminiRateLimitResetTime(body []byte) *int64 {
 		}
 	}
 
-	// 遍历 error.details 查找 quotaResetDelay
+	// 遍历 error.details 查找 quotaResetDelay（AI Studio）或 RetryInfo.retryDelay（Vertex）
 	var found *int64
 	gjson.GetBytes(body, "error.details").ForEach(func(_, detail gjson.Result) bool {
-		v := detail.Get("metadata.quotaResetDelay").String()
-		if v == "" {
+		if v := detail.Get("metadata.quotaResetDelay").String(); v != "" {
+			if dur, err := time.ParseDuration(v); err == nil {
+				// Use ceil to avoid undercounting fractional seconds (e.g. 10.1s should not become 10s),
+				// which can affect scheduling decisions around thresholds (like 10s).
+				ts := time.Now().Unix() + int64(math.Ceil(dur.Seconds()))
+				found = &ts
+				return false
+			}
 			return true
 		}
-		if dur, err := time.ParseDuration(v); err == nil {
-			// Use ceil to avoid undercounting fractional seconds (e.g. 10.1s should not become 10s),
-			// which can affect scheduling decisions around thresholds (like 10s).
-			ts := time.Now().Unix() + int64(math.Ceil(dur.Seconds()))
-			found = &ts
-			return false
+		// Vertex 标准错误格式：google.rpc.RetryInfo.retryDelay（"39s" / "1.5s"）
+		if detail.Get("@type").String() == geminiRetryInfoTypeURL {
+			if dur, err := time.ParseDuration(detail.Get("retryDelay").String()); err == nil && dur > 0 {
+				if dur > geminiRetryInfoMaxDelay {
+					dur = geminiRetryInfoMaxDelay
+				}
+				ts := time.Now().Unix() + int64(math.Ceil(dur.Seconds()))
+				found = &ts
+				return false
+			}
 		}
 		return true
 	})

backend/internal/service/gemini_messages_compat_service_test.go

@@ -1017,6 +1017,40 @@ func TestParseGeminiRateLimitResetTime(t *testing.T) {
 			wantNil:     false,
 			approxDelta: 10,
 		},
+		{
+			name:        "Vertex RetryInfo retryDelay",
+			input:       `{"error":{"code":429,"message":"Resource exhausted. Please try again later.","status":"RESOURCE_EXHAUSTED","details":[{"@type":"type.googleapis.com/google.rpc.RetryInfo","retryDelay":"39s"}]}}`,
+			wantNil:     false,
+			approxDelta: 39,
+		},
+		{
+			name:        "Vertex RetryInfo 小数秒向上取整",
+			input:       `{"error":{"code":429,"status":"RESOURCE_EXHAUSTED","details":[{"@type":"type.googleapis.com/google.rpc.RetryInfo","retryDelay":"1.5s"}]}}`,
+			wantNil:     false,
+			approxDelta: 2,
+		},
+		{
+			name:        "Vertex RetryInfo 超过上限被截断到 15 分钟",
+			input:       `{"error":{"code":429,"status":"RESOURCE_EXHAUSTED","details":[{"@type":"type.googleapis.com/google.rpc.RetryInfo","retryDelay":"3600s"}]}}`,
+			wantNil:     false,
+			approxDelta: 900,
+		},
+		{
+			name:    "retryDelay 存在但 @type 不是 RetryInfo 则忽略",
+			input:   `{"error":{"code":429,"details":[{"@type":"type.googleapis.com/google.rpc.ErrorInfo","retryDelay":"39s"}]}}`,
+			wantNil: true,
+		},
+		{
+			name:    "RetryInfo retryDelay 非法值忽略",
+			input:   `{"error":{"code":429,"details":[{"@type":"type.googleapis.com/google.rpc.RetryInfo","retryDelay":"soon"}]}}`,
+			wantNil: true,
+		},
+		{
+			name:        "quotaResetDelay 与 RetryInfo 并存时先到者优先",
+			input:       `{"error":{"code":429,"details":[{"metadata":{"quotaResetDelay":"120s"}},{"@type":"type.googleapis.com/google.rpc.RetryInfo","retryDelay":"39s"}]}}`,
+			wantNil:     false,
+			approxDelta: 120,
+		},
 	}
 
 	for _, tt := range tests {