fix(routing): bench a denied credential across all models, not one

server/dist/providers/google.d.ts

@@ -13,6 +13,21 @@ export declare class GoogleProvider extends BaseProvider {
     createSpeech(apiKey: string, request: SpeechRequest, modelId: string): Promise<SpeechResult>;
     createRealtimeSession(apiKey: string, request: RealtimeSessionRequest, modelId: string): Promise<RealtimeSessionResponse>;
     streamChatCompletion(apiKey: string, messages: ChatMessage[], modelId: string, options?: CompletionOptions): AsyncGenerator<ChatCompletionChunk>;
+    /**
+     * Liveness only — this proves the key authenticates, NOT that its project may
+     * run inference.
+     *
+     * Verified 2026-08-23 across 16 live keys: a project answering
+     * "403: Your project has been denied access" on generateContent still returns
+     * 200 from both `GET /models` and `countTokens`, so neither free endpoint can
+     * detect the denial. Only generateContent can, and health.ts sweeps every 5
+     * minutes — probing it here would burn ~4,600 free-tier requests a day just
+     * to check keys.
+     *
+     * The denial is caught where it actually surfaces instead: classifyProviderError
+     * maps it to an auth failure with a key-scoped 24h cooldown, so one 403 benches
+     * that credential across every model rather than once per model.
+     */
     validateKey(apiKey: string): Promise<boolean>;
 }
 //# sourceMappingURL=google.d.ts.map

server/dist/services/provider-errors.d.ts

@@ -4,6 +4,13 @@ export interface ClassifiedProviderError {
     retryable: boolean;
     skipModel: boolean;
     keyCooldownMs: number;
+    /**
+     * How wide the cooldown should be. 'model' benches the credential for the
+     * one model that failed; 'key' benches it for every model on the platform,
+     * which is the only correct scope when the failure is a property of the
+     * account rather than the model.
+     */
+    cooldownScope?: 'model' | 'key';
 }
 export declare function classifyProviderError(err: unknown): ClassifiedProviderError;
 export declare function canRetryProviderFailure(failure: ClassifiedProviderError, requestedModel?: string): boolean;

server/dist/services/ratelimit.d.ts

@@ -11,6 +11,16 @@ export declare function canUseTokens(platform: string, modelId: string, keyId: n
 export declare function recordRequest(platform: string, modelId: string, keyId: number): void;
 export declare function recordTokens(platform: string, modelId: string, keyId: number, tokens: number): void;
 export declare function setCooldown(platform: string, modelId: string, keyId: number, durationMs?: number): void;
+/**
+ * Bench a credential across ALL models on its platform.
+ *
+ * Some failures are properties of the account, not the model: a Google project
+ * denied access answers 403 for every model it owns. A per-model cooldown
+ * cannot express that, so the router kept re-selecting the bad credential once
+ * per model — one denied Google project produced 50 logged 403s across the
+ * catalog while its dashboard row still read "healthy".
+ */
+export declare function setKeyCooldown(platform: string, keyId: number, durationMs?: number): void;
 export declare function isOnCooldown(platform: string, modelId: string, keyId: number): boolean;
 export declare function getRateLimitStatus(platform: string, modelId: string, keyId: number, limits: {
     rpm: number | null;