fix: always store to S3 for PTB when usage extraction fails

For pass-through billing requests, if we fail to extract usage from
the response body in the Worker, we must store the body to S3 so
Jawn can extract it for accurate billing.

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

Author: 2 people

worker/src/lib/dbLogger/DBLoggable.ts

@@ -820,10 +820,13 @@ export class DBLoggable {
       reasoning_tokens?: number;
     } = {};
     let extractedModel: string | undefined;
+    let failedToGetUsage = true;
     try {
       const responseText = rawResponseBody.join("");
       const parsedResponse = JSON.parse(responseText);
       extractedUsage = this.getDetailedUsage(parsedResponse);
+      // Check if we actually got usage tokens
+      failedToGetUsage = !extractedUsage.prompt_tokens && !extractedUsage.completion_tokens;
       // Extract model from response (OpenAI format)
       if (
         typeof parsedResponse === "object" &&
@@ -833,12 +836,15 @@ export class DBLoggable {
         extractedModel = (parsedResponse as { model?: string }).model;
       }
     } catch {
-      // Ignore parsing errors - usage will be extracted later by Jawn if body is stored
+      // Parsing failed - Jawn will need to extract usage from the body
+      failedToGetUsage = true;
     }
 
     // Skip S3 storage only if BOTH request and response are omitted
     // If only one is omitted, still store to S3 - Jawn will respect the omit flags
+    // IMPORTANT: For PTB, if we failed to get usage, always store to S3 so Jawn can extract it for billing
     const skipS3Storage =
+      !(isPassthroughBilling && failedToGetUsage) &&
       requestHeaders?.omitHeaders?.omitRequest === true &&
       requestHeaders?.omitHeaders?.omitResponse === true;