fix: serialize Claude API calls via queue to prevent 429 rate-limit spam

Tier-1 Anthropic API has 40K TPM — with ~20K tokens per pipeline step,
concurrent calls immediately hit the limit. enqueueClaude() serializes
all generateClaude() calls so only one runs at a time, eliminating
the flood of 429-retry-429-retry loops.

packages/api/src/llm/client.ts

@@ -31,6 +31,16 @@ function sleep(ms: number): Promise<void> {
 // ANTHROPIC CLAUDE PROVIDER
 // ═══════════════════════════════════════════════════════
 
+// Serialize Claude API calls to stay within TPM limits
+// Tier-1 has 40,000 TPM — with ~20K tokens/step, only 1 concurrent call safe
+let claudeQueue: Promise<unknown> = Promise.resolve();
+
+function enqueueClaude<T>(fn: () => Promise<T>): Promise<T> {
+  const result = claudeQueue.then(() => fn());
+  claudeQueue = result.catch(() => {});
+  return result;
+}
+
 async function generateClaude(
   systemPrompt: string,
   userPrompt: string,
@@ -40,6 +50,7 @@ async function generateClaude(
     throw new Error("ANTHROPIC_API_KEY not set — cannot use Claude provider");
   }
 
+  return enqueueClaude(async () => {
   const startTime = Date.now();
 
   const resp = await fetch("https://api.anthropic.com/v1/messages", {
@@ -90,6 +101,7 @@ async function generateClaude(
     totalDuration: duration * 1_000_000, // ns for compat
     evalCount: data.usage.output_tokens,
   };
+  }); // end enqueueClaude
 }
 
 // ═══════════════════════════════════════════════════════