fix(llm-client): retry Cloudflare edge timeouts 522 and 524

src/campaign/external-optimizer-callback.ts

@@ -179,15 +179,19 @@ async function handleCallback<TResponse>(
     let result: TResponse
     try {
       result = await args.evaluate({ candidate, exampleId: body.exampleId }, signal)
-    } catch {
+    } catch (error) {
       observe({
         kind: 'refusal',
         reason: 'evaluation-failed',
         candidate,
         candidateHash,
         exampleId: body.exampleId,
       })
-      sendJsonIfOpen(response, 500, { error: 'evaluation failed' })
+      // The thrown detail is the only diagnostic for a failed evaluation —
+      // an opaque 500 costs the caller a blind debugging round.
+      sendJsonIfOpen(response, 500, {
+        error: `evaluation failed: ${String(error).slice(0, 400)}`,
+      })
       return
     }
     let encoded: string

src/llm-client.ts

@@ -349,7 +349,10 @@ function providerTokenCount(value: unknown): number | undefined {
   return typeof value === 'number' && Number.isSafeInteger(value) && value >= 0 ? value : undefined
 }
 
-const RETRYABLE_STATUS = new Set([429, 502, 503, 504])
+// 522/524: Cloudflare edge timeouts (origin connect/respond) — transient
+// exactly like 504. Router deployments serve through Cloudflare, so a long
+// non-streaming call can draw these from the edge, not the origin.
+const RETRYABLE_STATUS = new Set([429, 502, 503, 504, 522, 524])
 
 /**
  * Transient transport/network error signatures, matched against an error's

tests/campaign/external-optimizer-lifecycle.test.ts

@@ -38,6 +38,22 @@ describe('external optimizer server lifecycle', () => {
     expect(activeEvaluations).toBe(0)
   })
 
+  it('carries the thrown detail in the evaluation-failed 500 body', async () => {
+    const callback = await startExternalOptimizerCallback({
+      token: 'secret',
+      maxEvaluations: 1,
+      evaluate: async () => {
+        throw new Error('router 503 after 3 attempts')
+      },
+    })
+    const response = await postEvaluation(callback.url, callback.token)
+    const body = (await response.json()) as { error: string }
+    await callback.close()
+
+    expect(response.status).toBe(500)
+    expect(body.error).toContain('router 503 after 3 attempts')
+  })
+
   it('closes the callback and aborts active evaluation work with the owner signal', async () => {
     const owner = new AbortController()
     const started = deferred<void>()

tests/llm-transient-status.test.ts

@@ -0,0 +1,23 @@
+import { describe, expect, it } from 'vitest'
+import { isTransientLlmError, LlmCallError } from '../src/llm-client'
+
+describe('isTransientLlmError — HTTP status classification', () => {
+  const call = (status: number) => new LlmCallError('upstream error', status, '', 'test-model')
+
+  it.each([429, 502, 503, 504])('retries gateway-transient status %i', (status) => {
+    expect(isTransientLlmError(call(status))).toBe(true)
+  })
+
+  it.each([522, 524])('retries Cloudflare edge timeout %i like a 504', (status) => {
+    expect(isTransientLlmError(call(status))).toBe(true)
+  })
+
+  it.each([400, 401, 404, 422])('never retries deterministic status %i', (status) => {
+    expect(isTransientLlmError(call(status))).toBe(false)
+  })
+
+  it('reads a retryable numeric status off a foreign (non-LlmCallError) error', () => {
+    const err = Object.assign(new Error('edge timeout'), { status: 524 })
+    expect(isTransientLlmError(err)).toBe(true)
+  })
+})