Address Copilot review on PR #25

- Drop unused isJudgeTimeout flag from JudgeTimeoutError; instanceof is
  sufficient since the spawn function and the error class live in the
  same module realm.
- Clamp bytesCapMultiplier to a finite positive number (defaulting to 1)
  so NaN / Infinity / negative inputs cannot produce empty or runaway
  prompt sections.
- Strengthen the timeout-retry test: capture the full retry prompt and
  explicitly assert the schema-retry hint header is absent (not just
  that the prompt is shorter).

Author: claude

src/server/judge.ts

@@ -31,7 +31,6 @@ export interface JudgeInput {
 }
 
 export class JudgeTimeoutError extends Error {
-  readonly isJudgeTimeout = true;
   constructor(message: string) {
     super(message);
     this.name = 'JudgeTimeoutError';
@@ -128,7 +127,11 @@ export function buildJudgePrompt(
   opts: BuildJudgePromptOptions = {},
 ): JudgePromptArtifacts {
   const { runConfig, transcript, variantContent, outputs } = input;
-  const m = opts.bytesCapMultiplier ?? 1;
+  // Default to 1 if the option is missing or non-finite/non-positive; otherwise
+  // multiplier values like NaN or Infinity would propagate through Math.floor and
+  // produce empty or runaway prompt sections.
+  const rawMultiplier = opts.bytesCapMultiplier ?? 1;
+  const m = Number.isFinite(rawMultiplier) && rawMultiplier > 0 ? rawMultiplier : 1;
   // Floors keep retries useful even if a future caller passes a very small multiplier.
   const promptCap = Math.max(256, Math.floor(JUDGE_PROMPT_CAP_BYTES * m));
   const variantCap = Math.max(512, Math.floor(JUDGE_VARIANT_CAP_BYTES * m));

test/judge.spec.ts

@@ -523,9 +523,9 @@ async function withTmpRunDir<T>(fn: (runDir: string) => Promise<T>): Promise<T>
 
 scenario('runJudge: timeout on first attempt triggers retry with halved caps', async () => {
   await withTmpRunDir(async (runDir) => {
-    const calls: { promptLen: number }[] = [];
+    const calls: string[] = [];
     const spawnFn: SpawnJudgeFn = async (_bin, prompt) => {
-      calls.push({ promptLen: prompt.length });
+      calls.push(prompt);
       if (calls.length === 1) {
         throw new JudgeTimeoutError('judge subprocess timed out after 120s');
       }
@@ -539,13 +539,16 @@ scenario('runJudge: timeout on first attempt triggers retry with halved caps', a
       throw new Error(`expected 2 spawn calls (initial + retry), got ${calls.length}`);
     }
     // Halved caps must yield a strictly shorter retry prompt.
-    if (!(calls[1]!.promptLen < calls[0]!.promptLen)) {
+    if (!(calls[1]!.length < calls[0]!.length)) {
       throw new Error(
-        `retry prompt should be shorter (halved caps); first=${calls[0]!.promptLen} retry=${calls[1]!.promptLen}`,
+        `retry prompt should be shorter (halved caps); first=${calls[0]!.length} retry=${calls[1]!.length}`,
       );
     }
     // The retry must NOT include the schema-retry hint — that hint is only for
     // parse failures, not timeouts.
+    if (calls[1]!.includes('# Retry required')) {
+      throw new Error('timeout-retry prompt must not include the schema-retry hint header');
+    }
     const judgeFile = JSON.parse(readFileSync(join(runDir, 'judge.json'), 'utf8')) as {
       status: string;
       scores?: { accuracy: number };