Pin JUDGE_MODEL to a concrete model ID, fix #13

The `haiku` CLI alias can be repointed to a future generation or removed;
either case silently shifts judge score baselines or breaks the judge
entirely, breaking historical comparisons. Pin to `claude-haiku-4-5`
and route the preflight auth smoke test through the same constant so it
exercises the model the judge actually uses.

Author: claude

src/server/preflight.ts

@@ -3,7 +3,7 @@ import { promisify } from 'node:util';
 import { homedir, tmpdir } from 'node:os';
 import { join } from 'node:path';
 import { pathExists, atomicWriteFile, ensureDir, readJsonIfExists } from './fsUtil.js';
-import { PROJECT_MARKERS, STORAGE_DIR_NAME } from '@shared/constants.js';
+import { JUDGE_MODEL, PROJECT_MARKERS, STORAGE_DIR_NAME } from '@shared/constants.js';
 import { readdir, readFile, writeFile, unlink } from 'node:fs/promises';
 import { log } from './log.js';
 
@@ -104,7 +104,7 @@ export async function authSmokeTest(bin: string): Promise<void> {
     '-p',
     'ping',
     '--model',
-    'haiku',
+    JUDGE_MODEL,
     '--output-format',
     'json',
     '--tools',

src/shared/constants.ts

@@ -20,7 +20,11 @@ export const PROJECT_MARKERS = [
   'pyproject.toml',
 ];
 
-export const JUDGE_MODEL = 'haiku';
+// Pin to a concrete model ID, not the `haiku` CLI alias: aliases can be
+// repointed to a future generation or removed, which would silently shift
+// score baselines or break the judge entirely (issue #13). Bump manually
+// when a new Haiku ships and rebaselining is acceptable.
+export const JUDGE_MODEL = 'claude-haiku-4-5';
 
 // Judge input budgets (plan § Judge flow step 2).
 export const JUDGE_PROMPT_CAP_BYTES = 4 * 1024;