Isolate judge subprocess from project settings (closes #3)

Spawn the judge with cwd=os.tmpdir() so it cannot walk up from the
user's project applying their .claude/settings.json, hooks, and
project-scoped tool overrides. Switch --setting-sources from
'project' to 'user' for predictable, project-independent behavior.

Without this, judge scores were silently contaminated by per-project
hooks (which could log/mutate the prompt or response), permission
rules, and a project-level defaultModel that could override --model
haiku — making scores quietly differ between projects for reasons
unrelated to the variants under test.

Author: slaFFik

docs/PLAN.md

@@ -217,8 +217,12 @@ The judge is **per-variant**, powered by **Haiku**. It runs independently for ea
      --output-format json \
      --tools        "" \
      --allowedTools "" \
-     --strict-mcp-config
+     --strict-mcp-config \
+     --setting-sources user
    ```
+   - `cwd` = `os.tmpdir()` (the judge must NOT inherit the user's project cwd, or it would walk up applying their `.claude/settings.json`, hooks, and project-scoped tool overrides — silent score contamination).
+   - `--setting-sources user` keeps the source explicit and predictable across projects.
+
    Judge uses no tools — it reads its inputs from the prompt and emits structured JSON. Output shape enforced via prompt + ajv validation (with `--json-schema` fallback if supported by the installed CLI).
 5. On schema failure, retry once with an error-explaining follow-up. Still invalid → `judge.json.status = "errored"` with the error surfaced in the UI; the column's run results are untouched.
 6. Write `<run>/judge.json`. Push `judge.updated` (or `judge.errored`) for that column over SSE.

src/server/judge.ts

@@ -1,4 +1,5 @@
 import { spawn } from 'node:child_process';
+import { tmpdir } from 'node:os';
 import { join } from 'node:path';
 import JSON5 from 'json5';
 import {
@@ -260,13 +261,19 @@ function spawnJudge(claudeBin: string, prompt: string, opts: SpawnJudgeOptions):
       '',
       '--strict-mcp-config',
       '--setting-sources',
-      'project',
+      'user',
       '--disable-slash-commands',
     ];
     if (opts.jsonSchema) {
       args.push('--json-schema', JSON.stringify(JUDGE_MODEL_JSON_SCHEMA));
     }
-    const proc = spawn(claudeBin, args, { stdio: ['ignore', 'pipe', 'pipe'] });
+    // Spawn from os.tmpdir() so the judge never inherits the user's project cwd.
+    // Combined with `--setting-sources user`, this prevents project `.claude/settings.json`,
+    // hooks, and tool overrides from silently contaminating judge scores (issue #3).
+    const proc = spawn(claudeBin, args, {
+      stdio: ['ignore', 'pipe', 'pipe'],
+      cwd: tmpdir(),
+    });
     let stdout = '';
     let stderr = '';
     proc.stdout.on('data', (d) => (stdout += d.toString()));