fix: claude CLI version resolves in the runner parent, never on a test thread

Eng-review finding: getClaudeCliVersion's fallback is a SYNCHRONOUS
spawnSync on the same thread that polls concurrent PTY/session tests — the
judgePtyState blocking class this overhaul kills elsewhere. The paid runner
parent now resolves it once (cached) and stamps GSTACK_CLAUDE_CLI_VERSION
into every shard's env; eval-store short-circuits on the env var, and the
fallback spawn's budget tightens 10s -> 3s (bounded one-time stall, records
'unknown' on a slow CLI).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

scripts/test-paid-shards.ts

@@ -64,7 +64,7 @@ import {
 } from './test-strict-output';
 import { PAID_TEST_GLOBS, isPaidTestFile } from '../test/helpers/paid-test-set';
 import { PERIODIC_CI_EXCLUDE } from '../test/helpers/periodic-exclude-data';
-import { getProjectEvalDir } from '../test/helpers/eval-store';
+import { getProjectEvalDir, getClaudeCliVersion } from '../test/helpers/eval-store';
 import { preflightAnthropicApi } from '../test/helpers/anthropic-preflight';
 import {
   detectBaseBranch,
@@ -493,6 +493,12 @@ export async function runPaidShard(
   if (options.evalDirBase) {
     env.GSTACK_EVAL_DIR = path.join(options.evalDirBase, 'shards', shardSlug(files));
   }
+  // Resolve `claude --version` ONCE in the parent (cached across shards) and
+  // hand it to every child: eval-store's fallback is a synchronous spawn on
+  // the same thread that polls PTY sessions, so children must never pay it.
+  if (!env.GSTACK_CLAUDE_CLI_VERSION) {
+    env.GSTACK_CLAUDE_CLI_VERSION = getClaudeCliVersion();
+  }
 
   const startedAt = Date.now();
   log(`${label} START ${files.join(' ')} (timeout ${Math.round(timeoutMs / 1000)}s)`);

test/helpers/eval-store.ts

@@ -786,11 +786,23 @@ function getVersion(): string {
 // claude-CLI TUI drift only after long flake hunts — stamping the version
 // into every run record makes that correlation a grep instead of an
 // archaeology dig.
+//
+// GSTACK_CLAUDE_CLI_VERSION short-circuits the spawn entirely: the paid
+// runner's parent resolves the version once and passes it to every shard,
+// so test processes never block on it. The fallback spawn is SYNCHRONOUS on
+// the same thread that polls PTY sessions — the judgePtyState blocking
+// class — so its budget is a tight 3s, not a generous one: a slow/hung CLI
+// costs one bounded stall per process and records 'unknown'.
 let claudeCliVersionCache: string | null = null;
 export function getClaudeCliVersion(): string {
   if (claudeCliVersionCache !== null) return claudeCliVersionCache;
+  const fromEnv = process.env.GSTACK_CLAUDE_CLI_VERSION;
+  if (fromEnv) {
+    claudeCliVersionCache = fromEnv;
+    return claudeCliVersionCache;
+  }
   try {
-    const result = spawnSync('claude', ['--version'], { stdio: 'pipe', timeout: 10_000 });
+    const result = spawnSync('claude', ['--version'], { stdio: 'pipe', timeout: 3_000 });
     claudeCliVersionCache = result.stdout?.toString().split('\n')[0].trim() || 'unknown';
   } catch {
     claudeCliVersionCache = 'unknown';