fix(cli): stop a test's real MCP warm, and let the temp guard see every leftover

cli/__tests__/adapters.claude.environment.test.mjs

@@ -1806,10 +1806,14 @@ describe('claude adapter: the public seat Commonly MCP witness (TASK-256)', () =
     }));
     fs.writeFileSync(path.join(pkgDir, 'src', 'index.js'), '');
     fs.writeFileSync(path.join(home, 'current'), '0.3.13\n');
-    // A fresh registry answer naming the installed build, so no turn here starts
-    // a background warm: a detached `npm install` outlives the suite and writes
-    // into TMPDIR after it ends (the CLI temp-cleanup guard caught it).
     fs.writeFileSync(path.join(home, '.registry.json'), JSON.stringify({ version: '0.3.13', checkedAt: Date.now() }));
+    // A held warm lock, so no turn here starts a background warm. A fresh
+    // registry answer isn't enough: a seat running a hand-run copy resolves no
+    // version from this home, and a spawn with no version always wants a warm.
+    // That warm is a detached `npm install` from the real registry. It outlived
+    // the suite, refilled this home after the teardown below removed it, and
+    // raced the CLI temp-cleanup guard's own rm (ENOTEMPTY, two dequeued PRs).
+    fs.writeFileSync(path.join(home, '.warm.lock'), 'warming\n');
     prior = process.env.COMMONLY_MCP_HOME;
     process.env.COMMONLY_MCP_HOME = home;
     Object.defineProperty(process, 'platform', { value: 'darwin' });

cli/scripts/verify-temp-cleanup.mjs

@@ -2,10 +2,17 @@
 
 /**
  * Run the CLI Jest files that create temp homes twice in an isolated TMPDIR and
- * fail if any test-created temp directory survives either run. The full CLI
+ * fail if anything the tests created survives either run. The full CLI
  * suite runs in CI separately; this is a cheap regression guard for the
  * teardown contract. The default list is the files that leaked ~40 GB
  * (TASK-436); pass file paths to check others.
+ *
+ * Everything left in the probe root counts except the tools' own caches. It
+ * used to count a fixed list of prefixes, and a home under a new prefix
+ * (`cli-claude-mcp-home-`) leaked 29 MB unseen: a detached MCP warm that a test
+ * started kept writing into it after Jest exited, and the only symptom was this
+ * script's final rm failing with ENOTEMPTY on whichever PR was in the queue.
+ * That rm still fails the run, now naming what was still being written.
  */
 import { spawn } from 'node:child_process';
 import { lstat, mkdtemp, readdir, rm } from 'node:fs/promises';
@@ -20,13 +27,11 @@ const targets = process.argv.length > 2
     '__tests__/adapters.codex.test.mjs',
     '__tests__/adapters.pi.test.mjs',
   ];
-const prefixes = [
-  'opencode-adapter-test-',
-  'cli-mcp-home-',
-  'cli-mcp-pin-',
-  'cli-mcp-home-old-',
-  'commonly-mcp-home-jest-',
-];
+// Caches Jest and Node write to TMPDIR for themselves, not for a test.
+const toolCaches = [/^jest_/, /^node-compile-cache$/];
+// A detached child a test started gets this long to show itself in the probe
+// root before the root is read.
+const SETTLE_MS = 1500;
 
 const bytesUnder = async (path) => {
   const info = await lstat(path);
@@ -38,7 +43,7 @@ const bytesUnder = async (path) => {
 };
 
 const snapshot = async (root) => {
-  const entries = (await readdir(root)).filter((name) => prefixes.some((prefix) => name.startsWith(prefix)));
+  const entries = (await readdir(root)).filter((name) => !toolCaches.some((cache) => cache.test(name)));
   let bytes = 0;
   for (const entry of entries) bytes += await bytesUnder(join(root, entry));
   return { count: entries.length, bytes, entries };
@@ -71,12 +76,26 @@ try {
   for (let pass = 1; pass <= 2; pass += 1) {
     const before = await snapshot(probeRoot);
     await run(probeRoot);
+    await new Promise((resolve) => { setTimeout(resolve, SETTLE_MS); });
     const after = await snapshot(probeRoot);
     console.log(`CLI temp cleanup pass ${pass}: before=${before.count} dirs/${before.bytes} bytes, after=${after.count} dirs/${after.bytes} bytes`);
     if (after.count !== 0 || after.bytes !== 0) {
       throw new Error(`CLI test temp directories survived pass ${pass}: ${after.entries.join(', ')}`);
     }
   }
 } finally {
-  await rm(probeRoot, { recursive: true, force: true });
+  try {
+    await rm(probeRoot, { recursive: true, force: true, maxRetries: 5, retryDelay: 200 });
+  } catch (error) {
+    // A root that won't go is a leak the snapshots missed. ENOTEMPTY after the
+    // retries means something kept writing after the last snapshot, the
+    // detached-child case above. exitCode, not a throw, so a leak check's own
+    // error stays the one reported when both happen.
+    const left = await readdir(probeRoot).catch(() => []);
+    const hint = error.code === 'ENOTEMPTY'
+      ? ' Something kept writing there after the last snapshot, usually a process a test started that outlived Jest.'
+      : '';
+    console.error(`CLI temp cleanup could not remove ${probeRoot} (${error.code || error.message}); left in it: ${left.join(', ') || 'nothing readable'}.${hint}`);
+    process.exitCode = 1;
+  }
 }