test: de-flake whole-group teardown; make CI failures diagnosable

.github/workflows/ci.yml

@@ -5,6 +5,9 @@ on:
 jobs:
   test:
     strategy:
+      # Without this, the first failing job cancels the rest — so a flake on one
+      # platform hides whether the others passed.
+      fail-fast: false
       matrix:
         os: [ubuntu-latest, macos-latest]
         node: [20, 22]

test/process.test.ts

@@ -45,7 +45,15 @@ describe("whole-group teardown", () => {
       });
       expect(pid).toBeGreaterThan(0);
       expect(pgid).toBe(pid); // detached ⇒ group leader
-      expect(await waitForPort(GROUP_PORT, 8000)).toBe(true);
+
+      // This fixture starts TWO node processes (server + a fanned-out child)
+      // before it binds, so it needs more headroom than a one-liner server on a
+      // loaded CI runner. On failure, surface the server's own log — a bare
+      // "expected false to be true" tells you nothing.
+      if (!(await waitForPort(GROUP_PORT, 20_000))) {
+        const log = readFileSync(join(dir, "out.log"), "utf8");
+        throw new Error(`server never bound port ${GROUP_PORT}. Its log:\n${log || "(empty)"}`);
+      }
       expect(existsSync(pidfile)).toBe(true);
       const childPid = Number(readFileSync(pidfile, "utf8"));