fix(scripts): observe the final polling interval before declaring a timeout

Copilot's suppressed round-3 comment is right. The attempt-counted loop read
the output and then slept, so nothing ever observed the last interval: an
announcement arriving at 29.9s of a 30s budget was reported as a timeout, and
a spawn error or early exit landing in that window was misattributed the same
way. Inherited from the original smoke, but wrong either way.

Replace it with a deadline loop that re-reads state after every wait, the last
one included, and clamp the final wait to the deadline so the budget is a real
bound rather than one poll longer.

The regression test sets pollMs === timeoutMs, which makes the distinction
deterministic rather than a race: the attempt-counted loop gets exactly one
check, at t≈0 before the child has finished starting, then throws; the deadline
loop re-reads after the wait and sees the announcement. The child announces
immediately, ~2s inside the window, so a loaded runner can't flip it. Verified
both directions — the test fails against the old loop and passes against the
new one — and `npm run test:scripts` was repeated to confirm it is stable under
the full 116-test parallel run (an earlier margin-based version of this test
was not).

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013hzwzS8UvBsr4yZm7o7sev
Signed-off-by: cliffhall <cliff@futurescale.com>

scripts/lib/announced-child.mjs

@@ -69,15 +69,24 @@ export async function startAnnouncedChild({
   child.on("exit", () => (exited = true));
   child.on("close", () => (exited = true));
 
-  const attempts = Math.max(1, Math.ceil(timeoutMs / pollMs));
-  for (let attempt = 0; attempt < attempts; attempt++) {
+  // A deadline loop, not a fixed attempt count: state is re-read *after* every
+  // wait, including the last one. Counting attempts and throwing straight after
+  // the final `delay` leaves the whole last polling interval unobserved, so an
+  // announcement landing at 29.9s of a 30s budget is reported as a timeout —
+  // and a spawn error or early exit in that window is misattributed the same
+  // way. The final wait is also clamped to the deadline so the budget is a
+  // real bound rather than one poll longer.
+  const deadline = Date.now() + timeoutMs;
+  for (;;) {
     const match = out.match(pattern);
     if (match) return { child, match };
     if (spawnError) {
       throw new Error(`could not spawn the ${what}: ${spawnError.message}`);
     }
     if (exited) throw new Error(`${what} exited early:\n${out}`);
-    await delay(pollMs);
+    const remaining = deadline - Date.now();
+    if (remaining <= 0) break;
+    await delay(Math.min(pollMs, remaining));
   }
   throw new Error(
     `${what} did not start within ${Math.round(timeoutMs / 1000)}s:\n${out}`,

scripts/lib/announced-child.test.mjs

@@ -70,6 +70,30 @@ test("scans stdout as well as stderr", async (t) => {
   assert.equal(match[1], "4242");
 });
 
+test("sees an announcement that lands inside the final polling interval", async (t) => {
+  // The off-by-one this guards: a fixed attempt count checks the output and
+  // then sleeps, so nothing ever observes the last interval — an announcement
+  // at 29.9s of a 30s budget reads as a timeout.
+  //
+  // Setting pollMs === timeoutMs makes that deterministic rather than a race:
+  // the attempt-counted loop gets exactly one check, at t≈0 before the child
+  // has even finished starting, and then throws. The deadline loop re-reads
+  // after the wait, so *any* announcement before the deadline is seen. The
+  // child announces immediately, ~2s inside the window on either side, so a
+  // loaded runner cannot flip the outcome. (Verified: this fails against the
+  // attempt-counted loop and passes against the deadline loop.)
+  const { match } = await startAnnouncedChild({
+    command: process.execPath,
+    args: ["-e", 'console.error("listening at http://127.0.0.1:8080")'],
+    pattern: /listening at (http:\/\/\S+)/i,
+    onSpawn: killOnTeardown(t),
+    what: "probe",
+    timeoutMs: 2000,
+    pollMs: 2000,
+  });
+  assert.equal(match[1], "http://127.0.0.1:8080");
+});
+
 test("a non-announcing child is still reachable and killable after the timeout", async (t) => {
   // The regression: a child that is alive but never announces. The throw must
   // not be the only thing that happens — the caller must already hold the