fix(test): v1.3.0's release gate failed on Linux — a test helper counted event-loop turns

.github/workflows/release.yml

@@ -40,30 +40,9 @@ jobs:
         with:
           node-version: "24"
       - name: Extension unit tests
-        # `shopt` is a bash builtin, so pin the shell instead of relying on the runner default (bash on
-        # Linux/macOS, but pwsh on Windows — where this step would break if the job were ever copied).
-        # Explicit `shell: bash` also upgrades the default `bash -e` to
-        # `bash --noprofile --norc -eo pipefail`, so a stray profile file can't perturb the gate either.
-        shell: bash
-        # DISCOVERS suites — it used to `cd extensions/levelcode-ai`, so levelcode-updater's tests never
-        # ran here, including the one guarding the updater's Download button against serving a raw
-        # .app.zip. Globbing every extension means a new suite is gated the moment it is added, with no
-        # list here to keep in sync. Requires are file-relative, so running from the repo root is fine.
-        run: |
-          shopt -s nullglob
-          count=0
-          for t in extensions/*/test/*.test.js; do
-            echo "── $t"
-            node "$t"          # `-e` (from `shell: bash` above) aborts the job on the first failure
-            count=$((count + 1))
-          done
-          # A zero-match glob would otherwise report success and gate nothing — the exact failure this
-          # step is fixing. Fail loudly instead.
-          if [ "$count" -eq 0 ]; then
-            echo "::error::No suites matched extensions/*/test/*.test.js — the gate would pass vacuously."
-            exit 1
-          fi
-          echo "──────── $count test files passed ────────"
+        # The gate itself lives in a script, so the pull-request workflow (test.yml) runs exactly what
+        # this does. It discovers every extensions/*/test/*.test.js and fails if it finds none.
+        run: ./scripts/test-extensions.sh
 
   build:
     name: Build ${{ matrix.arch }}

.github/workflows/test.yml

@@ -0,0 +1,33 @@
+name: Extension tests
+
+# The release gate (scripts/test-extensions.sh), on every pull request and on develop.
+#
+# release.yml runs the same script, but only when a tag is pushed — which made a tag the first time
+# the suites ran on Linux. v1.3.0's gate failed there on a suite that had only ever run on a Mac,
+# after the pull requests that wrote it had merged. Here a suite that passes on one machine and not
+# on the runner is found on the pull request that adds it.
+
+on:
+  pull_request:
+  push:
+    branches: [develop]
+
+permissions:
+  contents: read
+
+# A newer push to the same branch or pull request supersedes the run before it.
+concurrency:
+  group: extension-tests-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  test:
+    name: Extension unit tests
+    runs-on: ubuntu-latest # the runner the release gate uses
+    steps:
+      - uses: actions/checkout@v7
+      - uses: actions/setup-node@v7
+        with:
+          node-version: "24" # as release.yml
+      - name: Extension unit tests
+        run: ./scripts/test-extensions.sh

extensions/levelcode-ai/test/authRetryCallers.test.js

@@ -210,11 +210,23 @@ const { registerInlineComplete } = require('../inlineComplete');
 
 // ── the network: the only stand-in ───────────────────────────────────────────────────────────────
 const tick = () => new Promise((resolve) => setImmediate(resolve));
+/** How long anything here is given to happen. Reached only by what never does. */
+const PATIENCE_MS = 2000;
+/**
+ * Wait for `cond` — by the CLOCK, not by counting turns of the event loop. Inline completion waits
+ * out its debounce on a real timer, a millisecond even when the setting is 0, and how many turns fit
+ * into a millisecond is the machine's business: a few dozen on a Mac, well over a thousand on
+ * Linux, where the release gate runs. A count that was plenty on one ran out on the other before
+ * the timer had fired.
+ */
 async function until(cond, what) {
-	for (let i = 0; i < 400; i++) { if (cond()) { return; } await tick(); }
-	assert.fail('never happened: ' + what);
+	const deadline = Date.now() + PATIENCE_MS;
+	while (!cond()) {
+		if (Date.now() > deadline) { assert.fail('never happened: ' + what); }
+		await tick();
+	}
 }
-/** Let everything already in motion get as far as it can. */
+/** Let everything already in motion get as far as it can. Turns, not time — so only for what no timer stands in the way of. */
 async function settle() { for (let i = 0; i < 40; i++) { await tick(); } }
 function deferred() { let resolve; const promise = new Promise((r) => { resolve = r; }); return { promise, resolve }; }
 function within(promise, ms, what) {

scripts/test-extensions.sh

@@ -0,0 +1,30 @@
+#!/usr/bin/env bash
+#
+# Run every extension's unit suites. This is the gate: the release workflow runs it before spending
+# an hour of macOS build minutes, and every pull request runs it too (.github/workflows/test.yml).
+#
+# It DISCOVERS suites — extensions/*/test/*.test.js — so a new one is gated the moment it is added,
+# with no list to keep in sync. (The gate used to `cd extensions/levelcode-ai`; levelcode-updater's
+# suites never ran, including the one guarding the Download button against serving a raw .app.zip.)
+# Requires are file-relative, so the suites run from the repo root.
+#
+# Usage:
+#   ./scripts/test-extensions.sh
+#
+set -euo pipefail
+shopt -s nullglob
+
+cd "$(dirname "$0")/.."
+
+count=0
+for t in extensions/*/test/*.test.js; do
+  echo "── $t"
+  node "$t"          # `set -e` stops at the first suite that fails
+  count=$((count + 1))
+done
+# A glob that matches nothing would otherwise report success and gate nothing. Fail loudly instead.
+if [ "$count" -eq 0 ]; then
+  echo "::error::No suites matched extensions/*/test/*.test.js — the gate would pass vacuously."
+  exit 1
+fi
+echo "──────── $count test files passed ────────"