[schmitch]

The npm-e2e job (from #463) can fail on a race in its own script, as it did in the merge queue's run of #465. That run had the same tree that had passed on #465's own PR run.

── legacy login with Yarn, and logout ──
Error: write EPIPE
    at Wt.writeLine (…/yarn/package/bin/yarn.js:177:6121)
FAIL: yarn npm whoami

node yarn.js npm whoami | grep -q "$admin_user": grep -q exits at its first match. When Yarn writes another line after that, it writes into a closed pipe, dies of the unhandled EPIPE, and with set -o pipefail the check fails although the whoami was right. Whether Yarn's last write lands before or after grep exits is timing, hence the intermittent failure.

Change

.github/scripts/npm-e2e.sh now reads a client's output whole and then compares it, as its other checks already do ([ "$(npm whoami …)" = … ]). The same applies to the other checks that piped a client into grep -q:

• yarn npm whoami;
• the two npm view … version | grep -qx checks;
• npm dist-tag ls … | grep -q legacy && fail. Besides the race, that check passed silently when npm dist-tag ls itself failed.

No server code changes, so there is no changelog fragment.

Tests

• .github/scripts/npm-e2e.sh locally with npm 11.21, pnpm 11.5 and Yarn 4 (fetched by the script): passed.
• bash -n on the script.

---

Generated by Claude Code

[schmitch]

The merge queue's run of this PR (run 776) failed one check, test, in crates/cache/tests/l1_unit.rs. This PR only changes .github/scripts/npm-e2e.sh, and the same check passed on its PR run.

a_factory_chosen_lifetime_shortens_the_ttl ... FAILED
assertion `left == right` failed: still live  (left: 2, right: 1)   l1_unit.rs:524

The test gives an entry a 100 ms lifetime and expects it still live on the very next call. On a runner busy with the parallel test binaries, more than 100 ms passed between the two calls, so the entry had expired and the loader ran again. No fix for it exists yet. One that keeps the test's meaning with half-second margins (about 5 s instead of 0.8 s):

-    assert_eq!(load(Duration::from_millis(100)).await, 1);
-    assert_eq!(load(Duration::from_millis(100)).await, 1, "still live");
-    tokio::time::sleep(Duration::from_millis(200)).await;
-    assert_eq!(load(Duration::from_secs(60)).await, 2, "expired at 100 ms");
-    tokio::time::sleep(Duration::from_millis(300)).await;
-    assert_eq!(load(Duration::from_secs(60)).await, 2, "capped at the TTL");
-    tokio::time::sleep(Duration::from_millis(300)).await;
-    assert_eq!(load(Duration::from_secs(60)).await, 3, "expired at 500 ms");
+    // TTL 3 s (the `Duration::from_millis(500)` above becomes `from_secs(3)`).
+    assert_eq!(load(Duration::from_secs(1)).await, 1);
+    assert_eq!(load(Duration::from_secs(1)).await, 1, "still live");
+    tokio::time::sleep(Duration::from_millis(1500)).await;
+    assert_eq!(load(Duration::from_secs(60)).await, 2, "expired at 1 s");
+    tokio::time::sleep(Duration::from_secs(1)).await;
+    assert_eq!(load(Duration::from_secs(60)).await, 2, "capped at the TTL");
+    tokio::time::sleep(Duration::from_millis(2500)).await;
+    assert_eq!(load(Duration::from_secs(60)).await, 3, "expired at 3 s");

It belongs in a pull request of its own, not in this one. Re-queueing this PR should pass, since the race rarely hits.

---

Generated by Claude Code