Prove can't-fail test verdicts with a mutation

[kmcginnes]

Summary

A review recommended deleting cross-tab isolation tests as "can't fail" after reasoning about the current code, when the tests only needed to mount the open tab's atom. These two rules make that verdict something a reviewer proves rather than argues.

 REVIEW.md › Tests that can't fail
-  Ask whether the test would fail if the change were reverted.
+  Name a plausible regression of the behavior the test's name promises,
+  and prove the verdict with a mutation. A test that stays green is the
+  finding; tighten it until the mutation turns it red. Delete only when
+  no regression of that behavior could reach the test.

 docs/agents/testing.md › Special cases
+  Jotai mount effects: store.get doesn't mount an atom, so onMount never
+  runs. Mount it with store.sub(atom, () => {}) when the behavior depends on it.

The Jotai note is the gotcha behind the near-miss: the test fidelity PR (#2389) showed that an onMount cross-tab sync left the old tests green and turned the subscribed ones red.

Merge Danger

Door: two-way

Impact: none

Review and testing guidance only.

Related

• Related to Tighten test fidelity in connection, proxy, and session storage tests #2389

[kmcginnes]

Approved