@@ -287,6 +287,15 @@ describe("McpServerRestriction subcomponent — flicker regressions", () => {
 	 *     React.memo + the stable `customMode`/`onCommit` props.
 	 *
 	 * Fake timers prevent the debounced flush from firing during the test.
+	 *
+	 * NOTE on (b): we assert `<= 1` rather than `=== 0`. `<Profiler onRender>`
+	 * fires whenever the Profiler boundary commits, which happens whenever its
+	 * parent re-renders — even when every child inside bails out via
+	 * React.memo. So an equivalent-heartbeat rerender of the parent `Tree` will
+	 * still produce one Profiler callback with `phase: "update"` and ~0
+	 * `actualDuration`, regardless of memo. The real anti-flicker guarantee is
+	 * verified by Test 2 (DOM-node identity preserved across heartbeat); here
+	 * we just bound the child's render work to at most one commit.
 	 */
 	it("Test 3: profiler — 1 commit per click, 0 commits on equivalent mcpServers heartbeat", () => {
 		vitest.useFakeTimers()
@@ -322,7 +331,11 @@ describe("McpServerRestriction subcomponent — flicker regressions", () => {
 			// which is what a properly-memoized parent would do.
 			rerender(<Tree servers={mcpServers} />)
 			const afterHeartbeatCommits = onRender.mock.calls.length
-			expect(afterHeartbeatCommits - afterClickCommits).toBe(0)
+			// See JSDoc NOTE on (b): Profiler commits with its parent even when
+			// React.memo bails out, so this delta is 1 (the Profiler callback itself),
+			// not 0. The child's render work is still bounded — Test 2 proves the
+			// server-list DOM node is preserved across the heartbeat.
+			expect(afterHeartbeatCommits - afterClickCommits).toBeLessThanOrEqual(1)
 		} finally {
 			vitest.useRealTimers()
 		}