diff --git a/packages/server/scripts/measure-file-observer.test.ts b/packages/server/scripts/measure-file-observer.test.ts
new file mode 100644
index 0000000000..345da462bb
--- /dev/null
+++ b/packages/server/scripts/measure-file-observer.test.ts
@@ -0,0 +1,59 @@
+import { execFile } from "node:child_process";
+import { mkdtemp, rm } from "node:fs/promises";
+import { tmpdir } from "node:os";
+import { join } from "node:path";
+import { fileURLToPath } from "node:url";
+import { promisify } from "node:util";
+import { expect, test } from "vitest";
+
+const execFileAsync = promisify(execFile);
+
+// Linux delivers the existing inventory as one batch. Other backends have
+// different startup semantics, so they do not exercise this regression.
+test.skipIf(process.platform !== "linux").each(["healthy", "directory-churn"])(
+  "%s benchmark collects a startup batch larger than the argument limit",
+  async (mode) => {
+    const temporaryRoot = await mkdtemp(join(tmpdir(), "paseo-watch-benchmark-test-"));
+    try {
+      const { stdout } = await execFileAsync(
+        process.execPath,
+        [
+          // Lower the argument limit instead of creating hundreds of thousands
+          // of files. This still runs the real observer and benchmark CLI.
+          "--stack-size=256",
+          "--import",
+          "tsx",
+          fileURLToPath(new URL("./measure-file-observer.ts", import.meta.url)),
+          "--child",
+        ],
+        {
+          timeout: 60_000,
+          env: {
+            ...process.env,
+            TMPDIR: temporaryRoot,
+            PASEO_WATCH_BENCH_MODE: mode,
+            PASEO_WATCH_BENCH_ROOTS: "1",
+            PASEO_WATCH_BENCH_DIRS: "10",
+            PASEO_WATCH_BENCH_PREEXISTING_FILES: "40000",
+            PASEO_WATCH_BENCH_IGNORED_FILES: "1",
+            PASEO_WATCH_BENCH_EDITS: "1",
+            PASEO_WATCH_BENCH_SUSTAINED_EDITS: "1",
+            PASEO_WATCH_BENCH_SUSTAINED_INTERVAL_MS: "1",
+            PASEO_WATCH_BENCH_CHURN_DIRECTORIES: "10",
+            PASEO_WATCH_BENCH_CHURN_FILES_PER_DIRECTORY: "4000",
+            PASEO_WATCH_BENCH_CHURN_ROUNDS: "1",
+            PASEO_WATCH_BENCH_CHURN_DIRECTORIES_PER_ROUND: "2",
+          },
+        },
+      );
+      const measurement = JSON.parse(stdout);
+      expect(measurement.startupEvents).toBe(40_000);
+      expect(measurement.deliveredEvents).toBeGreaterThan(measurement.startupEvents);
+      expect(measurement.missedTrackedPaths).toBe(0);
+      expect(measurement.ignoredEvents).toBe(0);
+    } finally {
+      await rm(temporaryRoot, { recursive: true, force: true });
+    }
+  },
+  90_000,
+);
diff --git a/packages/server/scripts/measure-file-observer.ts b/packages/server/scripts/measure-file-observer.ts
index 49a54faf4d..2e37f9a414 100644
--- a/packages/server/scripts/measure-file-observer.ts
+++ b/packages/server/scripts/measure-file-observer.ts
@@ -199,8 +199,8 @@ async function measure(run: number): Promise<Measurement> {
         root,
         (nextError, events) => {
           error ??= nextError;
-          delivered.push(...events);
           for (const event of events) {
+            delivered.push(event);
             if (!firstObservedAt.has(event.path))
               firstObservedAt.set(event.path, performance.now());
           }
@@ -456,7 +456,7 @@ async function measureDirectoryChurn(run: number): Promise<Measurement> {
     const setupStarted = performance.now();
     subscription = await observer.subscribe(root, (nextError, events) => {
       error ??= nextError;
-      delivered.push(...events);
+      for (const event of events) delivered.push(event);
     });
     const setupMs = performance.now() - setupStarted;