diff --git a/.changeset/3597-windows-argv-overflow.md b/.changeset/3597-windows-argv-overflow.md
new file mode 100644
index 000000000..976b81859
--- /dev/null
+++ b/.changeset/3597-windows-argv-overflow.md
@@ -0,0 +1,5 @@
+---
+type: Fixed
+pr: 3649
+---
+**`scripts/run-tests.cjs` no longer fails on Windows when invoking many test files at once** — Windows `CreateProcess` caps `lpCommandLine` at 32,767 chars, so an unchunked spawn of `node --test <546 paths>` aborted instantly with exit 1 and no test output (Linux/macOS allow ~2 MB so the same path worked there). The harness now batches selected files into chunks whose total argv stays under 28,000 chars and runs each chunk sequentially, reporting `run-tests: chunk N/M — K files` to stderr. The ceiling is overridable via `RUN_TESTS_MAX_CMDLINE_CHARS` for tuning and tests. Adds a cross-platform regression test that forces chunking with a low ceiling.
diff --git a/scripts/run-tests.cjs b/scripts/run-tests.cjs
index 55038b934..e38f47f50 100644
--- a/scripts/run-tests.cjs
+++ b/scripts/run-tests.cjs
@@ -130,14 +130,49 @@ function main() {
     ? `--test-concurrency=${process.env.TEST_CONCURRENCY}`
     : '--test-concurrency=4';
 
-  try {
-    execFileSync(process.execPath, ['--test', concurrency, ...selected], {
-      stdio: 'inherit',
-      env: { ...process.env },
-    });
-  } catch (err) {
-    process.exit(err.status || 1);
+  // Windows `CreateProcess` caps the full command line at 32,767 chars
+  // (lpCommandLine). With 500+ test paths the spawn fails instantly with no
+  // test output. Linux/macOS allow ~2 MB (ARG_MAX) so unchunked spawns are
+  // fine there. Split into chunks sized for the tightest target so behavior
+  // is identical across platforms. (#3597)
+  // Operator override (also used by tests to force chunking with short paths).
+  const MAX_CMDLINE_CHARS = process.env.RUN_TESTS_MAX_CMDLINE_CHARS
+    ? Number(process.env.RUN_TESTS_MAX_CMDLINE_CHARS)
+    : 28000; // headroom below the 32,767 Windows ceiling
+  const FIXED_OVERHEAD = process.execPath.length + '--test'.length + concurrency.length + 8;
+  const chunks = [];
+  let current = [];
+  let currentLen = FIXED_OVERHEAD;
+  for (const file of selected) {
+    const add = file.length + 1; // +1 for the inter-arg separator
+    if (current.length > 0 && currentLen + add > MAX_CMDLINE_CHARS) {
+      chunks.push(current);
+      current = [];
+      currentLen = FIXED_OVERHEAD;
+    }
+    current.push(file);
+    currentLen += add;
   }
+  if (current.length > 0) chunks.push(current);
+
+  let firstFailureExit = 0;
+  for (let i = 0; i < chunks.length; i++) {
+    if (chunks.length > 1) {
+      console.error(`run-tests: chunk ${i + 1}/${chunks.length} — ${chunks[i].length} files`);
+    }
+    try {
+      execFileSync(process.execPath, ['--test', concurrency, ...chunks[i]], {
+        stdio: 'inherit',
+        env: { ...process.env },
+      });
+    } catch (err) {
+      const code = err.status || 1;
+      // Run every chunk so the operator sees all failures in one pass; report
+      // the first non-zero exit at the end.
+      if (firstFailureExit === 0) firstFailureExit = code;
+    }
+  }
+  if (firstFailureExit !== 0) process.exit(firstFailureExit);
 }
 
 main();
diff --git a/tests/run-tests-harness.test.cjs b/tests/run-tests-harness.test.cjs
index ad579c4e6..75cb12155 100644
--- a/tests/run-tests-harness.test.cjs
+++ b/tests/run-tests-harness.test.cjs
@@ -34,10 +34,10 @@ function seed(dir, names) {
   }
 }
 
-function runHarness(testDir, args = []) {
+function runHarness(testDir, args = [], extraEnv = {}) {
   // Clear node:test parent-context env so the harness's child `node --test`
   // doesn't refuse to run with "recursive run() skipping running files".
-  const env = { ...process.env, GSD_TEST_DIR: testDir };
+  const env = { ...process.env, GSD_TEST_DIR: testDir, ...extraEnv };
   delete env.NODE_TEST_CONTEXT;
   return spawnSync(process.execPath, [HARNESS, ...args], {
     cwd: path.join(__dirname, '..'),
@@ -207,4 +207,34 @@ test('boom', () => { throw new Error('intentional'); });
       );
     });
   });
+
+  describe('Windows argv-overflow chunking (issue #3597)', () => {
+    // Windows CreateProcess caps lpCommandLine at 32,767 chars. With ~550
+    // tests the unchunked spawn fails instantly on Windows with no test
+    // output. Linux/macOS allow ~2 MB so the same path works there. The
+    // harness chunks selected files so each spawn stays under the ceiling,
+    // and chunking is observable via the `run-tests: chunk N/M …` stderr
+    // line. Long filenames force chunking even with a modest file count so
+    // the test stays fast on every platform.
+    test('chunks when total argv would exceed configured ceiling', () => {
+      // Use a deliberately low MAX_CMDLINE_CHARS so the test is independent
+      // of tmp-path length (varies by OS). With a 2000-char ceiling and 30
+      // tests at ≥100 char paths, chunking must engage and at least one
+      // `chunk N/M …` marker must appear in stderr.
+      const longPrefix = 'a-deliberately-long-test-filename-to-force-chunking-behavior-cross-platform-';
+      const names = Array.from({ length: 30 }, (_, i) => `${longPrefix}${String(i).padStart(4, '0')}.test.cjs`);
+      seed(tmpDir, names);
+      const r = runHarness(tmpDir, [], { RUN_TESTS_MAX_CMDLINE_CHARS: '2000' });
+      assert.strictEqual(
+        r.status,
+        0,
+        `expected zero exit; got status=${r.status} signal=${r.signal}\nSTDERR (tail):\n${r.stderr.split('\n').slice(-20).join('\n')}`,
+      );
+      assert.match(
+        r.stderr,
+        /run-tests: chunk \d+\/\d+ — \d+ files/,
+        `expected chunking marker in stderr; STDERR (tail):\n${r.stderr.split('\n').slice(-20).join('\n')}`,
+      );
+    });
+  });
 });