test_runner: backport #64706 to v22.x — non-ASCII test stdout can silently drop a whole test file

[positivef]

Request

Please consider backporting #64706 (fix for #64061) to the v22.x release line.

State as of 2026-09-09, checked with the GitHub contents API against lib/internal/test_runner/runner.js, looking for ) >>> 0) + kSerializedSizeHeader:

ref	has the fix
main	yes
v26.x	yes
v24.x, v24.x-staging	yes
v22.x, v22.x-staging	no

#64706 carries no dont-land-on-v22.x label, and I could not find an existing backport request or backport PR for v22.x.

The line is actively released — the most recent v22 release is v22.23.2 (2026-07-29), three days after #64706 landed on main (2026-07-26).

Why this may deserve more weight than the original report suggested

#64061 was reported as an intermittent CI flake. On v22.19.0 I traced the trigger, and it turns out not to be exotic: it is ordinary non-ASCII text written to stdout by a test.

FileTest.#processRawBuffer advances bufferHead past a consumed message and then reads the next four bytes as a length without re-checking for the FF 0F header:

bufferHead = TypedArrayPrototypeSubarray(concatenatedBuffer, fullMessageSize);
this.#rawBufferSize = TypedArrayPrototypeGetLength(bufferHead);
...
while (bufferHead?.length >= kSerializedSizeHeader) {
  const fullMessageSize = (
    bufferHead[kV8HeaderLength] << 24 |
    bufferHead[kV8HeaderLength + 1] << 16 |
    bufferHead[kV8HeaderLength + 2] << 8 |
    bufferHead[kV8HeaderLength + 3]
  ) + kSerializedSizeHeader;          // signed 32-bit

  if (this.#rawBufferSize < fullMessageSize) break;

When a test's own stdout follows a report message inside the same read, those bytes are interpreted as a length. Every UTF-8 lead/continuation byte is >= 0x80, so any non-ASCII output — CJK, emoji, accented Latin — can make fullMessageSize negative, which defeats the break guard and hands garbage to DefaultDeserializer.

In our repository the emitter was a server start-up banner containing an emoji, printed by the code under test. Three sightings over two days, always the same test file.

Minimal deterministic reproduction

No load, no concurrency, no flakiness. One write():

// repro.mjs   —   node --test repro.mjs
import { writeSync } from 'node:fs';
import test from 'node:test';
import { DefaultSerializer } from 'node:v8';

test('a passing test', () => {});

const s = new DefaultSerializer();
s.writeHeader();
s.writeValue({ __proto__: null, type: 'test:diagnostic',
               data: { __proto__: null, nesting: 0, message: 'hi', file: 'repro.mjs' } });
const payload = s.releaseBuffer();
const size = Buffer.alloc(4);
size.writeUInt32BE(payload.length);
const message = Buffer.concat([Buffer.from([0xff, 0x0f]), size, payload]);

// a report message, immediately followed by ordinary test stdout, in ONE write
const tail = process.env.ASCII ? '\nserver listening -> http://localhost:3456\n'
                               : '\n\u{1F7E1} 서버 실행 중 -> http://localhost:3456\n';
writeSync(1, Buffer.concat([message, Buffer.from(tail, 'utf8')]));

On v22.19.0, Windows 11 x64:

$ node --test repro.mjs
not ok 1 - repro.mjs
  error: 'Unable to deserialize cloned data due to invalid or unsupported version.'
# fail 1

$ ASCII=1 node --test repro.mjs
ok 1 - a passing test
# fail 0

The only difference between the two runs is whether the trailing stdout is ASCII.

I only have v22.19.0 on this machine, so I have not run this against a fixed release line — the report above is strictly about v22.19.0.

Three symptoms, and one of them is silent

Also observed on v22.19.0, depending on where in the stream the corruption lands:

1. a file-level failure carrying the message above (the symptom in Error: Unable to deserialize cloned data due to invalid or unsupported version #64061);
2. the file's results vanish with no failure and no summary. Running a poisoned file together with a healthy one printed only the healthy file's ok 1 / ok 2 — no 1..N, no # tests, no # fail, no error text. The exit code was 1, but the TAP body reads as a clean pass;
3. with a length byte below 0x80 but large, the runner waits for a message that never arrives and hangs.

Symptom 2 is the reason I am filing this: on a line that is still in service, a whole test file can drop out of a run while the output looks green, and only the exit code disagrees.

Offer

If the missing step is just a backport-requested-v22.x label or an approval, I am happy to open the [v22.x backport] PR against v22.x-staging — please say so.

[abhinav-phi]

I'll handle the v22.x backport of #64706 - apply the fix to the v22.x test_runner code path, resolve whatever drift exists against the branch, and verify the regression test passes on v22.x before opening the PR.

/assign
take