[miguel-heygen]

The BeginFrame capability probe > reports an empty screenshot as unsupported test ran the probe against a real 120 ms deadline. On a slow runner (windows-latest) the deadline passed before the loop reached its last attempt, so the message read beginFrame probe timeout before ... instead of returned 0 bytes after N attempts.

The test now runs under fake timers and advances time itself, so the outcome no longer depends on how fast the machine is. No source change.

fails without the fix: BeginFrame capability probe > reports an empty screenshot as unsupported (the old version, run with every clock read made 25 ms slower, fails with expected 'beginFrame probe timeout before warm-…' to match /returned 0 bytes after \d+ attempts/; the new version passes under the same slow clock)

The file passes 3 runs in a row (117 tests), oxfmt clean.

Size

One test, 11 lines added.

[heygengenesis]

Verdict: APPROVE · 0 blocker · 0 important · 0 nits · head bf74cbbb

Test-only. Vitest 4 fake timers also fake Date, which is the clock the probe uses for its deadline. Each 10 ms retry delay is advanced on purpose, so the loop always stops at Date.now() + 10 >= deadline and throws the 0 bytes after N attempts error, however slow the machine is. try/finally puts real timers back.

Checks

• Ran browserManager.test.ts at head 3 times: 117/117 passed each run.
• The 'fails without the fix' claim holds. With the old test body and each Date.now() read made 25 ms slower, the test fails with expected 'beginFrame probe timeout before warm-…'. Making each CDP send take 40 ms of real time also breaks the old body (…timeout before scree…). The new body passes under the same 40 ms slow-down.
• CI: the one red check, Test (support), fails in studio-server projectHistory.test.ts > never removes a lock another process holds (HistoryBusyError at the lock deadline). That is unrelated to this one-file engine test change. The Windows shards were still running when I reviewed.