GeezerVison

Fixes #6. test_writer_bad_message calls timestamp() on a naive datetime, so its expected bytes depend on the host timezone. The two hard-coded timestamps only cover some environments; adding another timestamp as in the closed #7 would leave the underlying problem.

Use a UTC-aware JsonDatetime and keep a single timestamp expectation (1546304461), while retaining the existing empty-output allowance for the standard JSON fallback. Instantiate the subclass directly instead of permanently replacing the global datetime.datetime class.

Validation on Python 3.13.5:

• Before the change, the test passes in UTC and America/New_York but fails in Europe/Luxembourg, Asia/Kathmandu and Pacific/Chatham.
• After the change, all 27 tests pass in each of those five timezones with both ujson 6.0.0 and the standard-library JSON fallback (10 separate processes). Also verified datetime.datetime retains its original identity after each suite run.
• Repository-wide pylint (10.00/10), pycodestyle, pyflakes and git diff --check pass.

This changes tests and macOS CI only; it does not alter production serialization or the separate orjson migration in #29. No Windows/OpenBSD run was performed locally.

CI follow-up:

• The old macOS jobs requested the retired macos-13 runner and were cancelled without executing any steps.
• Switch only the runner label to macos-15-intel, preserving the x64 architecture and Python 3.8/3.9/3.10 matrix.
• YAML semantic comparison confirms only the runner label changed. git diff --check passes, and all 27 tests pass on Linux / Python 3.13.15 with TZ=Europe/Luxembourg. GitHub macOS execution is still required.

AI disclosure: prepared and tested by OpenAI Codex acting on behalf of @saitakarcesme. No independent human review is claimed.