[Daomiyo]

Draft contribution with automated validation completed. Contributor self-review, personal verification, and human-authored rationale remain pending; this is not ready for maintainer review under AI_POLICY.md.

Observed behavior

On the unmodified baseline, a UTC datetime ending in 1970-01-01T00:00:00.000249 is written with 248 microseconds instead of 249. Other dates also lose microseconds, and decoding the maximum Python datetime can overflow.

Automated implementation notes: preserve integer epoch seconds and microseconds separately in Python and Cython, retain local-time conversion for naive values, and decode the bounded fractional component with ties-to-even rounding. The existing 12-byte seconds/nanoseconds representation and public API are unchanged. Added tests inspect exact bytes and values; five Java/Python fixtures exercise both directions.

Automated validation

Validated local snapshot: d2f1d708ddc44492e7dadb5e0b2874df8455919c
Published branch uses this exact commit; all five remote file hashes match the validated manifest.
Base: 80bcf64e68defb2de70e8e9047a1bb5f5d050883

macOS arm64, CPython 3.12.13, NumPy 1.26.4, PyArrow 20.0.0, JDK 25.0.2, Maven 3.10.0. Source and native-extension fingerprints checked before the test run.

• Full Python suite: pure mode 1091 passed / 35 skipped; Cython mode 1096 passed / 30 skipped. Arrow tests ran; remaining skips are Java-driven cases and mode-specific cases.
• Java-driven PythonXlangTest plus PyCrossLanguageTest#testCrossLanguageSerializer: each mode 68 passed / 20 skipped / 0 failures. New timestamp cases executed; skips are existing duplicate exclusions and unsupported union cases.
• Changed Python files pass Ruff format/check; Java test formatting and git diff --check pass.
• Earlier focused timestamp regression run against the unmodified serializers: 20 failures / 26 passes per mode.
• Windows execution was not performed. Its existing current-DST offset behavior is retained.

Commands, from python/:

ENABLE_FORY_CYTHON_SERIALIZATION=0 python -m pytest -q -ra .
ENABLE_FORY_CYTHON_SERIALIZATION=1 python -m pytest -q -ra .

From java/, after the reactor installation and clean test compilation, with the local Python environment on PATH, for MODE=0 and MODE=1:

FORY_PYTHON_JAVA_CI=1 ENABLE_FORY_CYTHON_SERIALIZATION=$MODE ENABLE_FORY_DEBUG_OUTPUT=1 mvn -B -ntp -pl fory-core test '-Dtest=org.apache.fory.xlang.PythonXlangTest,org.apache.fory.xlang.PyCrossLanguageTest#testCrossLanguageSerializer'

Benchmark

Adjacent baseline/change pairs with 120,000 loops × 8 values × 15 repetitions per case, identical frames, and matching before/after source/native fingerprints. These are local CPython 3.12 microbenchmarks. Earlier results are retained in the evidence archive.

Mode	Case	Operation	Baseline ns/value	Changed ns/value	Change
Cython	utc	serialize	235.97	141.49	-40.04%
Cython	utc	deserialize	330.88	181.28	-45.21%
Cython	offset	serialize	359.28	257.78	-28.25%
Cython	offset	deserialize	326.81	183.25	-43.93%
Cython	naive	serialize	398.08	387.28	-2.71%
Cython	naive	deserialize	329.47	177.41	-46.15%
Cython	pre_epoch	serialize	232.29	140.07	-39.70%
Cython	pre_epoch	deserialize	323.59	179.61	-44.49%
Python	utc	serialize	1151.91	1102.96	-4.25%
Python	utc	deserialize	1043.34	1047.84	+0.43%
Python	offset	serialize	1359.34	1266.73	-6.81%
Python	offset	deserialize	1052.47	1029.19	-2.21%
Python	naive	serialize	1336.17	1338.39	+0.17%
Python	naive	deserialize	1056.54	1039.46	-1.62%
Python	pre_epoch	serialize	1195.15	1093.16	-8.53%
Python	pre_epoch	deserialize	1061.24	1054.32	-0.65%

AI Usage Disclosure

• substantial_ai_assistance: yes
• scope: investigation, implementation, tests, performance measurement, automated reviews, and these draft technical notes
• affected_files_or_subsystems: Python/Cython timestamp serializers; Python serializer tests and Java/Python cross-language fixtures
• ai_review: two fresh clean-context automated reviews on the validated local snapshot above reported no actionable findings. Human self-review has not occurred, so the required human-first final review sequence is not claimed complete.
• ai_review_artifacts: automated workflow labels and exact diff identity are recorded below. They are preliminary to the project's required human-first sequence.
• human_verification: pending. All checks described above were executed and inspected by the coding agent, not personally by the submitter.
• performance_verification: adjacent baseline/change local microbenchmarks; raw measurements and before/after source/native fingerprints are prepared with the contribution artifacts.
• provenance_license_confirmation: modifications are to existing Apache-licensed files and introduce no copied third-party code or new dependency; contributor confirmation remains pending.

Automated review records:

• Fory-guided workflow /root/fory_fresh_guided_oct22: no actionable findings after inspecting all five files, repository guidance, arithmetic, timezone/fold semantics, C API bounds, protocol and test evidence.
• Independent general workflow /root/fory_fresh_independent_oct22: no actionable findings after independently inspecting all five files, arithmetic, boundaries, compatibility and coverage.
• Both reviewed the validated local snapshot/base above; retained local git diff SHA-256: 46c175d8c1baa4c8b7aaad72fc23a44addb1bedb2b842f2212ec0e85e2439312.

Remaining contributor steps

• Disclose substantial AI assistance and distinguish automated checks from human checks.
• Personally review all changed lines and explain the implementation and tradeoffs.
• Personally run relevant checks and inspect results.
• Supply human-authored rationale, risk assessment and review responses.
• Complete the two-fresh-reviewer sequence after human self-review and persist final evidence.
• Confirm provenance and license obligations before marking ready for review.

Retained automated evidence

Source fingerprints, raw test output, benchmark measurements, and preliminary automated review records:

fory-timestamp-verification.zip

[Daomiyo]

Closing this draft because I won't be continuing work on it. Anyone interested is welcome to pick up the timestamp precision issue; the patch and automated validation artifacts above remain available as a starting point. The contributor self-review and personal verification listed above remain incomplete.

Automated CI triage for run 39639380150: seven jobs stopped during Python setup because Python 3.8 x64 was unavailable on Ubuntu 26.04. The remaining failure was compilation of the unchanged C++ ResultTest.Assignment test with -Werror=maybe-uninitialized, before the C++ tests ran; the validity of that warning remains unverified.

Thanks to anyone who takes this forward.