[heathsanchez]

Summary

Fix the one-second GPS-to-UTC conversion error immediately before positive leap-second insertions. The existing tow_to_datetime() selects a leap-second offset by comparing a GPS-clock timestamp against UTC transition dates, applying the new offset prematurely.

This change:

• Selects the offset using the corresponding transition on the GPS-clock timescale.
• Preserves the public API, get_leap_seconds() behavior, and existing dependencies.
• Preserves the existing lossy convention for the inserted leap second; Python datetime cannot represent 23:59:60, and this PR does not introduce that representation.

Reproduction

Before the fix:

utc = datetime(2016, 12, 31, 23, 59, 59, tzinfo=timezone.utc)
week, tow = datetime_to_tow(utc)  # 1930, 16.0
tow_to_datetime(week, tow)       # Incorrectly returns 23:59:58 UTC

After the fix, the inverse returns 2016-12-31 23:59:59 UTC. Dependent public datetime and Unix-millisecond conversions also produce the expected values.

Tests and evidence

• Tested all 18 recorded positive leap-second insertions, with 30 ordinary UTC timestamps immediately before each (540 cases). Incorrect round trips dropped from 153 on the original release snapshot to zero with this patch.
• Added representative 1982, 2015, and 2016 round-trip regressions; independently specified expected values for four public conversion paths; and tests for the inserted-second mapping and pre-GPS-epoch rejection.
• Added a synthetic downstream prepare_kaggle_submission() regression checking UnixTimeMillis. No claim is made that real competition data was affected.
• Local Python 3.12 offline suite: 690 passed, 4 deselected, 2 dependency deprecation warnings. The four deselected tests require external FTP/HTTP; upstream CI remains to be verified.
• Coverage compared with the original release-branch snapshot: 97.6816% → 97.6870% overall; 100% in time_conversions.py.
• Pylint 2.17.7: 10.00/10 for the modified production file. A Python-3.12-compatible lint comparison found no new errors or warnings in the touched files.
• git diff --check passed.

This PR targets the current v1.1.1 release branch, following the contributing guide. There are no API changes, new dependencies, or directory changes.

Fixes #192

AI assistance: OpenAI Codex helped reconstruct and independently test the change.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
⚠️ Please upload report for BASE (v1.1.1@3ca808f). Learn more about missing BASE report.

Additional details and impacted files

@@            Coverage Diff            @@
##             v1.1.1     #201   +/-   ##
=========================================
  Coverage          ?   99.61%           
=========================================
  Files             ?       27           
  Lines             ?     3908           
  Branches          ?        0           
=========================================
  Hits              ?     3893           
  Misses            ?       15           
  Partials          ?        0           

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.