[heathsanchez]

Problem

Rounding a UTC instant near the end of a minute can produce a civil second that does not exist. On current master:

from skyfield.api import load
ts = load.timescale()

print(ts.utc(2016, 12, 30, 23, 59, 59.9995).utc_iso(places=3))
# Before: 2016-12-30T23:59:60.000Z (no leap second that day)

print(ts.utc(2016, 12, 31, 23, 59, 60.9995).utc_iso(places=3))
# Before: 2016-12-31T23:59:61.000Z (no second 61, even on a leap day)

utc_jpl() exhibits the same rollover problem at four decimal places, with 59.99995 and 60.99995. These are formatting labels, not evidence that Skyfield's internal time calculations are incorrect.

Cause and fix

_utc_seconds(offset) adds half a display unit before splitting the TAI seconds. Near the boundary, the fractional remainder can be just below 1.0 (for example, 0.9999999999999976), then round up when _utc_tuple() adds it to an integer civil second. That last rounding happens after the leap-offset/calendar decision, yielding impossible :60 or :61 labels.

This patch normalizes the near-integer carry before applying the leap-second offset, and adds native regression tests for ordinary-minute and leap-minute rollovers in ISO and JPL formatting. The change is limited to skyfield/timelib.py and skyfield/tests/test_timelib.py.

Verification

• Original source at 31a6ed7584cd09b6e48e07546a4b5921e1cd5bb4: existing native suite 584/584 pass, but the new boundary witness fails.
• This patch: native assay --batch skyfield.tests 586/586 pass (Python 3.12, NumPy 1.26.4), pyflakes skyfield clean, and git diff --check clean.
• Separate controls exercised scalar/vector agreement, ordinary midnight, fractional leap seconds, and the 1973, 2012 and 2016 leap-second insertions.
• Upstream GitHub CI and behavior across all floating-point edge cases remain to be verified by this PR.

The report and correction were prepared with AI assistance. The unchanged upstream source was used for the original-head reproduction, and native controls and tests were run before submission. The claim is limited to impossible UTC labels. Maintainer feedback on the near-carry threshold or a preferred alternative is welcome.

[brandon-rhodes]

Oh, that's a very bad behavior — thanks for letting me know! I'll plan to look at this sometime this week, and hopefully get it merged. I'll have to plan for a Skyfield release near the end of the month, I suppose.

[heathsanchez]

No problem Brandon, it's a brilliant project- happy to help. Have a great week!