On 29 December 2011, Samoa moved from one side of the date line to the other. Clocks in Apia went from 23:59:59 on Thursday 29 December straight to 00:00 on Saturday 31 December. Friday 30 December never happened there. The tz database records it as a jump from UTC-10:00 to UTC+14:00.
A daily calendar event at 12:00 in Pacific/Apia still has to land somewhere on that Friday. RFC 5545, section 3.3.5, says what to do with a local time that doesn't exist: read it with the UTC offset in force before the gap. So 30 December 12:00 is read at -10:00, which is 22:00 UTC on the 30th, the same instant as Saturday's 12:00 at +14:00.
What the RRULE explainer did instead
The explainer lists each occurrence of a rule with its UTC instant and marks times that fall in a clock-change gap. To find "the offset before the gap" it looked three hours before the moment of the shift. That works for every daylight-saving change, since those gaps are an hour or less. A 24-hour gap is longer than three hours, so the look-back still landed after the shift and picked up +14:00. The rule
DTSTART;TZID=Pacific/Apia:20111228T120000
RRULE:FREQ=DAILY;COUNT=5
listed 30 December at 2011-12-29T22:00:00Z, the same instant as the 29 December occurrence the line above it. The UTC column went backwards. Kiritimati, which skipped 31 December 1994 the same way, did the same.
The fix doesn't use a fixed look-back. Read with the later offset, any skipped wall time lands on an instant before the shift, so the offset at that instant is the one the RFC wants, however long the gap. New York's skipped 02:30 and Lord Howe Island's half-hour change give the same answers as before. The test pins all three cases.
How the bug turned up
This was a red-team run: a script seeds small faults into the source, one at a time, and checks whether the test suite fails. Of 12 faults in rrule.mjs, 8 were caught. One of the 4 that got through changed a two-day look-back to twelve hours. Working out whether that could ever matter led to zones more than 12 hours ahead of UTC, then to Apia's history, then to the real bug in the line below it.
That fault turned out to matter too. With a twelve-hour look-back, Apia's last daylight-saving end (4 April 2021, 04:00 back to 03:00) lost its earlier reading of 03:30, and the explainer would have called a repeated hour an ordinary one. There's a test for that now as well. Two of the other escaped faults were plain test gaps: BYMINUTE=-5 (minutes take no sign) and the point where the English description stops listing clock times. The fourth fault couldn't change anything a user sees.
After the new tests, all 11 faults that could change behaviour are caught. A second campaign on the Australian redundancy pay calculator caught 10 of 10.
These dates are rare, but people do use the explainer to check rules against historical data and odd zones, and a checker that disagrees with the RFC on the rare cases is the kind people stop trusting.