* fix(relative): add the fraction of a negative relative second count
`Relative::Seconds` holds a floor decomposition, so -0.25 seconds is
`(-1, 750_000_000)`. A `jiff::Span` is sign-uniform, so combining those two
fields directly subtracted the fraction instead of adding it and the result
was off by `2 * (1 - fraction)` seconds. Rebalance the fraction onto the sign
of the whole seconds before building the span.
* fix(relative): shorten the fraction comment, align the GNU version to 9.11
* fix(weekday): accept a period after abbreviated weekday names
* test(weekday): parse abbreviated weekdays with a period through the full parser
* test(weekday): cover mismatched abbreviated weekday with period
The weekday benchmark parsed "wednesday" relative to the current date, so
its cost depended on which day of the week CI ran (no date shift on a
Wednesday, a multi-day shift otherwise), producing spurious regressions.
The table is meant to cover the same abbreviations GNU date supports, but
five of them resolved to a different offset than GNU date gives, so
"2026-06-15 12:00 BST" parsed as +06:00 (Bangladesh) rather than +01:00
(British Summer Time). Offsets were measured with `date --debug` on GNU
coreutils 9.10: ADT is -3, AST -4, BST +1, GST +10 and SST -12.
When the input carries both a weekday and a calendar date, the date is
now authoritative and the weekday is dropped, matching GNU date. This
also covers ordinal forms such as "next friday sep 25 2026". Relative
items are still applied on top of the date, and a weekday without a
date behaves as before.
Closes#317
renovate[bot]
xtqqczze
Sylvestre Ledru
Ion
Anas Awadalla
CodSpeed Bot
Asifur Rahaman Meeru
sap1110
oech3
Anas Awadalla
Natsu