NEWS.md
parse_ggir_date() rounds timestamps to the nearest minute before taking the date. GGIR 2.5-1 stamps the last partial day a few seconds before midnight (e.g. 23:59:55), so it collided with the previous day.collate_outputs() and assemble_datasheet() read GGIR output as text. Previously readr guessed part 5 clock times (M16TIME, L16TIME, …) as datetimes and rewrote them in UTC.assemble_datasheet() errors when part 2 dates end in Z. These were converted to UTC before collation and the local date cannot be recovered.assemble_datasheet() now reads calendar_date as text and parses it with a new internal helper, so part 2 dates keep their local day. Previously readr converted part 2’s timestamps to UTC, which moved every midnight day east of UTC back by one day. Parts 4 and 5 are parsed the same way.assemble_datasheet() now keeps part 5 files whose names carry a prefix (e.g. a_part5_daysummary_WW_...). collate_outputs() already accepted these, but the datasheet silently dropped every part 5 column.assemble_datasheet() now errors when no part 5 rows match, and warns on duplicated part 2 dates, part 2 weekdays that disagree with the date, missing SleepRegularityIndex values, missing MM rows, and part 5 data that does not join.assemble_datasheet() on the existing collated output. GGIR does not need to be re-run.