ISO 8601 durations: pd.Timedelta('PT1.5H') is 5 h 1 s

I checked 6000 random ISO durations (no Y/M) against an exact Fraction oracle. pandas 3.1.1 Timedelta gets every fraction on a non-seconds component wrong: 'PT1.5H' gives 05:00:01, 'P1.5D' gives 5 days 00:00:01, 'PT0.01M' gives 60 s (open since 2023 as pandas#51882, which also lists P1M read as minutes, PT1S1S1S1S and so on). In JS, dayjs 1.11.23's duration plugin drops the leading minus ('-P1W' is +7 days) and turns a decimal comma into 0 ('PT1,5S' is 0 s); upstream already has several open PRs. iso8601-duration 2.2.0 also drops the sign and returns NaN from toSeconds past about 2.7e8 days. pydantic 2.14 rejects PT9999999H as 'too large' but accepts P999999999D (the seconds field looks like it's u32), and it accepts duplicate or out-of-order units (PT1H1H = 2 h). If you need exact durations, convert them yourself or use isodate, tinyduration or moment; all three were within float/µs rounding on valid input.