• [fix] Make the interpolation of calendar-day percentile thresholds from 365
  to 366 days exact (PR #566). With SciPy 1.18.1, interp1d returns
  inexact values for constant multi-dimensional data (scipy/scipy#26609), so
  DataArray.interp, which icclim used through xclim's
  adjust_doy_calendar, turned neighbouring thresholds of exactly 10.0 into
  values such as 10.000000000000002 or 9.999999999999998. Observations that tie
  with a threshold, for example temperatures reported to 0.1 degC, were then
  counted as exceeding it in count-type percentile indices such as TX90p.
  In a test with temperatures rounded to 0.1 degC over 30 years, the summed
  TX90p and TX10p counts were 25970 instead of 25926 (44 additional
  exceedance days). Affected: 7.2.0, 7.3.0, 7.3.1 and 7.3.2 installed with
  SciPy 1.18.1 (run on a day-of-year percentile count test; SciPy 1.17.1 gives
  the exact result). Results computed with SciPy 1.18.x should be recomputed.
  Other SciPy 1.18 releases were not tested.
• [fix] Compute the calendar-day percentile windows with wrap-around at the
  edges of the percentile base period, for nominal and bootstrap thresholds,
  with the same thresholds in eager and chunked execution (issue #559,
  PR #564). Before, the 5-day window of the first and last days of the base
  period was truncated, which left CD, WD and WW one day off in the
  independent check. Behaviour change: thresholds for the first and last days
  of the year can change slightly, and day-of-year percentile indices with
  them; period percentiles do not change. Affected: 7.3.2 (run, listed there
  as a known issue); earlier releases were not checked.
• [fix] Rolling-window indices such as RX5day no longer fail when a Dask
  time chunk is shorter than the rolling window: only the chunks that are too
  short are merged, other chunk layouts are kept (issue #526, PR #565).
  Confirmed failing with xarray 2026.2.0 and dask 2026.1.2 (run); it did not
  reproduce with xarray 2026.9.0 and dask 2026.8.0, so the affected releases
  are not determined.
• [doc] Add the behaviour-change, ambiguity and change-detector rules to the
  release process (PR #563).