[headtr1ck]

Rolling argmin/argmax return the position within the padded window, so windows that extend beyond the array count the padding. This is most visible for cumulative, whose windows span the whole dimension: da.cumulative("time").argmax() returned [4, 4, 3, 4, 4] instead of [0, 1, 1, 3, 4] for [1, 2, 1.5, 3.5, 4].

• The indices now start at the first value of the window within the array, which is consistent with the windows of Rolling.__iter__. Interior windows are unchanged, and cumulative returns absolute indices without needing a special case.
• Fixed in DataArrayRolling._array_reduce, so it applies to the numpy, bottleneck and dask paths, and to Dataset.
• The window-local indices go back to FIX: reverse index output of bottleneck move_argmax/move_argmin functions #8552 (GH rolling argmin() or argmax() gives the wrong result (are they inverted?) #8541), which made the bottleneck path consistent with the padded numpy path; padding at the edges wasn't considered there.

Supersedes #11412.

• Closes cumulate+argmax uses padded index instead of absolute index #11336
• Tests added
• User visible changes (including notable bug fixes) are documented in whats-new.rst

AI Disclosure

• This PR contains AI-generated content.
  • I have tested any AI-generated content in my PR.
  • I take responsibility for any AI-generated content in my PR.
    Tools: Claude Code

[This is Claude Code on behalf of Michael Niklas]

🤖 Generated with Claude Code