Checklist

• The title and commit message(s) are descriptive.
• Small commits made to fix the PR have been squashed.
• Tests have been added for the bug fix.
• The affected API contract is documented.

Description

Fixes #2595 by documenting the ordering precondition for xt::drop: indices must be in increasing order after negative indices are resolved relative to the sliced axis. For an axis of size 6, drop(0, -1) resolves to {0, 5}, whereas drop(-1, 0) is outside this contract. Callers ensure the order; the library performs no sorting or order check.

Reset the cumulative offset map before rebuilding a reused drop slice. With drop(-2, -1), normalization for axis sizes 4 and then 6 uses ordered positions in both cases, but an old offset entry makes slice(2) return 4 instead of 2. A single two-assertion regression covers this stale state. Existing keep behavior is unchanged.

Validation on macOS arm64 with Apple Clang 21 and C++20:

• The regression fails against the original upstream implementation and passes with the fix.
• Official complete test_xtensor_lib: 1,251 cases / 19,719 assertions pass, including optional JSON/MIME coverage; no cases skipped.
• test_xview and test_xdynamic_view pass in Release, combined ASan/UBSan, and default column-major builds. Sanitizer coverage is limited to these related targets; leak detection is disabled on macOS and the existing CMake configuration disables signed-integer-overflow instrumentation.
• Pinned clang-format 17.0.6, standalone-header compilation, new RST paragraph syntax, and git diff --check pass. Full Sphinx/Doxygen output was not built.

The existing xsort compilation warnings and non-failing xinfo type-name warning remain unchanged. AI assistance was used for investigation, implementation, tests, validation, and this description. Duplicate and out-of-range indices remain outside this change's scope.