Fix numpy view py/id desync when base is shared (#449) - #612

• If this PR changes any non-documentation portion of the codebase, I have updated the CHANGES.rst file for this change.
  • In the CHANGES.rst update, I added a link to the PR number that this will be (e.g. (+611)) and (if applicable) added a link to the issue that this fixes (e.g. (#608)).
• If this PR fixes a bug or adds a feature, I have added appropriate tests.
• I have run the tests with pytest and run formatting with ruff format.
• If generative AI of any kind was used in creating this PR, I have followed the these guidelines:
  • I ensured that this PR is not entirely written by any generative AI model.
  • I ensured that any portions of the code written by any generative AI model are free from copyrighted code.
  • I ensured that I included a Co-authored-by or Assisted-by tag in the footer containing the name of all generative AI models used.
  • I ensured that all code in this PR submitted verbatim from generative AI output is limited to creating/fixing tests/documentation and/or fixing bugs in extensions, and does not impact the core code.
  • I understand that if I used AI and the above checkmarks are not all true when I submit this PR, my PR will be closed.
• I understand that if this checklist is deleted from the PR body, my PR will be closed.

Summary

Fixes #449. Round-tripping a non-contiguous NumPy view whose base is shared with another view raised:

ValueError: strides is incompatible with shape of requested array and size of buffer

The maintainer confirmed an off-by-one in the encode step (a py/id emitted as N+1 instead of N), so the base-array reference resolved to the wrong object during restore — a 0-dim array wrapping an unpickler._Proxy, as reported in the issue thread.

Root cause

When a view is stored by reference to a base array, NumpyNDArrayHandlerView.flatten also emits a decorative values entry, documented as "not used in restore since base is present, but include values for human-readability". That decoration went through self.context.flatten(obj.tolist(), reset=False), which registers a py/id reference in the pickler's bookkeeping.

Restore, however, ignores values entirely when a base is present (it rebuilds the array from buffer/offset/strides) and therefore never registers a matching reference. The encode and decode id counters then drift apart by one for every referenced array carrying a decorative values. A later view stored as {"py/id": N} to a shared base then resolves to the wrong object during restore, producing the strides error.

This is easy to trigger with e.g. sklearn's cv_results_ (the issue's repro): a masked array registers an unmirrored reference and the splitN_test_score views share a single base.

Fix

Emit the decorative values without allocating a py/id slot, by snapshotting and restoring the pickler's reference bookkeeping (_objs / _flattened) around the call. This keeps the encode and decode id counters in lockstep while preserving the human-readable values in the output. The change is confined to the numpy extension (jsonpickle/ext/numpy.py); no core code is touched.

Tests / Verification

• Added tests/numpy_test.py::test_shared_base_after_referenced_object, a minimal regression that reproduces the exact ValueError on the pristine code and passes with the fix (masked array + two non-contiguous views sharing a base).
• The original sklearn cv_results_ repro from the issue now round-trips correctly with all arrays equal.
• pytest tests/numpy_test.py tests/sklearn_test.py tests/jsonpickle_test.py — all green (221 passed).
• Verified object-dtype views, 2D non-contiguous views, and unpicklable=False still round-trip correctly.
• ruff check, ruff format --check, and isort --profile=black are clean; mypy --strict introduces no new errors in the changed file.

Assisted-by: Claude (Anthropic)

Disclosure: prepared with AI assistance; reviewed and verified locally.