[ananya-kaul]

FIX Preserve NumPy masked-array fill values

Saving and loading a masked array currently discards its custom fill value. For example, an integer array with fill_value=-99 uses 999999 after loading, changing the result of filled().

Persist the fill value using the existing NumPy node and restore it when constructing the masked array. Older files without this field retain their default behavior. The new child is included in the type audit and restricted to NdArrayNode.

Validation: 12 round-trip cases fail before the fix and pass afterward, covering integer, floating-point and string data, masked/unmasked arrays, and file/bytes APIs. A compatibility test covers files without the new field, and an audit regression rejects an unexpected child type. Persistence, legacy persistence and audit suites: 1083 passed, 1 skipped, 1 xpassed. Ruff lint/format and focused pyrefly check pass. Tested on macOS with Python 3.13.5; the full project suite and other platforms were not run.

[Copilot]

🟡 Changes recommended

The fill-value node restriction rejects supported object-dtype masked arrays with Python string fills, breaking round-trip loading and untrusted-type inspection.

1 open issue

• Object-dtype masked arrays with string fills fail serialization round-trip · New

What changed in this PR

This PR updates NumPy masked-array persistence to preserve custom fill values, while retaining default behavior for older files without a saved fill value.

Changes:

• Save and restore fill values and include them in type auditing.
• Add round-trip, backward-compatibility, and unexpected-child-type tests.
• Document the fix in the changelog.

File	Description
skops/​io/​tests/​test_masked_array.py	Adds fill-value round-trip and backward-compatibility tests.
skops/​io/​tests/​test_audit.py	Adds rejection coverage for unexpected fill-value node types.
skops/​io/​_numpy.py	Serializes, audits, and restores masked-array fill values.
docs/​changes.rst	Records the persistence fix.

_{Review effort: Balanced}