Cached numpy / numeric-column values are read-only and alias the L1 buffer

[27Bslash6]

Summary

np.frombuffer is used to reconstruct numpy arrays and numeric DataFrame/Series columns. The returned arrays are writeable=False, owndata=False, and alias the source bytes. On an L1 hit they alias the live L1-cached bytes.

Evidence

• src/cachekit/serializers/auto_serializer.py:642 (numpy), :705 (numeric DataFrame col), :766 (numeric Series)
• Empirically: np.frombuffer(...).reshape(...) → flags.writeable=False; mutation raises ValueError: assignment destination is read-only

Impact

1. Behavioral regression: a caller who mutates a returned array/frame gets a crash that an uncached call would not produce.
2. Aliasing risk: on L1 hits the returned array shares memory with the cached entry; mutating-by-copy patterns can be surprising. DATA-IS-SACRED adjacency.

Fix

Return a writable copy (np.array(frombuffer(...))) or document the read-only contract explicitly and copy-on-return for the L1 path.