NumPy ↔ Production Python Runtime: np.load allow_pickle interoperability

NumPy I/O for .npy files and production Python runtime security policy interact at the allow_pickle boundary. The goal is to define a portable persistence approach for NumPy arrays that remains valid when the same code and files move from development to a hardened production environment.

The relevant constraint is that object arrays require pickle support to deserialize, while production runtimes often restrict pickle for security. NumPy version differences between local and production also affect default handling and error semantics for np.load.

What is the documented default and version-specific behavior of np.load with respect to allow_pickle for object arrays? When does deserialization of an object array become policy dependent rather than format dependent? What persistence alternatives are supported by NumPy for object dtype data that avoid pickle reliance across environments?