[max-models]

Copy and device semantics

• copy=None | True | False on every interop conversion (as in the array API): False raises when the conversion cannot share memory, True always copies.
• Memory is read from __dlpack_device__ (Apple's unified memory counts as host). from_framework copies between host and device explicitly instead of falling back to asarray silently; a framework that cannot take an array's device raises a ValueError naming the device.
• Transfer profiler: a conversion that shares memory records nothing (so assert_no_transfers checks zero copy), host↔device records to_host / to_device, a copy within one memory device_copy.
• from_jax raises a clear TypeError for traced values.

Explicit dtypes

• No silent float64 → float32: JAX without jax_enable_x64, PyTorch on MPS and MLX for the GPU raise a TypeError with advice. MLX keeps float64 for device=mlx.core.cpu.
• supports_float64(device) in the JAX, PyTorch and MLX adapters, interop.jax.enable_x64().
• Fixes to_mlx(device=...), which passed device to mlx.core.array (it takes none).

Julia adapter (host only)

• cunumpy.interop.julia (extra cunumpy[julia], juliacall): to_julia / from_julia between NumPy arrays and Julia Arrays sharing memory, kept alive across both GCs. C order arrives with reversed dimensions (a component-major (3, n) array is a Julia n×3 matrix), F order with the same; from_julia(order="C" | "F"). Strided or read-only arrays are copied; non-Array Julia arrays are collected.
• CuPy ↔ CUDA.jl is left for later (needs a GPU).

Tests

Shared rules tested with fake device arrays; the framework and Julia tests run with the real JAX, PyTorch, MLX and juliacall and are skipped without them (CI installs none of them yet, except MLX on macOS). Locally: 810 passed, interop coverage 98%.