1Implementors must uphold all of the following for **every** call:
 2
 3* **Memory domains**: Pointers produced by to_ref() / to_mut() must be valid
 4  in the target execution domain for Self (e.g., CPU host memory for CPU,
 5  device memory for a specific GPU). If host↔device transfers are required,
 6  perform them inside the implementation; do not assume the caller synchronized.
 7
 8* **Alignment & layout**: All data must match the layout, stride, and element
 9  size expected by the kernel. size(), rows(), cols_in(), cols_out(),
10  n(), etc... must be interpreted identically to the reference CPU implementation.
11
12* **Scratch lifetime**: Any scratch obtained from scratch.tmp_slice(...) (or a
13  backend-specific variant) must remain valid for the duration of the call; it
14  may be reused by the caller afterwards. Do not retain pointers past return.
15
16* **Synchronization**: The call must appear **logically synchronous** to the
17  caller. If you enqueue asynchronous work (e.g., CUDA streams), you must
18  ensure completion before returning or clearly document and implement a
19  synchronization contract used by all backends consistently.
20
21* **Aliasing & overlaps**: If res, a, b, etc... alias or overlap in ways
22  that violate your kernel’s requirements, you must either handle safely or reject
23  with a defined error path (e.g., debug assert). Never trigger UB.
24
25* **Numerical contract**: For modular/integer arithmetic, results must be
26  bit-exact to the specification. For floating-point, any permitted tolerance
27  must be documented and consistent with the crate’s guarantees.