[no-ci] cuda.core: fix host memcpy source pointer cast

[rwgk]

Description

This extracts the one-line host memcpy source-pointer fix from #2470 so it can merge independently while the broader CI dependency-selection work continues through review and coordination with #2464.

Landing the correction separately prevents this known source-build failure from continuing to block work based on current main.

Root cause

PR #2395 refactored graph memcpy parameter initialization and inadvertently changed the CUDA_MEMCPY3D.srcHost assignment from a void * cast to a const void * cast.

The current in-tree cuda-bindings declaration exposes CUDA_MEMCPY3D.srcHost as void *. Cython therefore rejects the assignment because it would discard the const qualifier:

Assigning to 'void *' from 'const void *' discards const qualifier

The regression initially escaped CI because cuda.core was being Cythonized against a published cuda-bindings wheel from PyPI rather than the wheel built from the same checkout. That published declaration accepted the const-qualified pointer. The current bindings wheel was installed only after the core wheel had already been Cythonized, so subsequent tests could not detect the mismatch.

Fix

Cast the source address to void *, matching the current in-tree declaration.

A mutable pointer remains assignable to an older declaration accepting const void *, so the correction remains compatible with the intentional published-bindings compatibility builds. It does not change structure layout, ABI, ownership, or runtime copy behavior.

Why this is separate and [no-ci]

This exact commit was extracted unchanged from #2470. There, the local-wheel dependency-selection fix forced cuda.core to build against the same-checkout cuda-bindings wheel, exposed the original compiler error, and fully exercised this correction across the relevant Linux and Windows wheel builds. The sole failure in the latest #2470 run occurred later in an unrelated Windows NumPy source build.

A standalone PR based on main does not contain #2470's dependency-selection fix. Its normal CI can therefore take the same published-wheel path that masked the regression and would not independently validate this compatibility boundary. Re-running the full matrix would provide little additional signal over the directly relevant coverage already completed in #2470.

Splitting the correction out lets the blocker merge quickly. PR #2464 independently contains the same one-line correction as part of broader selective-wheel-build work; whichever lands second can drop the overlapping patch.

Validation

• The exact change was exercised by same-checkout bindings/core builds in Fix three recent cuda.core build and CI issues #2470.
• pre-commit run --all-files passes on this extracted branch.
• git diff --check passes.

Related: #2395, #2468, #2464, #2470.

Checklist

• Existing build coverage in Fix three recent cuda.core build and CI issues #2470 exercises this compile-time correction.
• Documentation is up to date; this internal type correction has no user-facing behavior change.

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.