fix(macos): wrap before_build with arch -x86_64 under Rosetta

[henryiii]

🤖 AI text below 🤖

Problem

On a macOS arm64 host building x86_64 wheels via Rosetta (common now that GHA deprecated macos-13), the command hooks were wrapped with arch -x86_64 inconsistently:

• before_test and the test command run under arch -x86_64 (test loop: shell_with_arch = functools.partial(call, *arch_prefix, "/bin/sh", "-c") with arch_prefix = ["arch", "-x86_64"]).
• before_build ran via a plain shell(before_build_prepared, env=env) — no prefix.

The wheel itself still cross-compiles via ARCHFLAGS, but hooks that probe the architecture or invoke arch-specific tooling (platform.machine(), uname -m, compilers, etc.) behaved differently in before_build vs before_test. The NumPy maintainers hit this (#2592).

Fix

• Extract the prefix logic into a small shared helper rosetta_arch_prefix(machine_arch, target_arch) so the build and test paths stay in sync.
• Wrap before_build with arch -x86_64 when the host is arm64 and the build target is x86_64-only, mirroring the test loop.
• Refactor the test loop to use the same helper (no behavior change there — it returns [] for native and the previously-handled cases).

The native paths (arm64-on-arm64, x86_64-on-x86_64) are unchanged: the helper returns an empty prefix, so before_build still goes through plain shell().

Design decisions

• universal2: A universal2 config compiles both arches in a single invocation via ARCHFLAGS="-arch arm64 -arch x86_64". There is no single arch prefix that is correct for that hook, so I left universal2 unwrapped (runs natively as arm64). Forcing arch -x86_64 would be wrong for the arm64 slice. This matches how the build itself is invoked (one process, both arches).
• before_all: Left unwrapped for now. It runs once before all builds, keyed off python_configurations[0], which does not necessarily represent the arch of every build in a mixed selection. Wrapping it based on the first config could be misleading. Tradeoff noted below as an open question.

Open questions for maintainers

1. Confirm the desired semantics: wrap before_build for x86_64-only configs on arm64 hosts, but not universal2. Is leaving universal2 native correct, or would you prefer wrapping it as the native arch explicitly (a no-op today) for clarity?
2. Should before_all also be wrapped? If so, based on what — the first config's arch, or only when all selected configs target x86_64? Or leave it native (current behavior)?
3. The fix is verified by a unit test on the pure helper. Full end-to-end behavior needs a real arm64-host / x86_64-target (Rosetta) environment, which isn't available in unit tests — see limitation below.

Tests

Added unit_test/macos_rosetta_test.py covering rosetta_arch_prefix for the arm64->x86_64 (Rosetta), native, and unsupported-direction cases. The wrapping wiring inside build() itself isn't unit-testable without a real cross-arch build (it requires real Python.org installs and Rosetta); this is noted as a limitation.

uv run pytest unit_test -> 795 passed. Ruff + mypy (strict) clean on the changed file.

Closes #2592

[joerick]

I'm not sure the conclusion of that issue thread #2592 was to emulate the before-build calls. There is indeed an inconsistency between build and test phases (cross-compile vs emulation) but isn't it worse to run before-build under emulation and then the build under the native arch?

[henryiii]

Okay, I wasn't sure.