[zhangzhenghao]

Summary

• Keep scikit-build-core as the direct packaging backend, with CMake disabled and automatic py3-none platform wheel tags. No custom build backend or hook.
• Add one repository-only Jinja2 python/pyproject.toml.j2 template for CPU, CUDA, ROCm and Vulkan packages; publish only rendered metadata, never templates.
• Delete stage_runtime.py, _build_backend.py, prepare_backend.py; do not introduce hatch_build.py.
• GitHub Actions directly copies externally compiled runtime files and licenses, then validates them before wheel assembly.
• Exclude tests from both wheels and sdists. Release CI obtains tests separately from the exact release checkout, preserving the installed-wheel Python 3.9–3.14 matrix.
• Update archive/release validators and documentation; preserve the shared native build and existing Node/Rust jobs.

Verification

• Independent review: no blocking findings.
• Installed real CPU wheel + packaging/release tests: 109 passed, 77 subtests passed, including real inference, ownership, concurrency and typing.
• Actual release workflow rendering/source-build blocks exercised: CPU/CUDA/ROCm/Vulkan sdists built and passed strict Twine checks; inspected wheels and sdists contain no tests.
• Compiler/CMake/Ninja PATH sentinels confirmed compilation-free packaging on Linux.
• Native payload bytes match the external runtime; NumPy absent.
• actionlint and git diff --check passed.

Limits

• GPU namespace packaging checks use CPU runtime bytes and are not GPU execution evidence.
• No local macOS/Windows runner execution, GPU compilation/inference, manylinux repair, or registry publishing.
• Removing the custom backend intentionally moves missing/invalid-runtime validation to Actions; standalone direct-backend builds must prepare the documented runtime and license payload themselves.

Template exclusion follow-up

• Both wheels and sdists exclude all .j2 files; GPU rendering and installed-wheel tests receive the template from the exact checkout, not published archives.
• Inspected four wheels and four sdists: no templates or tests. Actual workflow sandbox verified; 91 installed-wheel tests, 19 release tests and 7 packaging tests passed. Independent review found no blockers.

Remote CI

• CI run 42280272477 passed on latest head ffa1e9c5514f094cf16617d137b27a965923e579: all 10 jobs succeeded, including Linux x64/ARM64, macOS ARM64 and Windows x64 Python assembly/tests; Windows ARM64 retains C++-only coverage. Release-only GPU/repair/publication jobs remain unexecuted.