[harbinresearcher]

Closes #240.

When the metadata fallback builds a wheel in the directory later passed to build_wheel() or build_editable(), cached-wheel reuse copies the wheel onto itself. The subprocess fails even though the requested wheel is already present.

Skip the cached-wheel copy when the existing destination is the same file. Separate output directories retain the existing copy behavior. No backend hook or public API changes.

Four regression cases exercise wheel/editable fallbacks with same/separate directories. A trailing marker verifies the cached wheel is preserved rather than rebuilt. On unmodified main, the two same-directory cases fail and the separate-directory cases pass.

Validation:

• Linux Python 3.12: baseline 42 passed; patched 46 passed.
• Windows Python 3.8 and 3.12: baseline 41 passed / 1 failed; patched 45 passed / the same 1 failed. The pre-existing failure is test_path_pollution_symlinked_in_process because this local process lacks symlink privileges (WinError 1314); the new regressions pass on both versions.
• Pinned pre-commit hooks: all passed, including Ruff, Black and mypy.
• Strict Sphinx documentation build: passed.
• Source distribution and wheel built from that source distribution: passed.

Prepared with AI assistance, including tests and local validation. No maintainer approval or production usage is claimed.

[agammann]

I checked the two fallback build paths with the metadata and wheel directories both shared and separate. The new samefile guard skips an invalid self-copy only for the shared directory, while the normal copy still produces the cached wheel in a distinct output directory. All four focused cases pass, and this matches the reproduction in #240.