fix: preserve borrowing for Windows drive-relative paths

[hyfdev]

Summary

• add an independent bounded oracle for public normalize and into_normalized semantics
• cover 6,560 Unix and 14,040 Windows cases across roots, drive-relative paths, components, separator spellings, trailing separators, Unicode, and case preservation
• verify exact results, Cow ownership, borrow source, and consuming results
• preserve borrowing and owned-buffer reuse for normalized Windows drive-relative leading-parent paths such as C:..\..

Why

The existing classifier test compares the production classifier with the production normalizer. It protects output equality, but cannot independently validate public normalization semantics or detect an unnecessary allocation when both paths produce the same value.

The new model exposed a real classifier gap: the generic Windows scan treated the second \.. in C:..\.. as requiring normalization. Drive-relative leading parents cannot be collapsed, so the normalizer rebuilt an identical path and returned an owned result.

The classifier now recognizes canonical drive-relative leading-parent tails while still rejecting duplicate separators, current-directory components, trailing separators in strip mode, and parents that follow a normal component.

Impact

• public output semantics are unchanged
• canonical Windows drive-relative leading-parent paths now return Cow::Borrowed
• into_normalized reuses the original owned buffer for those paths
• paths that actually require normalization still return an owned result

Validation

• default and cached_current_dir workspace tests
• full Clippy with warnings denied
• documentation and package verification
• Windows MSVC library and integration-test metadata compilation
• native Linux, Windows, macOS, and WebAssembly CI
• Linux and Windows allocation snapshots
• CodSpeed: 228 unchanged benchmarks
• final adversarial review: pass

[codspeed-hq]

Merging this PR will not alter performance

✅ 228 untouched benchmarks
⏩ 38 skipped benchmarks¹

---

_{Comparing agent/public-normalize-model (74949da) with main (2c2b185)}

Footnotes

1. 38 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.