test: cover owned normalization contracts

[hyfdev]

Summary

• Add fixed, independent PathBuf::into_normalized input/output matrices for Unix and Windows instead of deriving expected values through the borrowed normalizer.
• Check exact native spelling and exact idempotence across clean and dirty relative paths, roots, unresolved parents, root clamping, current-directory collapse, trailing separators, non-ASCII components, and a dirty path beyond the inline component capacity.
• Cover Windows root-relative and drive-relative paths plus ordinary disk, UNC, verbatim disk, verbatim UNC, device namespace, and generic verbatim prefixes, including prefix-only output, collapse with and without a trailing separator, literal / inside verbatim components, and prefix-looking normal components.
• Preserve arbitrary native encoding through consuming normalization with raw Unix-byte assertions and Windows-wide assertions for ordinary components and all six encoded root-identifier slots.

Why

The existing dirty owned tests call normalize() to construct the expected value, so the borrowed and consuming APIs can agree on the same defect. Direct owned coverage otherwise protects only a small set of clean, current-directory, dirty, and invalid-component examples. This matrix fixes the expected native representation independently and exercises each behavior partition through the consuming wrapper.

Impact

This adds one integration-test file. It does not change production code, public APIs, dependencies, benchmarks, allocation snapshots, or documentation. The non-UTF-8 cases assert exact bytes or wide units and idempotence only; they deliberately do not turn pointer reuse, capacity, allocation count, or latency into a semantic merge gate.

Validation

• cargo +1.97.0 fmt --all --check
• cargo +1.97.0 test --locked -p sugar_path --test owned_normalization
• cargo +1.97.0 test --locked -p sugar_path --all-features --test owned_normalization
• cargo +1.97.0 test --locked -p sugar_path
• cargo +1.97.0 test --locked --workspace --all-features
• cargo +1.97.0 clippy --locked --workspace --all-targets --all-features -- -D warnings
• RUSTDOCFLAGS='-D warnings' cargo +1.97.0 doc --locked --workspace --all-features --no-deps
• The same test source passes cargo check and Clippy for x86_64-pc-windows-msvc in a dependency-only harness without the host-incompatible allocation-tracker C dependency.
• The dependency-only harness also passes Clippy with warnings denied for wasm32-wasip1-threads.

Adversarial review

• Review found that the shared helper remained compiled after every test was gated away on non-Unix/Windows targets, causing a dead_code error under WASI Clippy with warnings denied.
• A file-level Unix/Windows gate now removes the complete integration target elsewhere. The same reviewer verified the fix and reported PASS with no remaining blocker, P1, or P2 finding.

Hosted validation

• All 8 GitHub Actions jobs pass: lint, three native test hosts, two allocation snapshots, and both benchmark runners.
• The native Windows log shows both owned-normalization matrices executing successfully in both test steps. Ubuntu and macOS likewise execute both Unix matrices in both steps.
• The external CodSpeed check reports the exact previously observed opposing pair across different runtime environments: owned_receiver/string_result repeats 401.7 ns to 460 ns (-12.68%), while as_path/string repeats 255.8 ns to 226.7 ns (+12.87%). This test-only diff changes no production source, dependency, benchmark, or benchmark input, so that comparison is not attributable to this PR.

[codspeed-hq]

Merging this PR will not alter performance

✅ 228 untouched benchmarks
⏩ 38 skipped benchmarks¹

---

_{Comparing agent/owned-normalization-contract-matrix (531ece8) with main (f66d7c1)}

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.