Ferralk / Archive
Deferred follow-up
Ferralk 1.1.0 is published. There are no external release gates and no noisy wall-time thresholds. Deterministic allocation-count invariants run with the ordinary test suite; GitHub tracks the remaining follow-up work.
The native filesystem parity corpus that used to sit in this table closed with #12: eleven parity families now compare the native and portable backends over whole trees, on both platforms and under the sanitizers.
Platform state
Scroll sideways for more columns.
| Platform | Backend | Exercised by |
|---|---|---|
| macOS | getdirentries64, behind native-macos | macOS native and AddressSanitizer jobs, every pull request |
| Linux | getdents64, behind native-linux | Linux native, AddressSanitizer and Miri jobs, every pull request |
| Windows | Portable reader only | Windows test job, every pull request |
Both native backends degrade per entry rather than per directory, and fall back to the portable reader when a filesystem cannot answer them; the fallback paths carry their own tests. Windows is deliberately portable-only, including the follow-symlinks cycle key, which needs a file index Rust does not expose on stable there — ADR-0005 and ADR-0010 carry the reasoning.
The native validation workflows run automatically for pull requests. Their purpose is safety and behavioral parity, not a performance threshold.
Benchmark evidence lives in benchmark evidence: what each lane measures, how to reproduce it, and how Ferralk compares with the Rust baselines and with zlob. None of it gates a release.