docs(audits): cross-platform fast-I/O gap benchmark plan (#1386)

[oferchen]

Summary

• Plan-only audit at docs/audits/cross-platform-fast-io-gap-bench.md for task Streamline CI and expose brand profiles #1386.
• Catalogs the three fast paths under test: Linux io_uring writes (fast_io::io_uring), macOS clonefile/fcopyfile (fast_io::platform_copy), and Windows CopyFileExW + IOCP overlapped writes (fast_io::iocp, fast_io::copy_file_ex).
• Defines five workloads (1 MB / 100 MB / 1 GB single-file copy, 10 K small-file batch, modify-in-place delta apply), per-platform backend matrix, and metrics (wall time via hyperfine, syscall counts via strace/dtruss/ETW, peak RSS, bytes-on-disk).
• Predicts per-workload outcomes: APFS reflink dominates W1-W3, io_uring leads W4 and W5, IOCP closes the receiver-side gap on Windows; macOS small-file and delta-apply paths trail until the kqueue backend (Adjust CI coverage gate and extend cross-platform builds #1385) lands.

Test plan

• Land the harness changes called out in section 2.5 (extend crates/fast_io/benches/platform_copy.rs and add a delta_apply group).
• Wire IoUringPolicy / IocpPolicy overrides into the bench driver so each link in the fallback chain is exercised end-to-end.
• Run W1-W5 on Linux (Arch bench container), macOS (Mac mini M2 baseline), and Windows (CI runner; both NTFS and ReFS).
• Record results in docs/benchmarks/ and link back from this audit when measurements are taken.