rust-perf

Rust performance

Load this before optimizing. Do not sacrifice invariants (determinism, additive-only, no mandatory LLM) for speed.

Pre-registered targets (§7.3) — not measured results

Treat these as the optimization budget. Write them into benches as assert ceilings once harness exists. Until then, design so they are reachable. Never publish a number as measured unless sd-bench/RESULTS.md was produced by the harness on a tagged release.

Parity with Cisco/NVIDIA on synthetic TPR is the detection claim; adoption is this table. Competitors cannot move install footprint without leaving Python.

What to optimize (in order)

1. Do not compile YARA at scan time. build.rs once. Scan = match.
2. mmap skill files with memmap2 when size > ~64 KiB. Keep small files on the stack/heap.
3. Zero extra alloc on the miss path. A clean skill should not build a Vec<Finding> of zeros then throw it away — return empty slice / SmallVec.
4. Rayon over files and engines, not over lines inside one 2 KB SKILL.md.
5. Sorted walk once. Determinism and perf: ignore/walkdir → collect → sort → process.
6. Canonical digest over sorted (path, content) pairs; hash streaming, not a giant concatenated String.
7. No tokio, no async, no reqwest on the default path. Async startup cost kills cold-install.
8. Binary size: default features slim. LTO=thin in release profile. strip = true. Avoid tokio+reqwest+openssl in the default binary (use rustls only if L4 is on). Target ~12 MB stripped musl.
9. RSS: do not hold every file's bytes after L1 on that file finishes. Drop mmaps per file. Global caches are digest → findings, bounded.
10. Ruleset is embedded bytes, not a temp directory of .yar files at runtime.