Changelog

All notable changes to this project are documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning. Versions apply to the whole repository and are git tags (see docs/release.md). Kits are not released separately and carry no version of their own.

[Unreleased]Copy Anchor Link

• Docs guides (en + zh): plain-docs pass on performance guides — drop How-to-read / Scope-and-method coaching lists where the lead already says measure-first; replace heuristic-dogfood scope wording with lint-hit vs wall-time framing; prefer Lint: id (default …) lines; soften coaching bold; sync zh-CN guide-src (and zperf hand zh-CN/zig-performance-guide.md) to the English structure and voice.

• Docs zh-CN: site Markdown Chinese coverage for shared docs, kit guides (guide-src/zh-CN + build-guides stem pairing), package landings, corpus/evidence-policy/rule-ids, active evidence notes, and root pages published as *.zh.md via zh-CN/ parents. Tombstone evidence under evidence/_removed/ intentionally left English-only.

• Docs SEO / GEO (en + zh): description, Open Graph (og:locale + alternateLocale), Twitter Card, canonical and hreflang; static export forces blocking metadata into HTML (htmlLimitedBots); sitemap.xml pairs en/zh with language alternates; robots.txt allows common AI crawlers; llms.txt + llms.zh.txt; locale-aware JSON-LD (WebSite / WebPage / SoftwareApplication); post-build sets <html lang="zh-CN"> under /zh/ for crawlers; 1200×630 og.png. README opens with an English-only value-prop hook.

• Docs site: page links no longer keep a trailing .md (that 404s on GitHub Pages). Assemble rewrites them to trailing-slash directory URLs relative to the page directory, fails the build if any .md hrefs remain, emits static files as root-absolute paths so the /zh/ locale works, and publishes LICENSE / NOTICE as directories so trailingSlash can open them.

• All kits: default matching is --scope all where the kit CLI supports it (opt in to --scope loop for the former in-loop-only match). Finding messages no longer imply non-loop sites are out of scope. Rule slugs containing in-loop were renamed to *-hot (including leftover rs-map-new-in-loop → rs-map-new-hot; see per-kit Renamed tables); old slugs are tombstoned under each kit evidence/_removed/*.REMOVED.md and must never be reused. See each kit docs/rule-ids.md Renamed table. jsperf also keeps adopter FP cuts + new hints from earlier commits on this branch. No severity escalations.

• jsperf-lint: default --scope all reports former in-loop smells everywhere (extreme-perf default); opt in to --scope loop for the old in-loop-only match. Finding messages and docs no longer imply non-loop sites are out of scope. Still includes loop-header / nested-function FP cuts; path skips for e2e/, playwright-report/, and *.test.* / *.spec.*; new hints jp-query-all-star, jp-execjs-in-timer, and jp-clone-per-event (stay hint); scanner reads JS inside executeJavaScript / evaluate templates. No severity escalations; jp-string-plus-in-loop stays tombstoned.

• Documentation site uses a Fumadocs layout: sidebar, command search, and Geist. The pages are assembled with Node.

• GitHub Pages site for the guides and kit docs, built from the Markdown already in the tree.

• Removed the checked-in C# bench executable and its .NET build outputs.

• Restored rs-map-new-in-loop as hint. Two quiet rustc 1.99 runs of 20000 one-key inserts were about 1.53× and 1.51× versus clearing one map. The same two runs left the large Vec fill at about 1.00× and the large String fill at about 1.03×, so those removals stay.

• Quiet reruns on current toolchains. Rust 1.99: Instant::now about 20× (the old 439× after-side was 0.008 ms), File::open about 416×, Regex::new on regex 1.13.1 about 963×. All stay hint. Nim 2.2.12 &= versus join was about 0.79×; nm-string-amp-in-loop is removed and the slug is not reused.

• Hint text for compile, parse, open, and clock rules now states the bench premise: hoist only when the pattern, text, path, or timestamp stays the same. A call whose argument is the loop binding is not reported. Date / Stopwatch / Instant::now inside a loop that awaits is not reported in C#, Swift, and Rust. ArkTS no longer reports JSON.parse on a for / forEach / map head. Elixir Enum.map/filter says the cost grows with length.

• JavaScript jp-string-plus-in-loop is removed. push then join was not about 2× versus += on a short literal, String.fromCharCode, or a 64-character literal (at 22000 characters join was slower). jp-map-filter-chain now has JSONL and stays hint: 35000 rows about 4.6× and 9.8×, a 3-row filter+map 1.28×. Hint lines state the bench premise. readFileSync in a for iterable, Date.now in a loop that awaits, and new RegExp / JSON.parse / a DOM selector / Intl.DateTimeFormat whose argument is the loop binding are not reported.

• Measured the previously unbenched C and Swift shapes on this machine. strlen in a condition was about 125× a hoisted length when the string might change, and about 0.93× when -O2 already hoists it (cp-strlen-in-cond, still hint). strcat of fixed chunks was about 221× an end pointer (cp-strcat-hot, stays off). Swift reserveCapacity was about 5× on many small arrays and about 1.6× on one large fill (sp-array-no-reserve, still hint). print to /dev/null was about 10× a String (sp-print-hot, stays off). Removed sp-string-plus-in-loop (reserve was slower), sp-map-joined-chain (about 1.2×), and sp-force-unwrap-in-loop (same time as if let).

• Removed measured-near-1× matchers: at-object-delete (~1.01× on a Node proxy), cx-string-append-unreserved (~1.09× time, capacity under 1.4×), rs-string-plus-in-loop (~1.00× on a large fill), rs-vec-no-capacity (~1.00× time and ~1.05× bytes on a large fill), and zp-extern-struct-padding (~1.21× walk, 24 bytes versus 16). Slugs are tombstoned and not reused. format! rebind, value-class allocation, and alloca stack growth stay.

• Zig hint: allocPrint / Allocator.print of an unchanged format string and arguments was about 50× on the general-purpose allocator, about 12× on an arena, and about 14× for a constant string (zp-alloc-print-hot). bufPrint, Writer.print, byte copies, and a 256-byte struct copy stayed at or under about 7× and were not added. Empty-buffer writeByte was about 130×, but the same call on a 4KiB buffer was about 1.2×, so it was not added. Default stays hint.

• C++ hints: const std::string / const std::vector by value was about 15–21× versus a const reference when the copy allocates (cx-const-by-value). A 4-character string was about 1× and is still hinted. Walking vector<bool> was about 12× a vector<unsigned char> (cx-vector-bool); the packed container is smaller. Hoisting std::function was slower (about 0.6×). Copying shared_ptr stayed under 7×. map::operator[] on missing keys was about 30× but the same spelling on existing keys was slower than find, so it was not added. Defaults stay hint.

• C hints: alloca in a loop covers the stack (20000 × 64 bytes versus one 64-byte buffer) while the touch time stays about 1× (cp-alloca-hot). A 512-byte struct passed by value across two translation units was about 12.3× versus a pointer (cp-struct-by-value). The same shape in one translation unit was about 1×. fgetc / fputc / putc were faster as a block read (about 24–120×) but were not added: a line matcher cannot tell a byte parser from a bulk read. Defaults stay hint.

• Cross-language check of file opens and clock samples. New hints, each with JSONL, where the same shape was costly and the standard library can reuse a handle or a sample: Python open (~409×), JavaScript readFileSync (~480×), Date.now (~26×), and performance.now (~19×), Go os.Open (~2300×) and time.Now (~104×), Ruby File.read (~346×), C# File.ReadAllText (~1560×) and Stopwatch.GetTimestamp (~15×), Rust File::open (~2750×) and Instant::now (~439×), Swift contentsOfFile (~414×), Objective-C stringWithContentsOfFile (~450×). C# new Regex( now matches cs-regex-hot (~7×, both sides still match). Python perf_counter_ns (~2.9×), Ruby Process.clock_gettime (~2.3×), getattr (~1.4×), send (~1.9×), and JavaScript Reflect.get (~8×) were checked and not added. Defaults stay hint.

• C++ ifstream of one one-byte file was about 2025× versus opening once (cx-ifstream-hot, hint). Zig 0.17 openFileAbsolute was about 2134× and Clock.awake.now about 42× (zp-open-file-hot, zp-clock-hot, hint). C# GetField plus GetValue each iteration was about 62× versus a direct field (cs-field-by-name-hot, hint). Zig json.parseFromSlice of one small object was about 2683× versus parsing once (zp-json-hot, hint). Objective-C valueForKey: was about 124× and Swift value(forKey:) about 35× versus a direct field (oc-value-for-key-hot, sp-value-for-key-hot, hint). Fair rerun of Swift Date() (~244×) and Objective-C [NSDate date] (~769×), with the after loop kept alive. Both are hints with JSONL. A discarded first run had deleted that loop.

• Fair reruns: removed cx-sstream-in-loop, pp-regex-compile-in-loop, rs-map-new-in-loop, pp-datetime-format-in-loop, gp-time-parse-in-loop, and rb-date-parse-in-loop (near 1×, or the only win was skipping the call). cx-endl-hot is about 23× on an external stream and about 0.9× on a string stream. gp-reflect-hot now matches name lookup and reflect.New; TypeOf was about 1.3× and is not matched. cp-regex-hot, cx-clock-hot, cs-json-hot, rs-regex-new-hot, jp-dom-query-hot, and jp-intl-datetimeformat-hot have JSONL and stay hint. A JSON or date value that changes every iteration cannot be hoisted. Java SimpleDateFormat was not added: no JDK in this environment.

• Hint rules for repeated compile, parse, query, format, and fresh-map work in Python, JavaScript, Go, C++, C#, Ruby, and Rust. Paired fixtures only. No JSONL, so none of them is error. The 2026-10-22 corpus verdict does not cover these slugs.

• C hints for sprintf / fopen / clock reads in a loop, and for four or more full-width boolean flags in one struct (cp-wide-bool-flags). Paired fixtures only. No JSONL. Not corpus-triaged.

• Swift hints for constructing a date formatter, a JSON decoder, or NSRegularExpression in a loop. Both the Swift scanner and the text fallback carry them. Paired fixtures only. No JSONL. Not corpus-triaged.

• Objective-C hints for allocating a date formatter, calling NSJSONSerialization, or building NSRegularExpression in a loop. Paired fixtures only. No JSONL. Not corpus-triaged.

• Microbench JSONL for the new hint rules that could be timed on one shape (compile/parse/format/copy). Defaults stay hint. cx-endl-hot and gp-reflect-hot were about 1×. jp-dom-query-hot and rs-regex-new-hot were not timed.

[0.1.0] — 2026-10-22Copy Anchor Link

Umbrella first public-shaped snapshot (Asia/Shanghai).

• 19 language kits under packages/*, corpus-ready under AGENT.md G3.
• Rule IDs are semantic slugs only ({family}-{smell}); numeric codes removed (no aliases).
• Measured-useless rules deleted (tombstones), not left as zombie off.
• Root README slimmed; kit status → docs/kits.md; run commands → docs/quickstart.md.
• ~124 active slugs at tag time; see each kit docs/rule-ids.md.