sonicop 26.8.111-aarch64-linux → 26.8.112-aarch64-linux

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: ffcd9e8d5b49e4930702c201f4214fa4a3aca12d2d50c3307422488ea484fced
-  data.tar.gz: 0b25ee38f12c33914700b1f05e2628e494845ac2dc7b1c2538d2ae1ec128b4ba
+  metadata.gz: f1cab35205ef1b9545aa27aae10b033f94a6ef4285ad74f0f8439176570517ea
+  data.tar.gz: 8fa47739e28ec78bcd4ad740ffa2afdb4069a6fc673c9b11ef56f6a9ee6802fc
 SHA512:
-  metadata.gz: e1c2e822a493f07a7a887872dce6fbc5ee8b92aa5bb73a28d3c3b945a957da25c6baff4f3261fb4b7fa960a128faab4be59ad3a93d3668292c22364cc6ff04b5
-  data.tar.gz: 0a4aa06496d5a46b91ce0a4994d02ca969ef3c481fd9e1a6a5e136bb440f9c893dc012b97e561b4e6e7facf567ea50013f94bf8343badb16cd648429731043b0
+  metadata.gz: ee256fff60b527f0baf546b06a61ee57c0f12f3e4bdc25d109972a2376d8f723f66918de47b2c7c26305347c39be451c09422d37cb4d251e7abb4e9bce1ff7fa
+  data.tar.gz: 783be95e866e2ec1eeef816e1bd42500b01ff3d0b200073246d2c5e128fb80bd6722920c18d2591d2eb86eb0deb7928962baa9a3edc7d1c83f47352b59a9cde4

data/CONFORMANCE.md

@@ -55,17 +55,18 @@ sonicop --force-default-config --format json
 
 ## Results
 
-The two kinds of difference are counted separately because they mean different things. A
-`Lint/Syntax` difference says the two disagree about whether a file parses, and everything else in
-that file follows from it; a difference in any other cop says the port reads the same tree and
-draws a different conclusion. Only the second is a defect in a cop.
+Three kinds of difference are counted separately because they mean different things. The set of
+files with a fatal `Lint/Syntax` offense says whether the parsers disagree about accepting a file.
+`Lint/Syntax` positions inside a file both already rejected measure diagnostic recovery, not file
+acceptance. A difference in any other cop says the port reads the same tree and draws a different
+conclusion. Only the last category is a defect in a cop.
 
 | Corpus | Excess | Missing | of which `Lint/Syntax` | Other cops | Field differences | Measured |
 |---|---:|---:|---|---:|---|---|
 | rubocop/rubocop | 0 | 0 | — | 0 | correctable ×1 | 2026-08-17 |
 | rails/rails | 0 | 0 | — | 0 | none | 2026-08-17 |
 | mastodon/mastodon | 0 | 0 | — | 0 | none | 2026-08-17 |
-| Homebrew/brew | 263 | 997 | **all of them** | **0** | none | 2026-08-17 |
+| Homebrew/brew | 263 | 997 | **all of them** | **0** | none | 2026-08-18 |
 | ruby/ruby | 142 | 585 | 117 missing, 44 excess | 92 | 5 | 2026-08-16 |
 
 The last column is per row on purpose. A single date at the top of the file would say the five were
@@ -85,14 +86,19 @@ run starts from, and `TargetRubyVersion` is only inferred from the working direc
 tools therefore fall back to the default of 2.7 (RuboCop says so itself: `Using Ruby 2.7 parser`) and
 both call `dry_run:,` — a hash value omission, valid since 3.1 — a syntax error. **Run the same corpus
 at 3.1 and both report zero `Lint/Syntax`.** What is left is not disagreement about which files parse:
-the file sets are identical, 569 on each side, and the first diagnostic in each file lands at the same
-position. Only the second and later diagnostics diverge, because the two parsers recover from the
-error differently — which is what puts offenses on both sides of the ledger rather than only on
-RuboCop's. Those are not chased; see *Known divergences*.
-
-What is checked here is that every one of the 1,260 is a `Lint/Syntax` offense, in both directions.
-What is not checked is which recovery produced which diagnostic: the 263 Sonicop reports and RuboCop
-does not have not been read one by one. They are accounted for as a class, not individually.
+the file sets are identical, 569 on each side, with no file unique to either parser. The two parsers
+recover differently after encountering invalid syntax, which puts diagnostic positions on both sides
+of the ledger rather than only on RuboCop's.
+
+All 263 Sonicop-only positions were classified mechanically, not inferred from their cop name. They
+occur in 135 of those shared syntax-error files, and every one follows a `Lint/Syntax` position that
+both tools reported in the same file. Their messages are 113 end-of-input (`$end`), 86 right
+parentheses (`tRPAREN`), 53 commas (`tCOMMA`), seven `end` keywords (`kEND`), two equals signs
+(`tEQL`), one `when` (`kWHEN`) and one identifier (`tIDENTIFIER`). Thus the former “excess” is not a
+set of valid files rejected only by Sonicop; it is the other direction of the same diagnostic-recovery
+divergence as the 997 RuboCop-only positions. Suppressing every diagnostic after Sonicop's first one
+would make the excess zero, but would also discard 1,998 positions that already match and increase the
+missing side from 997 to 2,995. That is not a conformance fix. See *Known divergences*.
 
 Much of ruby/ruby's difference is the same shape,
 with 117 of the 585 missing and 44 of the 142 excess being `Lint/Syntax` itself,
@@ -181,10 +187,10 @@ but no longer reproduces fails the test, and so does one that appears without be
 ### Error recovery after a syntax error
 
 RuboCop parses with `parser`, an LALR parser that recovers from an error and keeps going, emitting
-further diagnostics from the recovered state. Sonicop parses with tree-sitter, which does not model
-that recovery, so the follow-on diagnostics cannot be reproduced. On Homebrew this accounts for the
-997 missing `Lint/Syntax` offenses: `class definition in method body`, `dynamic constant assignment`,
-`cannot assign to a keyword`, and repeated `unexpected token` inside one multi-line hash.
+further diagnostics from the recovered state. Sonicop parses with tree-sitter and reaches different
+recovery states. On Homebrew this produces 997 RuboCop-only and 263 Sonicop-only `Lint/Syntax`
+positions: examples on the RuboCop side include `class definition in method body`, `dynamic constant
+assignment`, `cannot assign to a keyword`, and repeated `unexpected token` inside one multi-line hash.
 
 What the divergence does not change is **which** files are held to be unparseable, and that is what
 decides whether a file is inspected at all: RuboCop runs no cop other than `Lint/Syntax` on a file
@@ -198,9 +204,10 @@ programs, one of which is English prose that happens to parse as Ruby — where
 error and `parser` does not. Those five hold 465 of that corpus's 773 differences.
 
 Within Homebrew's 569 files the agreement is partial, as recovery cannot be reproduced: 499 report the
-same first diagnostic (position and message), and 222 report an identical list end to end. The 277
-files in between are the divergence in its purest form — the two agree on where the file first goes
-wrong and part company on what follows. The 70
+same first diagnostic (position and message), and 222 report an identical list end to end. In all 569,
+Sonicop's earliest source-position diagnostic is also present in RuboCop's output. The 277 files in
+between are the divergence in its purest form — the two agree on at least one error position and part
+company as recovery proceeds. The 70
 files whose first diagnostic differs follow one shape — an endless method definition (`def to_s =
 to_str`, valid from Ruby 3.0, rejected by the default `TargetRubyVersion: 2.7`) leaves `parser`'s
 method context open, so the enclosing `class` emits `class definition in method body` when it is
@@ -299,17 +306,43 @@ defect, so the difference stands as it is.
 ## Limits
 
 A clean run is a property of the corpora, not a general claim. `known_divergences.yml` carries the
-current list of what these corpora never exercise; the main ones are non-default configuration
-values, extension plugins, and Windows line endings. Cops that never fire contribute nothing to a
-match count, and their silence is indistinguishable from agreement.
-
-That limit is not hypothetical, and one instance is now measured. `Style/HashSyntax` defaults
-`EnforcedShorthandSyntax` to `either`, under which neither tool reports anything; the port
-implements none of the other four values, so half the cop is missing. Both the corpus runs and a
-sweep of RuboCop's own specs report agreement for it, because both run at the default. A cop can be
-half absent and still match everywhere the default reaches. Where a cop's behaviour is selected by
-configuration, this document covers the default branch only — the others are neither measured nor
-claimed.
+current list of what these corpora never exercise; the main ones are extension plugins and Windows
+line endings. Cops that never fire contribute nothing to a match count, and their silence is
+indistinguishable from agreement.
+
+Non-default configuration values used to be the largest of those gaps, and are now measured. Every
+one of the 111 cops carrying an `Enforced*` setting was switched to a non-default value at once —
+`Layout/SpaceInsideParens` to `space`, `Style/HashSyntax` to `hash_rockets` with
+`EnforcedShorthandSyntax: always`, and so on — and rubocop/rubocop re-run against RuboCop under the
+same file. Of 622,317 reference offenses, **99.99% match**, and 84 of the 96 cops that fired match
+exactly.
+
+| Cop | Offenses differing | |
+|---|---:|---|
+| `Layout/HashAlignment` | 35 | **RuboCop's own bug**, see below |
+| `Style/BlockDelimiters` | 17 | |
+| `Style/ConditionalAssignment` | 12 | |
+| `Style/MixinGrouping` | 8 | |
+| `Layout/SpaceInsideStringInterpolation` | 6 | |
+| `Style/NumericPredicate` | 5 | |
+| `Style/PercentQLiterals` | 4 | |
+| `Naming/VariableName` | 3 | |
+| `Lint/SymbolConversion` | 3 | |
+| `Layout/SpaceInsideParens` | 1 | |
+
+`Layout/HashAlignment`'s 35 are not a defect here. Under `EnforcedHashRocketStyle: separator`
+RuboCop raises `Parser::ClobberingError` on `spec/rubocop/config_loader_spec.rb` — three times, at
+`1111:35`, `1130:38` and `1217:35` — and reports nothing for the hashes it crashed on. Sonicop
+inspects them and reports. Every one of the 35 sits in that file between lines 1114 and 1234.
+Reproduce with:
+
+```bash
+rubocop -c <variants.yml> --only Layout/HashAlignment -d spec/rubocop/config_loader_spec.rb
+```
+
+One value per cop is not the whole configuration space — a cop with four supported styles is
+measured at two of them — so this is a floor, not a ceiling. It does settle the shape of the old
+limit: a cop is no longer allowed to be half absent and still count as matching.
 
 The 215 cops RuboCop ships switched off are a limit of a different kind. They are implemented, but a
 default run never reaches them, so the corpus numbers above say nothing about them. What stands

data/NOTICE

@@ -15,9 +15,9 @@ licence position checkable rather than remembered.
   not the attribution. An entry that is added late is an entry that was, for a
   while, undocumented in a public repository.
 
-  Known to be coming: tests/fixtures/upstream_spec_cases.jsonl, which will hold
-  input text from RuboCop's specs. It is deliberately **not** listed yet -- this
-  file describes what is here now, not what is planned.
+  That inventory is now complete for what is here. The entry that used to be
+  listed as "coming" -- the spec-derived fixtures -- has landed and is recorded
+  below with the rest.
 
 ================================================================================
 WHAT IS TAKEN
@@ -32,6 +32,18 @@ RuboCop -- https://github.com/rubocop/rubocop -- MIT
       Regenerate: scripts/sync_default_yml.sh 1.89.0
       Full text also kept at licenses/RUBOCOP.txt
 
+  tests/fixtures/upstream_spec_cases.jsonl
+  tests/fixtures/upstream_spec_divergences.jsonl
+      Input text taken out of RuboCop's own cop specs, each paired with what
+      RuboCop 1.89.0 actually reports for it. The spec text is upstream's; the
+      expectation beside it is not -- upstream was run once and its output
+      recorded, because upstream does not always behave the way its own specs
+      say. The second file holds the cases whose `-A` output does not parse as
+      Ruby, kept apart so that nothing demands this port reproduce them.
+      Regenerate: make spec-fixtures  (runs upstream once per cop; needs the
+      rubocop gem, which the test run itself does not)
+      Read by tests/spec_fixtures.rs
+
 Ruby on Rails -- https://github.com/rails/rails -- MIT
   Copyright (c) David Heinemeier Hansson
 

data/README.ja.md

@@ -32,22 +32,80 @@ CI へ最小限の変更で導入できます。
 
 ## 主な機能
 
-Bundler、Gemspec、Layout、Lint、Metrics、Migration、Naming、Security、Style の各部門の Cop を実装しています。
-一覧はバイナリ自身が正本です。
+Bundler、Gemspec、Layout、Lint、Metrics、Migration、Naming、Security、Style の各デパートメントの
+Cop を実装しています。一覧はバイナリ自身が正本です。
 
 ```bash
 # 認識済み Cop と実装状況の一覧
 sonicop --show-cops
 ```
 
-**RuboCop 1.89 の全 609 Cop を実装しています。** 本家のレジストリと名前まで一致しており、
-ただしこれは**レジストリについての主張**であって、すべての設定についての主張ではありません。
-`Style/HashSyntax` は存在し既定の届く範囲では一致しますが、`EnforcedShorthandSyntax` の
-既定以外の 4 つの値はいずれも未実装です。**Cop は半分欠けたまま、既定の届く範囲では
-一致し続けられます** — [CONFORMANCE.md](CONFORMANCE.md) の *Limits* を参照してください。
+**RuboCop 1.89 の全 609 Cop を実装しています。** 本家のレジストリと名前まで一致しています。
 `Enabled: pending` の 159 個と `Enabled: false` の 56 個も含みます。この 215 個は本家でも
-既定の実行では走らないので、`--only` で名指しするか設定で有効にしてください。本家にも
-存在しない Cop 名だけをエラーにし、必要なら `--ignore-unrecognized-cops` で続行できます。
+既定の実行では走らないので、`--only` で名指しするか設定で有効にしてください。本家に存在しない
+Cop 名を書いた場合だけエラーで止まります（`--ignore-unrecognized-cops` で続行できます）。
+
+### Cop 別の一致状況
+
+全 609 Cop を両者で有効にし、本家の spec が供給する 37,491 ケースを、それぞれの spec が
+指定した `TargetRubyVersion` で走らせて比較した結果です。**完全一致**とは、その Cop の offense が
+位置・メッセージ・重大度・修正可否まで 1 件残らず一致し、どちらにも余りがないことを指します。
+
+<!-- conformance:start -->
+| デパートメント | Cop 数 | 検証済み | 完全一致 | 相違 |
+|---|---:|---:|---:|---:|
+| Bundler | 7 | 7 | **7 ✓** | 0 |
+| Gemspec | 10 | 10 | **10 ✓** | 0 |
+| Layout | 100 | 100 | 90 | 10 |
+| Lint | 157 | 157 | 147 | 10 |
+| Metrics | 10 | 10 | **10 ✓** | 0 |
+| Migration | 1 | 1 | **1 ✓** | 0 |
+| Naming | 19 | 19 | **19 ✓** | 0 |
+| Security | 7 | 7 | **7 ✓** | 0 |
+| Style | 298 | 298 | 288 | 10 |
+| **合計** | **609** | **609** | **579** | **30** |
+<!-- conformance:end -->
+
+**先に読むべきは「検証済み」の列です。** ここで一度も発火しなかった Cop は、沈黙が一致と
+見分けられないため、訊かれないまま一致に数えられてしまいます。**609 個すべてがここで発火します**。
+それが「完全一致」の列を意味あるものにしています。
+
+そのうち 3 個は専用の run を要しました。`Lint/DeprecatedReference`・`Lint/NameTypo`・
+`Lint/UnusedPrivateMethod` は `rubydex` のプロジェクトインデックスが無ければ何も報告せず、
+gem の導入と `AllCops/UseProjectIndex` の有効化が要ります。`Lint/DeprecatedReference` は
+さらに条件があり、`@deprecated` を持つメソッドの定義クラスを**継承したクラスの中**から
+呼ぶ必要があります。本家の spec は「インデックス無しでは offense を出さない」という
+`expect_no_offenses` で始まるので「到達不能」と読みたくなりますが、そうではありません。
+
+「完全一致」を 609 にし、「相違」をゼロにすることが現在の目標です。実コードを足しても
+届きません。RuboCop が既定で無効にしている 56 Cop と、pending として出荷している Cop の多くは、
+ツリーがどれだけ大きくても素の実行では発火しないためです。そこに届くのは本家の spec が
+供給する入力の方で、`tests/fixtures/upstream_spec_capture.jsonl` に記録したケースは
+**609 Cop すべてに到達します**（実測）。走らせ方には間違えやすい点が 2 つあり、
+**どちらも失敗せずに表を縮めます**。
+
+- **`TargetRubyVersion` は入力の一部であって、全体の設定ではありません。** 2.7 に固定すると
+  `Style/ArrayIntersect`・`Naming/BlockForwarding`・`Style/ItBlockParameter` ほか 11 個が
+  そもそも発火できません。各ケースはその spec が指定した版で走らせます。
+- **ファイル名そのものを見る Cop があります。** `Bundler/*` は `Gemfile`、`Gemspec/*` は
+  `.gemspec`、`Naming/FileName` は名前自体を読みます。全ケースを `.rb` で書き出すと、
+  この 17 Cop が何にも一致しませんでした。
+
+これとは別に、2026-08-29 に本家 Cop spec から直接抽出できた 11,506 ケースを oracle で
+照合しました。本家が入力を読めなかった 226 ケースとクラッシュした 1 ケースを除き、測定できた
+範囲では Sonicop の**検出差分・訂正差分ともにゼロ**でした。この結果は上の表には加えていません。
+直接抽出できるケースが無い Cop が 51 個あり、ディレクティブ系 3 Cop は `--only` では測定不能な
+ため、この掃引だけで全 609 Cop の完全一致を証明できないからです。また、この直接掃引は各例の
+`TargetRubyVersion` をすべて再現せず、中立的な既定条件で実行しています。そのため、今回変更した
+Ruby 3.4 依存の 4 Cop は 3.4 で別途比較し、メッセージと位置が本家に完全一致することを確認しました。
+
+設定値については別に測っています。既定値でしか一致しない Cop は半分しか実装していないのと
+同じだからです。`Enforced*` 系の設定を持つ 111 Cop すべてを**既定以外の値**に倒して同じコーパスを
+流すと、**622,317 件の offense のうち 99.995% が一致**し、発火した 96 Cop のうち 85 個が完全一致です。
+残るのは 10 Cop（いずれも 17 件以下）と、本家がクラッシュして sonicop が正常に検出する 1 Cop です。
+内訳は [CONFORMANCE.md](CONFORMANCE.md) にあります。
+
+どちらの表も `scripts/conformance_table.rb` で再現できます。
 
 ## インストール
 
@@ -70,7 +128,7 @@ cargo install --git https://github.com/owayo/sonicop
 # 現在のプロジェクトを検査
 sonicop
 
-# Cop／部門を選択
+# Cop／デパートメントを指定
 sonicop --only Layout,Style/StringLiterals app spec
 
 # 安全な自動修正／全自動修正
@@ -106,24 +164,30 @@ Style/StringLiterals:
   EnforcedStyle: single_quotes
 ```
 
-既存コマンドとの互換性を保つため、server/LSP/MCP、plugin、cache 系の引数も受理します。
-サーバートランスポート、Ruby プラグイン実行、キャッシュ再利用、カスタム Cop、実装済み以外の
-Cop は実行しません。
-
-そのうち大半は、その旨を出力します。`--server` / `--no-server` / `--lsp` / `--mcp` /
-`--plugin` は stderr に 1 行の注記を出します。cache 系は何も出しませんが、
-無言でよいのは片方だけです。
-
-- `--cache=false` はキャッシュを使わないことを求めるもので、sonicop はそれを満たしているため、
-  無言が正しい応答です。
-- `--cache=true` はキャッシュ再利用を求めるものですが、sonicop はこれを提供しておらず、
-  それでも無言です。
-
-Cop の設定値も後者と同じ挙動です。sonicop が実装していない設定値 — たとえば
-`Style/HashSyntax` の `EnforcedShorthandSyntax` — は**警告なしに無視されます**。
-名前を綴り間違えた設定値も同様に無視されます。つまり
+既存コマンドとの互換性を保つため、server/LSP/MCP、plugin 系の引数も受理します。
+サーバートランスポート、Ruby プラグイン実行、カスタム Cop、実装済み以外の Cop は実行しません。
+これらはその旨を出力します。`--server` / `--no-server` / `--lsp` / `--mcp` / `--plugin` は
+stderr に 1 行の注記を出します。
+
+cache 系の引数は、受理するだけでなく実際に効きます。sonicop は独自の結果キャッシュを持ち、
+検査時からサイズ・更新時刻・パーミッションのいずれも動いていないファイルには、
+保存済みのレポートをそのまま返します。
+
+- キャッシュは既定で有効です。`--cache false` で無効化できます。設定ファイルの
+  `AllCops/MaxFilesInCache: 0` でも同じです。
+- 置き場所は `--cache-root DIR` で指定します。省略時は `$XDG_CACHE_HOME/sonicop`、
+  macOS では `~/Library/Caches/sonicop`、それ以外は `~/.cache/sonicop` です。
+  `--cache-root` は `--cache false` とは併用できません。
+- 保持するレポート数の上限は `AllCops/MaxFilesInCache` で、既定は本家と同じ 20,000 件です。
+- autocorrect 実行、`--stdin`、`--profile`、`--memory` では読み書きしません。
+- 本家のキャッシュとは共有しません。形式が別物であり、書いたときとまったく同じ
+  ビルドの sonicop にしかエントリを返さないためです。
+
+無言なのは Cop の設定値のほうです。sonicop が実装していない設定値は**警告なしに無視されます**。
+名前を綴り間違えた設定値も同様です。つまり
 **offense が 0 件であることは、その設定が効いた証拠にはなりません**。
 無視された設定値と、違反の無いファイルが、同じ出力になるためです。
+どの設定値まで検証済みかは上の *Cop 別の一致状況* を参照してください。
 
 Cop の*名前*は検査されます。設定ファイルに未知の Cop 名があれば、実行はエラーで止まります。
 素通りするのは、既知の Cop の中の設定値です。
@@ -139,11 +203,13 @@ Mastodon（15,286 件）で、過剰も不足もメタデータ差もありま
 **件数だけでなくパスまで**一致します（集合として比較。どの側にも余りはありません）。
 残る差分は `Lint/Syntax` に集中しています。その大半は、本家の LALR パーサが
 構文エラーから回復して出す追加診断を tree-sitter では再現できないことによるもので、
-**これが説明するのは不足の向きだけ**です。**過剰の向きは別の問題で、未調査です** —
-Homebrew の 263 件はすべて `Lint/Syntax` であることまでは確認していますが、
-移植版だけが構文エラーと呼ぶその中身は調べていません。autocorrect は RuboCop 自身のツリーと
-Mastodon でバイト単位に一致します。この 2 つは死守ラインとして扱い、バイト一致が崩れた場合は
-既知差分ではなく退行として直します。
+診断位置の差は不足と過剰の両方に出ます。Homebrew の不足 997 件・過剰 263 件はすべて
+`Lint/Syntax` ですが、**構文エラーと判定したファイル集合は 569 対 569 で完全一致**し、
+移植版だけが拒否したファイルは 0 件です。過剰 263 件は共有エラーファイル 135 件にあり、
+すべて同じファイル内の共通診断より後ろにあるため、別の受理判定バグではなく回復後の診断位置差です。
+Homebrew を問題の構文をサポートする Ruby 3.1 として測ると、両者とも `Lint/Syntax` は 0 件になります。
+autocorrect は RuboCop 自身のツリーと Mastodon でバイト単位に一致します。この 2 つは死守ラインと
+して扱い、バイト一致が崩れた場合は既知差分ではなく退行として直します。
 
 コマンド、この数値を測ったコーパスのコミット、この種の計測が誤った結論を導く 2 つの罠は
 [CONFORMANCE.md](CONFORMANCE.md) にまとめています。
@@ -182,19 +248,22 @@ Mastodon でバイト単位に一致します。この 2 つは死守ライン
 Rails・Mastodon の 3 つで**すべての offense が一致**します（計 188,812 件、どちらの側にも
 残りません）。autocorrect は前者と後者でバイト単位に一致します。
 
-再現時に注意が必要な点が 2 つあります。RuboCop は **`--cache false` と併用すると
+再現時に注意が必要な点が 3 つあります。RuboCop は **`--cache false` と併用すると
 `--parallel` を黙って無効化します**。そのためここでの並列実行はキャッシュを有効にしたうえで
 実行ごとにキャッシュディレクトリを消しており、`--cache false --parallel` で計測すると
 単一プロセスを測ることになり差が過大に出ます。また RuboCop の既定は単一プロセス、
-Sonicop は `--no-parallel` を渡さない限り並列です。
+Sonicop は `--no-parallel` を渡さない限り並列です。そして**両方ともキャッシュを空にする**
+必要があります。Sonicop も既定でキャッシュするため、同じツリーを 2 回目に流すと自分の
+キャッシュが答えてしまい、エンジンについては何も測れません。どちらにも使い捨ての
+キャッシュディレクトリを渡してください。
 
 ```bash
 # RuboCop（並列・キャッシュは毎回空・既定の全 394 Cop）
 rubocop --force-default-config --cache true --cache-root "$(mktemp -d)" \
         --no-color --parallel -f quiet
 
-# Sonicop
-sonicop --force-default-config --format quiet
+# Sonicop（キャッシュは毎回空）
+sonicop --force-default-config --cache-root "$(mktemp -d)" --format quiet
 ```
 
 測定機は Apple M2（8 コア）、Ruby 4.0.6（YJIT 利用可）、RubyGems 導入の RuboCop 1.89.0。
@@ -220,8 +289,8 @@ make gem     # source gem
 
 ### Cop の追加
 
-Cop は `src/rules/<部門>/<cop>.rs` の 1 ファイルで、公開するのは `check(context, offenses)` の
-1 関数だけです。登録は部門の `mod.rs` に 1 行を足します。
+Cop は `src/rules/<デパートメント>/<cop>.rs` の 1 ファイルで、公開するのは
+`check(context, offenses)` の 1 関数だけです。登録はデパートメントの `mod.rs` に 1 行を足します。
 
 ```rust
 department_rules! {
@@ -245,6 +314,12 @@ Cop ごとの全走査はファイル規模ではなく Cop 数に比例して
 `scripts/sync_default_yml.sh <rubocop-version>` で行い、由来のバージョンがファイル先頭に
 記録されます。
 
+`src/display_width_table.rs` も生成物で、コミットします。RuboCop は表示桁を
+`unicode-display_width` gem で数えるため、この表は手書きせず gem から生成しています。
+手書きの例外表は実際にずれており、NFD 分解された日本語でキャレットの本数が合わなくなっていました。
+再生成は `ruby scripts/dump_display_width.rb > src/display_width_table.rs` で行い、
+gem と Unicode のバージョンがファイル先頭に記録されます。
+
 依存更新には `depup --install` を使います。Ruby grammar は再現可能性のため `Cargo.toml` で
 fork のコミットを固定しています。
 

data/README.md

@@ -41,15 +41,76 @@ sonicop --show-cops
 ```
 
 **All 609 RuboCop 1.89 cops are implemented**, matched name for name against the upstream registry.
-That is a claim about the registry, not about every configuration: `Style/HashSyntax` is present and
-matches everywhere its default reaches, yet none of the four non-default values of its
-`EnforcedShorthandSyntax` are implemented. **A cop can be half absent and still match everywhere the
-default reaches** — see *Limits* in [CONFORMANCE.md](CONFORMANCE.md).
 That includes the 159 shipped as `Enabled: pending` and the 56 shipped as `Enabled: false`, which a
 default run does not reach on either side — name them with `--only` or switch them on in a
 configuration, exactly as with RuboCop. Unknown cop names still fail validation unless
 `--ignore-unrecognized-cops` is supplied.
 
+### Cop conformance
+
+All 609 cops switched on, on both sides, over the 37,491 cases RuboCop's own specs supply, each
+run at the `TargetRubyVersion` its spec asked for. A cop counts as an **exact match** only when its
+offenses agree completely: every position, message, severity and correctable flag, with nothing
+extra on either side.
+
+<!-- conformance:start -->
+| Department | Cops | Exercised | Exact match | Diverging |
+|---|---:|---:|---:|---:|
+| Bundler | 7 | 7 | **7 ✓** | 0 |
+| Gemspec | 10 | 10 | **10 ✓** | 0 |
+| Layout | 100 | 100 | 90 | 10 |
+| Lint | 157 | 157 | 147 | 10 |
+| Metrics | 10 | 10 | **10 ✓** | 0 |
+| Migration | 1 | 1 | **1 ✓** | 0 |
+| Naming | 19 | 19 | **19 ✓** | 0 |
+| Security | 7 | 7 | **7 ✓** | 0 |
+| Style | 298 | 298 | 288 | 10 |
+| **Total** | **609** | **609** | **579** | **30** |
+<!-- conformance:end -->
+
+**Read the *Exercised* column first.** A cop nothing here made fire contributes neither way — its
+silence is indistinguishable from agreement, so it would be counted as agreement without ever
+being asked. **Every one of the 609 fires here**, which is what makes the *Exact match* column
+mean what it says.
+
+Three of them took a run of their own. `Lint/DeprecatedReference`, `Lint/NameTypo` and
+`Lint/UnusedPrivateMethod` report nothing without a `rubydex` project index, which needs the gem
+installed and `AllCops/UseProjectIndex` switched on — and `Lint/DeprecatedReference` needs more
+than that: the call has to sit inside a class inheriting the one whose method carries the
+`@deprecated` tag. Upstream's own specs open with an `expect_no_offenses` saying the cop is silent
+without the index, which is easy to read as "unreachable"; it is not.
+
+Getting *Exact match* to 609 while reducing *Diverging* to zero is the current goal. More real Ruby
+does not get there: the 56 cops RuboCop ships disabled, and much of what it ships as pending, never
+fire in a plain run however large the tree. What does reach every one of them is the input its own
+specs supply — the cases recorded in `tests/fixtures/upstream_spec_capture.jsonl` touch **609 of 609
+cops**, measured. Two things about running them are easy to get wrong, and both silently shrink the
+table rather than failing:
+
+- **`TargetRubyVersion` is part of the input, not a global.** Pinning everything at 2.7 leaves
+  `Style/ArrayIntersect`, `Naming/BlockForwarding`, `Style/ItBlockParameter` and eleven others
+  unable to fire at all. Each case is run at the version its spec asked for.
+- **The filename is what several cops inspect.** `Bundler/*` needs a `Gemfile`, `Gemspec/*` a
+  `.gemspec`, and `Naming/FileName` reads the name itself. Writing every case as `.rb` had all 17
+  of those cops matching nothing.
+
+A separate direct oracle sweep on 2026-08-29 examined 11,506 cases extractable from the upstream
+cop specs. RuboCop could not read 226 of those inputs and crashed on one; among the measurable
+cases, Sonicop had **zero detection differences and zero correction differences**. This result is
+not folded into the table above: 51 cops had no directly extractable case, and three directive cops
+cannot be measured under `--only`, so the sweep does not prove that all 609 cops are exact. The
+direct sweep also uses neutral/default conditions rather than preserving every example's
+`TargetRubyVersion`. The four Ruby 3.4-sensitive cops changed in this pass were therefore compared
+separately at 3.4, where their messages and locations matched RuboCop exactly.
+
+Configuration is measured separately, because a cop that only matches at its default value is half
+a cop. Every one of the 111 cops carrying an `Enforced*` setting was switched to a **non-default**
+value at once and the corpus re-run: **99.995% of 622,317 offenses match**, with 85 of the 96 cops
+that fired matching exactly. The residue is 10 cops of at most 17 offenses each, plus one where
+RuboCop crashes and sonicop does not; the list is in [CONFORMANCE.md](CONFORMANCE.md).
+
+Reproduce either table with `scripts/conformance_table.rb`.
+
 ## Installation
 
 ```bash
@@ -114,22 +175,29 @@ Style/StringLiterals:
   EnforcedStyle: single_quotes
 ```
 
-The CLI accepts RuboCop's server/LSP/MCP, plugin, and cache flags to keep existing command lines
-parse-compatible. Sonicop does not provide server transports, Ruby plugin execution, cache reuse,
-custom Ruby cops, or cops outside the implemented set.
-
-Most of those flags say so. `--server`, `--no-server`, `--lsp`, `--mcp`, and `--plugin` each print
-a one-line notice on stderr. The cache flags print nothing, and only one of them is silent for a
-good reason:
-
-- `--cache=false` asks for no caching, which sonicop already satisfies, so silence is the correct
-  answer.
-- `--cache=true` asks for cache reuse, which sonicop does not provide, and it is silent anyway.
-
-Cop settings behave like the second case. A setting sonicop does not implement — for example
-`EnforcedShorthandSyntax` under `Style/HashSyntax` — is ignored without any warning, and so is a
-setting whose name is simply misspelled. **A run that reports no offenses is therefore not evidence
-that a setting took effect**, because an ignored setting and a clean file produce the same output.
+The CLI accepts RuboCop's server/LSP/MCP and plugin flags to keep existing command lines
+parse-compatible. Sonicop does not provide server transports, Ruby plugin execution, custom Ruby
+cops, or cops outside the implemented set. Each of those flags says so: `--server`, `--no-server`,
+`--lsp`, `--mcp`, and `--plugin` print a one-line notice on stderr.
+
+The cache flags are honoured rather than merely parsed. Sonicop keeps a result cache of its own and
+serves a stored report for a file whose size, modification time and permission bits have not moved
+since it was inspected.
+
+- Caching is on by default. `--cache false` turns it off, as does `AllCops/MaxFilesInCache: 0` in a
+  configuration file.
+- `--cache-root DIR` chooses where it lives. Without it the root is `$XDG_CACHE_HOME/sonicop`, or
+  `~/Library/Caches/sonicop` on macOS, or `~/.cache/sonicop`. `--cache-root` cannot be combined
+  with `--cache false`.
+- `AllCops/MaxFilesInCache` bounds how many reports are kept, defaulting to RuboCop's 20,000.
+- Autocorrect runs, `--stdin`, `--profile` and `--memory` neither read nor write it.
+- It is not shared with RuboCop's cache: the formats are unrelated, and an entry is only served back
+  to a build of Sonicop identical to the one that wrote it.
+
+Cop settings are the silent case. A setting sonicop does not implement is ignored without
+any warning, and so is a setting whose name is simply misspelled. **A run that reports no offenses
+is therefore not evidence that a setting took effect**, because an ignored setting and a clean file
+produce the same output. *Cop conformance* above says which values have been measured.
 
 Cop *names* are checked: an unrecognised cop in a configuration file stops the run with an error.
 It is the settings inside a recognised cop that pass unvalidated.
@@ -146,9 +214,12 @@ Mastodon (15,286), with no excess, no shortfall and no metadata differences. The
 match exactly on all five — paths, not just counts, compared as sets. What remains is concentrated in
 `Lint/Syntax`. Most of it is RuboCop's
 LALR parser recovering from an error and emitting diagnostics a tree-sitter parse cannot
-reconstruct — that direction accounts for the shortfall. **The excess is a separate question and has
-not been investigated**; on Homebrew all 263 of it is `Lint/Syntax` too, but what Sonicop calls a
-syntax error there and RuboCop does not is unexamined.
+reconstruct, and the resulting position differences go in both directions. On Homebrew all 997
+missing and 263 excess positions are `Lint/Syntax`, but the **sets of files rejected as syntax
+errors are exactly the same: 569 versus 569**, with no file rejected only by Sonicop. The 263 excess
+positions occur in 135 shared syntax-error files and every one follows a diagnostic at a position
+shared by both tools, so they are recovery-position differences rather than a separate acceptance
+bug. At Ruby 3.1, which supports the syntax used there, both tools report zero `Lint/Syntax` offenses.
 Autocorrect is byte-identical on RuboCop's own tree and on Mastodon, the two corpora held as a hard
 line: a change that breaks byte equality there is a regression, not a new known divergence.
 
@@ -188,19 +259,21 @@ The speed is not bought by skipping work: over those same 394 cops the two agree
 on RuboCop's own tree, on Rails and on Mastodon — 188,812 offenses with nothing on either side of the
 ledger — and autocorrect is byte-identical on the first and the last.
 
-Two details matter for reproducing this. RuboCop **silently turns `--parallel` off when combined
+Three details matter for reproducing this. RuboCop **silently turns `--parallel` off when combined
 with `--cache false`**, so its parallel runs here use a cache directory that is deleted before each
 run rather than disabled; timing it with `--cache false --parallel` measures a single process and
 overstates the difference. RuboCop's default is a single process, while Sonicop is parallel unless
-`--no-parallel` is passed.
+`--no-parallel` is passed. And both sides need a **cold** cache: Sonicop caches by default too, so a
+second run over the same tree answers from its own cache and measures nothing about the engine.
+Give each tool a throwaway cache root.
 
 ```bash
 # RuboCop, parallel, cold cache, its full default set of 394 cops
 rubocop --force-default-config --cache true --cache-root "$(mktemp -d)" \
         --no-color --parallel -f quiet
 
-# Sonicop
-sonicop --force-default-config --format quiet
+# Sonicop, cold cache
+sonicop --force-default-config --cache-root "$(mktemp -d)" --format quiet
 ```
 
 Machine: Apple M2 (8 cores), Ruby 4.0.6 with YJIT available, RubyGems-installed RuboCop 1.89.0.
@@ -254,6 +327,13 @@ fails when the two disagree.
 `scripts/sync_default_yml.sh <rubocop-version>`, which records the source version in the file
 header.
 
+`src/display_width_table.rs` is generated and committed too. RuboCop measures display columns with
+the `unicode-display_width` gem, so the table is taken from the gem rather than restated by hand —
+an exception table written out by hand had already drifted far enough to draw the wrong number of
+carets under decomposed Japanese. Regenerate it with
+`ruby scripts/dump_display_width.rb > src/display_width_table.rs`, which records the gem and Unicode
+versions in the file header.
+
 Dependencies are updated with `depup --install`. The Ruby grammar dependency is pinned to an exact
 fork commit in `Cargo.toml` for reproducible builds.
 

data/lib/sonicop/runner.rb

@@ -15,7 +15,13 @@ module Sonicop
       MESSAGE
 
       begin
-        exec(binary, *arguments)
+        # `[binary, binary]` = exec 形式の強制。`exec(binary, *arguments)` は arguments が空、
+        # つまり「カレントディレクトリを見る素の `sonicop`」という最も普通の呼び出しのとき
+        # `exec(String)` = commandline 形式に落ちる。この形式は空白で単語分割し、メタ文字が
+        # あればまるごと /bin/sh に渡すため、binary のパス次第で ENOENT や exit 127 になる。
+        # パスの出所は SONICOP_BINARY と gem の install 先で、どちらもメタ文字を含まない保証がない。
+        # argv0 は multi-arg 形式が既定で入れるものと同じ値を明示している。
+        exec([binary, binary], *arguments)
       rescue SystemCallError => error
         # 実行できない典型は「platform gem は入ったが実行環境と ABI が違う」ケース
         # (musl 環境に glibc ビルドが入るなど)。素の errno だけでは原因が読めない。

data/lib/sonicop/version.rb

@@ -2,5 +2,5 @@
 
 # Generated from Cargo.toml by `rake version:sync`. Do not edit by hand.
 module Sonicop
-  VERSION = '26.8.111'
+  VERSION = '26.8.112'
 end

metadata

@@ -1,7 +1,7 @@
 --- !ruby/object:Gem::Specification
 name: sonicop
 version: !ruby/object:Gem::Version
-  version: 26.8.111
+  version: 26.8.112
 platform: aarch64-linux
 authors:
 - Yohei