@hyperfixi/testing-framework 4.2.0 → 4.4.0

package/src/multilingual/english-leaks.ts

@@ -0,0 +1,835 @@
+/**
+ * English left in a translation — the M2 naturalness gate
+ * --------------------------------------------------------
+ * Every other multilingual gate asks whether a translation means what the
+ * source means. None asks whether it reads as the language: a Spanish render
+ * that keeps `repeat`, `I match` or `debounced at` parses back perfectly and
+ * passes all of them. Measured 2026-10-08, 18% of corpus renders and 24% of
+ * the command-shape scripts' renders hold at least one English word.
+ *
+ * This gate counts three kinds of finding in a render of English `source`
+ * into a language:
+ *   - `word`   an English word the engine reads as grammar in the source (or a
+ *              reference: `me`, `my`, `I`, `it`, …), written as English in the
+ *              render;
+ *   - `event`  an event the source handles or sends, written in English where
+ *              the language's lexicon has a word for it and the pair is not on
+ *              the renderer's denylist (OPEN_ITEMS N2);
+ *   - `case`   the language's nominative `me` next to a role marker, in the
+ *              languages where that form is wrong (N3, see CASE_LANGUAGES).
+ *
+ * ## What counts as English
+ * Derived, never a hand list. A source's grammar words are the tokens the
+ * engine's own parser matched as literal words while reading it
+ * (`Parser.match`, which also consumes every command and feature keyword),
+ * plus the references in its parse. A render word counts when it is one of
+ * them (up to as many times as the source uses it as grammar, so a variable
+ * named like a keyword is never counted), and the language has no word spelled
+ * the same (es `a`, de `in`, it `in` are their own words). A property name
+ * (`#a's children`) counts when the language's lexicon has a word for it.
+ * Strings, `js` bodies and brace interiors (CSS, object literals) are code and
+ * are not read.
+ *
+ * Each finding carries a context, for the report: `clause` (inside a clause
+ * semantic keeps as written, M1's `verbatimClause` / `headClause`), `bracket`,
+ * `call` (a call's arguments), `property`, or `plain`; and whether the
+ * language has its own word for it (`hasWord`: then a renderer or reader can
+ * write it; otherwise it is a vocabulary decision, roadmap policy 5).
+ *
+ * ## Inputs and baseline
+ * Two halves, one rule. The corpus half renders every non-markup corpus row
+ * (`render(parse(en), L)`, what `populate` writes) and needs a freshly
+ * populated patterns.db, like the other corpus gates (`test:canonical`). The
+ * command-shape half reads the renders the command-shape gate already makes
+ * (`translate(src, 'en', L)`), in its shards. Each half's baseline lists, per
+ * source, the findings of each language; the gate fails on a finding a pair
+ * did not have, and on a listed finding that is gone (prune it with
+ * tools/regen-english-leaks-baseline.ts), so the lists only shrink.
+ */
+
+import { readFileSync } from 'node:fs';
+import path from 'node:path';
+import { fileURLToPath } from 'node:url';
+import { readerOf, type CaseResult, type CommandShapeCase } from './command-shapes';
+import { FOREIGN_LANGUAGES } from './value-matrix';
+
+const here = path.dirname(fileURLToPath(import.meta.url));
+
+export const LEAK_BASELINE_PATHS = {
+  corpus: path.resolve(here, '../../baselines/english-leaks.corpus.json'),
+  shapes: path.resolve(here, '../../baselines/english-leaks.shapes.json'),
+} as const;
+
+export type LeakHalf = keyof typeof LEAK_BASELINE_PATHS;
+
+// ---------------------------------------------------------------------------
+// Findings
+// ---------------------------------------------------------------------------
+
+export type LeakKind = 'word' | 'event' | 'case';
+export type LeakContext = 'plain' | 'clause' | 'bracket' | 'call' | 'property';
+
+export interface LeakFinding {
+  kind: LeakKind;
+  /** The English word; for `event` the event name; for `case` the pronoun (`me`). */
+  word: string;
+  context: LeakContext;
+  /** Does the language have its own word for it? */
+  hasWord: boolean;
+  /** English a recorded decision keeps ({@link KEPT_ENGLISH}): not counted toward M2's targets. */
+  kept?: true;
+}
+
+/**
+ * English a recorded decision keeps, so M2's words target does not count it
+ * (the gate still does: it only shrinks). Contexts policy 6 keeps (roadmap §5),
+ * and the vocabulary sheet's keep rows (`docs-internal/multilingual/
+ * VOCABULARY_SHEET.md`, C1–C3, decided 2026-10-08).
+ */
+export const KEPT_ENGLISH: {
+  readonly contexts: ReadonlySet<LeakContext>;
+  readonly phrases: ReadonlyArray<{ readonly phrase: string; readonly why: string }>;
+} = {
+  // Policy 6: bracket and brace interiors, a call's arguments, a property name
+  // its reader cannot bring back, and a clause no pattern models (M1).
+  contexts: new Set(['bracket', 'call', 'property', 'clause']),
+  phrases: [
+    { phrase: 'debounced at', why: 'C1: a term of art with no settled native word' },
+    { phrase: 'throttled at', why: 'C1: a term of art with no settled native word' },
+    { phrase: 'url', why: 'C2: an international loanword (policy 6 keeps push/replace url)' },
+    { phrase: 'dom', why: 'C3: an acronym' },
+    {
+      phrase: 'view transition',
+      why: "C3: the web API's name (policy 6 keeps using view transition)",
+    },
+    { phrase: 'using', why: 'policy 6: `using view transition` (the phrase above holds the rest)' },
+  ],
+};
+
+/** A finding as the baseline writes it: the word, or `event:<name>` / `case:<pronoun>`. */
+export const findingKey = (f: Pick<LeakFinding, 'kind' | 'word'>): string =>
+  f.kind === 'word' ? f.word : `${f.kind}:${f.word}`;
+
+/**
+ * Languages whose `me` word is nominative while a marker next to it wants
+ * another case: es `a yo` (a mí), pt `a eu` (a mim), it `a io` (a me), de
+ * `zu ich` (zu mir), pl `do ja` (do mnie), ru `к я` (ко мне), uk `до я`
+ * (до мене), tr `ben e` (bana), ar `إلى أنا` (إليّ), he `אל אני` (אליי),
+ * tl `sa ako` (sa akin), hi `मैं में` (मुझमें), bn `আমি তে` (আমাতে). The other
+ * ten are right as written (fr `à moi`, ja `自分 に`, zh/id/ms/vi/th/ko/sw/qu
+ * do not inflect the pronoun for case here). OPEN_ITEMS N3.
+ */
+export const CASE_LANGUAGES: ReadonlySet<string> = new Set([
+  'ar',
+  'bn',
+  'de',
+  'es',
+  'he',
+  'hi',
+  'it',
+  'pl',
+  'pt',
+  'ru',
+  'tl',
+  'tr',
+  'uk',
+]);
+
+// ---------------------------------------------------------------------------
+// What a source holds
+// ---------------------------------------------------------------------------
+
+export type LeakReader = 'program' | 'statements';
+
+export interface SourceInfo {
+  /** Words the engine read as grammar, with how often. */
+  grammar: Map<string, number>;
+  /** Property names in the source (`my value`, `#a's children`), with how often. */
+  props: Map<string, number>;
+  /** Events the source handles, sends or triggers, with how often. */
+  events: Map<string, number>;
+  /** `js` blocks: never translated, never read. */
+  jsBodies: JsBlock[];
+  /** Words of the clauses semantic keeps as written. */
+  clauseWords: Set<string>;
+}
+
+/** References the engine reads as names; semantic translates them (`me` → es `yo`). */
+const REFERENCE_NAMES: ReadonlySet<string> = new Set([
+  'I',
+  'me',
+  'my',
+  'it',
+  'its',
+  'you',
+  'your',
+  'result',
+  'event',
+  'target',
+  'detail',
+  'body',
+  'document',
+  'window',
+]);
+
+const bump = (map: Map<string, number>, key: string, by = 1): void => {
+  map.set(key, (map.get(key) ?? 0) + by);
+};
+
+/** A `js` block: its parameter names (the script's, passed in) and its body (JavaScript). */
+export interface JsBlock {
+  params: string;
+  body: string;
+}
+
+/** The `js` blocks of a source: `js`, an optional parameter list, a body, `end`. */
+function jsBlocksOf(source: string): JsBlock[] {
+  return [...source.matchAll(/\bjs\b\s*(?:\(([^)]*)\))?([\s\S]*?)\bend\b/g)]
+    .map(m => ({ params: (m[1] ?? '').trim(), body: (m[2] ?? '').trim() }))
+    .filter(b => b.body || b.params);
+}
+
+const escapeRegExp = (text: string): string => text.replace(/[.*+?^${}()|[\]\\]/g, '\\$&');
+
+// ---------------------------------------------------------------------------
+// What a language holds
+// ---------------------------------------------------------------------------
+
+interface KeywordTranslationLike {
+  primary?: string;
+  alternatives?: readonly string[];
+  render?: boolean;
+}
+interface ProfileLike {
+  keywords?: Record<string, KeywordTranslationLike>;
+  roleMarkers?: Record<string, KeywordTranslationLike & { position?: string }>;
+  references?: Record<string, string>;
+  possessive?: {
+    marker?: string;
+    specialForms?: Record<string, string>;
+    keywords?: Record<string, string>;
+  };
+  eventHandler?: Record<string, unknown>;
+}
+type LexiconLike = Record<string, Record<string, KeywordTranslationLike> | undefined>;
+
+const LEXICON_CATEGORIES = ['values', 'expressions', 'logical', 'temporal', 'attributes'] as const;
+
+const wordsOf = (text: string | undefined): string[] => (text ?? '').split(/\s+/).filter(Boolean);
+const spellings = (t: KeywordTranslationLike | undefined): string[] =>
+  t ? [t.primary ?? '', ...(t.alternatives ?? [])].flatMap(wordsOf) : [];
+
+interface LanguageWords {
+  /** Every word the language writes: a render word spelled like one is its own. */
+  native: Set<string>;
+  /** English words the language has a word for (lexicon, keywords, references). */
+  translated: Set<string>;
+  /** Events its lexicon names, minus the renderer's denylist for it. */
+  events: Set<string>;
+  /** Its `me` word, and its role markers by side. */
+  me: string | undefined;
+  markersBefore: Set<string>;
+  markersAfter: Set<string>;
+}
+
+function languageWords(
+  profile: ProfileLike | undefined,
+  lexicon: LexiconLike | undefined,
+  denied: ReadonlySet<string> | undefined
+): LanguageWords {
+  const native = new Set<string>();
+  const translated = new Set<string>();
+  const add = (english: string, t: KeywordTranslationLike | undefined): void => {
+    const forms = spellings(t);
+    // An alternative spelled like the English word, beside a primary that is
+    // the language's own, is English the reader still takes (code-switched
+    // input), not a word the language writes: a render that writes it leaks.
+    // Counted as native, it hid every such leak (`forever` in the eight
+    // languages whose primary is their own, V1).
+    const englishWords = new Set(wordsOf(english));
+    const primary = wordsOf(t?.primary);
+    const ownPrimary = primary.some(w => !englishWords.has(w));
+    for (const w of forms)
+      if (!(ownPrimary && englishWords.has(w) && !primary.includes(w))) native.add(w);
+    if (forms.length && !forms.includes(english))
+      for (const w of wordsOf(english)) translated.add(w);
+  };
+  for (const [english, t] of Object.entries(profile?.keywords ?? {})) add(english, t);
+  for (const t of Object.values(profile?.roleMarkers ?? {}))
+    for (const w of spellings(t)) native.add(w);
+  for (const [english, n] of Object.entries(profile?.references ?? {}))
+    add(english, { primary: n });
+  for (const [english, n] of Object.entries(profile?.possessive?.specialForms ?? {}))
+    add(english === 'me' ? 'my' : english === 'it' ? 'its' : english === 'you' ? 'your' : english, {
+      primary: n,
+    });
+  // Native form → the English reference it stands for (es `mi` → `me`).
+  for (const n of Object.keys(profile?.possessive?.keywords ?? {}))
+    for (const w of wordsOf(n)) native.add(w);
+  for (const w of wordsOf(profile?.possessive?.marker)) native.add(w);
+  const handler = JSON.stringify(profile?.eventHandler ?? {});
+  for (const m of handler.matchAll(
+    /"(?:primary|alternatives|temporalMarkers)":(\[[^\]]*\]|"[^"]*")/g
+  ))
+    for (const w of (JSON.parse(m[1] ?? '""') as string | string[]).toString().split(/[\s,]+/))
+      if (w) native.add(w);
+  for (const category of LEXICON_CATEGORIES)
+    for (const [english, t] of Object.entries(lexicon?.[category] ?? {})) add(english, t);
+  const events = new Set<string>();
+  for (const [english, t] of Object.entries(lexicon?.events ?? {})) {
+    for (const w of spellings(t)) native.add(w);
+    if (t.render !== false && t.primary && t.primary !== english && !denied?.has(english))
+      events.add(english);
+  }
+  const markersBefore = new Set<string>();
+  const markersAfter = new Set<string>();
+  for (const t of Object.values(profile?.roleMarkers ?? {}))
+    for (const w of spellings(t)) (t.position === 'after' ? markersAfter : markersBefore).add(w);
+  return {
+    native,
+    translated,
+    events,
+    me: profile?.references?.me,
+    markersBefore,
+    markersAfter,
+  };
+}
+
+// ---------------------------------------------------------------------------
+// Reading a render
+// ---------------------------------------------------------------------------
+
+/** Blank out `{…}` groups: CSS blocks and object literals are code. */
+function blankBraces(text: string): string {
+  let out = '';
+  let depth = 0;
+  for (const c of text) {
+    if (c === '{') depth++;
+    out += depth > 0 ? ' ' : c;
+    if (c === '}' && depth > 0) depth--;
+  }
+  return out;
+}
+
+/** Strings, `js` bodies and brace interiors replaced by spaces (offsets kept). */
+export function codeOnly(render: string, jsBlocks: readonly JsBlock[]): string {
+  const blank = (m: string): string => ' '.repeat(m.length);
+  let text = render;
+  for (const { params, body } of jsBlocks) {
+    const list = params
+      ? `\\(\\s*${params
+          .split(/\s*,\s*/)
+          .map(escapeRegExp)
+          .join('\\s*,\\s*')}\\s*\\)\\s*`
+      : '';
+    text = text.replace(new RegExp(`${list}${escapeRegExp(body)}`, 'g'), blank);
+  }
+  return blankBraces(text)
+    .replace(/`[^`]*`/g, blank)
+    .replace(/"(?:[^"\\]|\\.)*"/g, blank)
+    .replace(/(?<![\p{L}\p{N}])'(?:[^'\\\n]|\\.)*'/gu, blank);
+}
+
+/** Where `[…]` groups and a call's `(…)` arguments sit, innermost last. */
+function groupSpans(text: string): Array<[number, number, 'bracket' | 'call']> {
+  const spans: Array<[number, number, 'bracket' | 'call']> = [];
+  const open: Array<[number, string, boolean]> = [];
+  for (let i = 0; i < text.length; i++) {
+    const c = text[i];
+    if (c === '[' || c === '(')
+      open.push([i, c, c === '(' && i > 0 && /[\w$]/.test(text[i - 1] ?? '')]);
+    else if ((c === ']' || c === ')') && open.length) {
+      const [start, kind, call] = open.pop()!;
+      if (kind === '[') spans.push([start, i, 'bracket']);
+      else if (call) spans.push([start, i, 'call']);
+    }
+  }
+  return spans.sort((a, b) => a[0] - b[0]);
+}
+
+interface RenderWord {
+  word: string;
+  offset: number;
+}
+
+/** The words of a render's code, each with its offset. */
+function renderWords(text: string): RenderWord[] {
+  const out: RenderWord[] = [];
+  for (const m of text.matchAll(/\S+/g)) {
+    let token = m[0];
+    let offset = m.index ?? 0;
+    const lead = /^[([,]+/.exec(token)?.[0].length ?? 0;
+    token = token.slice(lead);
+    offset += lead;
+    out.push({ word: token.replace(/[)\],;]+$/, '').replace(/'s$/, ''), offset });
+  }
+  return out;
+}
+
+// ---------------------------------------------------------------------------
+// The scanner
+// ---------------------------------------------------------------------------
+
+export interface LeakScanner {
+  /** What the engine and semantic read in a source; `null` when the engine rejects it. */
+  sourceInfo(source: string, reader: LeakReader): SourceInfo | null;
+  /** The findings of one render of `info`'s source into `language`. */
+  scan(render: string, info: SourceInfo, language: string): LeakFinding[];
+}
+
+interface EngineModule {
+  register(...modules: unknown[]): void;
+  everything: unknown[];
+  parse(source: string): unknown;
+  parseProgram(source: string): unknown;
+  Parser: { prototype: { match(value: string, type?: string): unknown } };
+}
+interface SemanticModule {
+  parse(source: string, language: string): unknown;
+  tryGetProfile(language: string): unknown;
+  getLexicon(language: string): unknown;
+  getEventLocalizationDenylist(): Record<string, ReadonlySet<string>>;
+}
+
+/** Load the engine and semantic (ESM-only, so imported here) and return a scanner. */
+export async function initLeakScanner(): Promise<LeakScanner> {
+  const engine = (await import('@hyperfixi/engine')) as unknown as EngineModule;
+  engine.register(...engine.everything);
+  const semantic = (await import('@lokascript/semantic')) as unknown as SemanticModule;
+  const denylist = semantic.getEventLocalizationDenylist();
+  const languages = new Map<string, LanguageWords>();
+  for (const language of FOREIGN_LANGUAGES) {
+    languages.set(
+      language,
+      languageWords(
+        semantic.tryGetProfile(language) as ProfileLike | undefined,
+        semantic.getLexicon(language) as LexiconLike | undefined,
+        denylist[language]
+      )
+    );
+  }
+
+  /** Parse with the engine, recording every token its grammar matched as a word. */
+  const readGrammar = (
+    source: string,
+    reader: LeakReader
+  ): { ast: unknown; matched: Set<object> } | null => {
+    const proto = engine.Parser.prototype;
+    const original = proto.match;
+    const matched = new Set<object>();
+    proto.match = function (this: unknown, value: string, type?: string) {
+      const token = original.call(this, value, type);
+      if (token && (type ?? 'IDENTIFIER') === 'IDENTIFIER') matched.add(token as object);
+      return token;
+    };
+    try {
+      return {
+        ast: reader === 'program' ? engine.parseProgram(source) : engine.parse(source),
+        matched,
+      };
+    } catch {
+      return null;
+    } finally {
+      proto.match = original;
+    }
+  };
+
+  return {
+    sourceInfo(source, reader) {
+      const read = readGrammar(source, reader);
+      if (!read) return null;
+      const grammar = new Map<string, number>();
+      for (const token of read.matched) bump(grammar, (token as { value: string }).value);
+      const references = new Map<string, number>();
+      const props = new Map<string, number>();
+      const events = new Map<string, number>();
+      const seen = new WeakSet<object>();
+      const walk = (value: unknown, key: string): void => {
+        if (typeof value === 'string') {
+          if (key === 'name' && REFERENCE_NAMES.has(value)) bump(references, value);
+          else if (key === 'prop') bump(props, value);
+          else if (key === 'on' || key === 'event') bump(events, value);
+          return;
+        }
+        if (!value || typeof value !== 'object' || seen.has(value)) return;
+        seen.add(value);
+        for (const [k, v] of Object.entries(value))
+          if (typeof v !== 'function') walk(v, Array.isArray(value) ? key : k);
+      };
+      walk(read.ast, '');
+      for (const [word, n] of references) grammar.set(word, Math.max(grammar.get(word) ?? 0, n));
+
+      const clauseWords = new Set<string>();
+      let node: unknown = null;
+      try {
+        node = semantic.parse(source, 'en');
+      } catch {
+        // No clause words: every finding reads as plain.
+      }
+      const clauses = (n: unknown): void => {
+        if (!n || typeof n !== 'object') return;
+        if (Array.isArray(n)) return n.forEach(clauses);
+        const record = n as Record<string, unknown>;
+        for (const k of ['verbatimClause', 'headClause'])
+          if (typeof record[k] === 'string')
+            for (const w of wordsOf(record[k] as string)) clauseWords.add(w);
+        for (const v of Object.values(record)) if (v && typeof v === 'object') clauses(v);
+      };
+      clauses(node);
+      return { grammar, props, events, jsBodies: jsBlocksOf(source), clauseWords };
+    },
+
+    scan(render, info, language) {
+      const lang = languages.get(language);
+      if (!lang) throw new Error(`no vocabulary for ${language}`);
+      const text = codeOnly(render, info.jsBodies);
+      const spans = groupSpans(text);
+      const contextAt = (offset: number): 'bracket' | 'call' | undefined => {
+        let found: 'bracket' | 'call' | undefined;
+        for (const [start, end, kind] of spans) if (start < offset && offset < end) found = kind;
+        return found;
+      };
+      const keptSpans: Array<[number, number]> = [];
+      for (const { phrase } of KEPT_ENGLISH.phrases)
+        for (const m of text.matchAll(
+          new RegExp(
+            `(?<![\\p{L}\\p{N}])${escapeRegExp(phrase).replace(/ /g, '\\s+')}(?![\\p{L}\\p{N}])`,
+            'gu'
+          )
+        ))
+          keptSpans.push([m.index ?? 0, (m.index ?? 0) + m[0].length]);
+      const isKept = (offset: number, context: LeakContext): boolean =>
+        KEPT_ENGLISH.contexts.has(context) ||
+        keptSpans.some(([start, end]) => start <= offset && offset < end);
+      const findings: LeakFinding[] = [];
+      const grammarLeft = new Map(info.grammar);
+      const propsLeft = new Map(info.props);
+      const eventsLeft = new Map(info.events);
+      const words = renderWords(text);
+      for (const { word, offset } of words) {
+        if (!/^[A-Za-z]+$/.test(word) || lang.native.has(word)) continue;
+        if ((eventsLeft.get(word) ?? 0) > 0) {
+          eventsLeft.set(word, eventsLeft.get(word)! - 1);
+          if (lang.events.has(word))
+            findings.push({ kind: 'event', word, context: 'plain', hasWord: true });
+          continue;
+        }
+        const grammar = (grammarLeft.get(word) ?? 0) > 0;
+        const prop = !grammar && (propsLeft.get(word) ?? 0) > 0 && lang.translated.has(word);
+        if (!grammar && !prop) continue;
+        (grammar ? grammarLeft : propsLeft).set(
+          word,
+          (grammar ? grammarLeft : propsLeft).get(word)! - 1
+        );
+        const context: LeakContext = info.clauseWords.has(word)
+          ? 'clause'
+          : (contextAt(offset) ?? (prop ? 'property' : 'plain'));
+        findings.push({
+          kind: 'word',
+          word,
+          context,
+          hasWord: lang.translated.has(word),
+          ...(isKept(offset, context) ? { kept: true as const } : {}),
+        });
+      }
+      if (CASE_LANGUAGES.has(language) && lang.me) {
+        const tokens = words.map(w => w.word);
+        for (let i = 0; i < tokens.length; i++) {
+          if (tokens[i] !== lang.me) continue;
+          const before = tokens[i - 1];
+          const after = tokens[i + 1];
+          if ((before && lang.markersBefore.has(before)) || (after && lang.markersAfter.has(after)))
+            findings.push({ kind: 'case', word: 'me', context: 'plain', hasWord: true });
+        }
+      }
+      return findings;
+    },
+  };
+}
+
+// ---------------------------------------------------------------------------
+// Baseline
+// ---------------------------------------------------------------------------
+
+/** One half's committed findings: per source, a findings string → the languages that have it. */
+export interface LeakBaseline {
+  description: string;
+  /** Renders read, renders with a finding, and findings, at the last regeneration. */
+  renders: number;
+  leaky: number;
+  findings: number;
+  /** The same, per kind: `word.renders` is the leak rate's numerator. */
+  kinds: Record<LeakKind, { renders: number; findings: number }>;
+  /** Per source id: `"<finding> <finding> …"` (sorted keys) → languages (`*` for all 23). */
+  entries: Record<string, Record<string, string>>;
+  /**
+   * Per source id, the languages with no render (the translation was refused).
+   * A lane that starts to render is progress, not a leak: the gate reports it
+   * as one to prune, and its English is counted from then on.
+   */
+  unrendered?: Record<string, string>;
+}
+
+/** The results of one half: per source id, per language, the findings (`null`: no render). */
+export type LeakResults = Map<string, Map<string, LeakFinding[] | null>>;
+
+const ALL = '*';
+export function compressLanguages(languages: readonly string[]): string {
+  const set = new Set(languages);
+  return FOREIGN_LANGUAGES.every(l => set.has(l))
+    ? ALL
+    : FOREIGN_LANGUAGES.filter(l => set.has(l)).join(' ');
+}
+export function expandLanguages(text: string): string[] {
+  return text === ALL ? [...FOREIGN_LANGUAGES] : text.split(' ').filter(Boolean);
+}
+
+export const keysOf = (findings: readonly LeakFinding[] | null): string[] =>
+  (findings ?? []).map(findingKey).sort();
+
+export function leakBaselineFrom(results: LeakResults, description: string): LeakBaseline {
+  const doc: LeakBaseline = {
+    description,
+    renders: 0,
+    leaky: 0,
+    findings: 0,
+    kinds: {
+      word: { renders: 0, findings: 0 },
+      event: { renders: 0, findings: 0 },
+      case: { renders: 0, findings: 0 },
+    },
+    entries: {},
+    unrendered: {},
+  };
+  for (const [id, byLanguage] of [...results].sort((a, b) => a[0].localeCompare(b[0]))) {
+    const groups = new Map<string, string[]>();
+    const unrendered: string[] = [];
+    for (const [language, findings] of byLanguage) {
+      if (!findings) {
+        unrendered.push(language);
+        continue;
+      }
+      doc.renders++;
+      if (!findings.length) continue;
+      doc.leaky++;
+      doc.findings += findings.length;
+      for (const kind of ['word', 'event', 'case'] as const) {
+        const n = findings.filter(f => f.kind === kind).length;
+        doc.kinds[kind].findings += n;
+        if (n) doc.kinds[kind].renders++;
+      }
+      const key = keysOf(findings).join(' ');
+      groups.set(key, [...(groups.get(key) ?? []), language]);
+    }
+    if (unrendered.length) doc.unrendered![id] = compressLanguages(unrendered);
+    if (!groups.size) continue;
+    doc.entries[id] = Object.fromEntries(
+      [...groups]
+        .sort((a, b) => a[0].localeCompare(b[0]))
+        .map(([key, ls]) => [key, compressLanguages(ls)])
+    );
+  }
+  return doc;
+}
+
+/** The findings a baseline lists for one (source, language). */
+export function listedKeys(
+  baseline: Pick<LeakBaseline, 'entries'>,
+  id: string,
+  language: string
+): string[] {
+  for (const [key, languages] of Object.entries(baseline.entries[id] ?? {}))
+    if (expandLanguages(languages).includes(language)) return key.split(' ');
+  return [];
+}
+
+export interface LeakChange {
+  id: string;
+  language: string;
+  /** Findings the pair has now and the baseline does not list. */
+  added: string[];
+  /** Findings the baseline lists and the pair no longer has. */
+  gone: string[];
+  /** The baseline lists the pair as unrendered and it renders now: `added` is its English. */
+  newRender?: boolean;
+}
+
+/** Multiset difference `a − b`. */
+function minus(a: readonly string[], b: readonly string[]): string[] {
+  const left = [...b];
+  return a.filter(x => {
+    const i = left.indexOf(x);
+    if (i < 0) return true;
+    left.splice(i, 1);
+    return false;
+  });
+}
+
+/** Compare results with a baseline, over the sources and languages that ran. */
+export function diffLeakBaseline(
+  results: LeakResults,
+  baseline: Pick<LeakBaseline, 'entries' | 'unrendered'>
+): LeakChange[] {
+  const changes: LeakChange[] = [];
+  for (const [id, byLanguage] of results) {
+    const unrendered = expandLanguages(baseline.unrendered?.[id] ?? '');
+    for (const [language, findings] of byLanguage) {
+      const wasUnrendered = unrendered.includes(language);
+      if (findings && wasUnrendered) {
+        changes.push({ id, language, added: keysOf(findings), gone: [], newRender: true });
+        continue;
+      }
+      const got = keysOf(findings);
+      const listed = listedKeys(baseline, id, language);
+      const added = minus(got, listed);
+      const gone = minus(listed, got);
+      if (added.length || gone.length) changes.push({ id, language, added, gone });
+    }
+  }
+  return changes;
+}
+
+/** Findings a pair gained: a regression (a lane that starts to render is not one). */
+export const gainedLeaks = (changes: readonly LeakChange[]): LeakChange[] =>
+  changes.filter(c => c.added.length > 0 && !c.newRender);
+
+/** Changes the baseline must be pruned for: findings gone, or a lane that renders now. */
+export const prunableLeaks = (changes: readonly LeakChange[]): LeakChange[] =>
+  changes.filter(c => c.gone.length > 0 || c.newRender);
+
+export function loadLeakBaseline(half: LeakHalf): LeakBaseline {
+  return JSON.parse(readFileSync(LEAK_BASELINE_PATHS[half], 'utf8')) as LeakBaseline;
+}
+
+// ---------------------------------------------------------------------------
+// M2's exit targets
+// ---------------------------------------------------------------------------
+
+/**
+ * M2's exit targets (decided 2026-10-08): the most of a language's renders that
+ * may hold a finding of each kind, in EVERY language. Per language because a
+ * reader reads one, and an average hid he (35% of corpus renders with an English
+ * word where the others sit near 18%). Words count only English no recorded
+ * decision keeps ({@link KEPT_ENGLISH}). Events are already net of the denylist,
+ * whose entries carry reasons.
+ */
+export const LEAK_TARGETS: Readonly<Record<LeakHalf, Readonly<Record<LeakKind, number>>>> = {
+  corpus: { word: 0.05, event: 0, case: 0 },
+  shapes: { word: 0.1, event: 0, case: 0 },
+};
+
+const KINDS: readonly LeakKind[] = ['word', 'event', 'case'];
+
+export interface LanguageLeakRates {
+  renders: number;
+  /** The share of the language's renders with a finding of each kind. */
+  rate: Record<LeakKind, number>;
+  /** Whether each rate is within {@link LEAK_TARGETS}. */
+  met: Record<LeakKind, boolean>;
+}
+
+/** Each language's leak rates against the half's targets; a lane with no render is not counted. */
+export function leakRatesByLanguage(
+  results: LeakResults,
+  half: LeakHalf
+): Map<string, LanguageLeakRates> {
+  const counts = new Map<string, { renders: number } & Record<LeakKind, number>>();
+  for (const byLanguage of results.values()) {
+    for (const [language, findings] of byLanguage) {
+      if (!findings) continue;
+      const row = counts.get(language) ?? { renders: 0, word: 0, event: 0, case: 0 };
+      row.renders++;
+      for (const kind of KINDS) if (findings.some(f => f.kind === kind && !f.kept)) row[kind]++;
+      counts.set(language, row);
+    }
+  }
+  const out = new Map<string, LanguageLeakRates>();
+  for (const [language, row] of counts) {
+    const rate = { word: 0, event: 0, case: 0 };
+    const met = { word: true, event: true, case: true };
+    for (const kind of KINDS) {
+      rate[kind] = row[kind] / row.renders;
+      met[kind] = rate[kind] <= LEAK_TARGETS[half][kind];
+    }
+    out.set(language, { renders: row.renders, rate, met });
+  }
+  return out;
+}
+
+// ---------------------------------------------------------------------------
+// The two inputs
+// ---------------------------------------------------------------------------
+
+/** The command-shape half: the renders the command-shape gate made for `results`. */
+export function shapeLeaks(
+  scanner: LeakScanner,
+  results: readonly CaseResult[],
+  cases: readonly CommandShapeCase[]
+): LeakResults {
+  const byId = new Map(cases.map(c => [c.id, c]));
+  const out: LeakResults = new Map();
+  for (const r of results) {
+    const shape = byId.get(r.id);
+    if (!shape) continue;
+    const info = scanner.sourceInfo(shape.source, readerOf(shape));
+    if (!info) continue;
+    const byLanguage = new Map<string, LeakFinding[] | null>();
+    for (const language of FOREIGN_LANGUAGES) {
+      const rendered = r.lanes[language]?.rendered;
+      byLanguage.set(
+        language,
+        rendered === undefined ? null : scanner.scan(rendered, info, language)
+      );
+    }
+    out.set(r.id, byLanguage);
+  }
+  return out;
+}
+
+interface CorpusRow {
+  id: string;
+  rawCode: string;
+}
+
+/**
+ * The corpus half: every non-markup row rendered as `populate` writes it
+ * (`render(parse(en), L)`). A row the engine rejects has no grammar to read,
+ * so it is skipped; the caller can count them from `skipped`.
+ */
+export async function corpusLeaks(
+  scanner: LeakScanner,
+  rows: readonly CorpusRow[]
+): Promise<{ results: LeakResults; skipped: string[] }> {
+  const { parse, render } = (await import('@lokascript/semantic')) as unknown as {
+    parse(source: string, language: string): unknown;
+    render(node: unknown, language: string): string;
+  };
+  const results: LeakResults = new Map();
+  const skipped: string[] = [];
+  for (const row of rows) {
+    if (row.rawCode.trimStart().startsWith('<')) continue;
+    const info = scanner.sourceInfo(row.rawCode, 'program');
+    let node: unknown = null;
+    try {
+      node = parse(row.rawCode, 'en');
+    } catch {
+      node = null;
+    }
+    if (!info || !node) {
+      skipped.push(row.id);
+      continue;
+    }
+    const byLanguage = new Map<string, LeakFinding[] | null>();
+    for (const language of FOREIGN_LANGUAGES) {
+      let rendered: string | null = null;
+      try {
+        rendered = render(node, language);
+      } catch {
+        rendered = null;
+      }
+      byLanguage.set(language, rendered ? scanner.scan(rendered, info, language) : null);
+    }
+    results.set(row.id, byLanguage);
+  }
+  return { results, skipped };
+}