champollion 0.3.3 → 0.4.0

package/lib/placeholders.js

@@ -0,0 +1,394 @@
+/**
+ * placeholders.js — the placeholders a value carries, and what a translation
+ * did to them. ONE rule for the sync quality gate (lib/validate.js) and for
+ * verify (lib/verify.js, through lib/integrity.js): sync refuses what verify
+ * would flag, so the fallback runs instead of the damage being written and
+ * then reported (Round 12, i18next: the gate did not check {{…}}, so sync
+ * wrote {{nom}} for {{name}}, verify flagged it, and the run exited 2).
+ *
+ * A leaf module (it imports only lib/icu-structure.js), so the gate and the
+ * integrity audit can both use it without importing each other.
+ */
+
+import { parseMessage, hasArguments } from './icu-structure.js';
+
+/**
+ * i18next interpolation — `{{name}}`, `{{ user.name }}`, `{{- html}}`
+ * (unescaped), `{{val, number}}` (formatted). A leading dot is Go template
+ * syntax (Hugo "{{ .Count }}") and is not matched.
+ */
+const I18NEXT_INTERPOLATION = /\{\{\s*(-\s*)?([A-Za-z_$][\w.$]*)\s*(?:,\s*([^{}]*?))?\s*\}\}/g;
+
+/**
+ * Extract ICU-style placeholders from a string.
+ *
+ * Handles:
+ *   - Simple: {name}, {count}            → "name", "count"
+ *   - Nested ICU: {count, plural, one {# item} other {# items}}
+ *   - React-intl: <bold>text</bold>       → "<bold>"
+ *   - i18next: {{name}}, {{val, number}}  → "{{name}}", "{{val, number}}"
+ *
+ * The token's form names its syntax (placeholderSyntax). An i18next token
+ * keeps its double braces: read as "name" it was indistinguishable from
+ * "{name}", which i18next prints literally — a `{{name}}` → `{name}` loss
+ * passed, and a `{{name}}` loss could not be reported as i18next (Round 12).
+ *
+ * @param {string} text - Translation string
+ * @returns {string[]} Sorted array of placeholder tokens
+ */
+function extractPlaceholders(text) {
+  if (typeof text !== 'string') return [];
+
+  const placeholders = new Set();
+
+  // i18next {{…}} first, then blanked out, so the single-brace scan below
+  // does not read the inner "{name}" of "{{name}}" as an ICU argument.
+  const rest = text.replace(I18NEXT_INTERPOLATION, (whole, unescaped, name, format) => {
+    placeholders.add(`{{${unescaped ? '- ' : ''}${name}${format ? `, ${format.trim().replace(/\s+/g, ' ')}` : ''}}}`);
+    return ' ';
+  });
+
+  // Simple ICU placeholders: {name}, {count}
+  // Match top-level braces only (not nested plurals)
+  const simplePattern = /\{(\w+)(?:[,}])/g;
+  let match;
+  while ((match = simplePattern.exec(rest)) !== null) {
+    placeholders.add(match[1]);
+  }
+
+  // React-intl XML tags: <bold>, </bold>, <link>, </link>
+  const xmlPattern = /<\/?(\w+)>/g;
+  while ((match = xmlPattern.exec(text)) !== null) {
+    placeholders.add(`<${match[1]}>`);
+  }
+
+  return [...placeholders].sort();
+}
+
+/**
+ * The syntax a token from extractPlaceholders is written in — read off the
+ * token's own form, which is how extractPlaceholders tells them apart:
+ *   'i18next'  "{{name}}"  (i18next interpolation)
+ *   'markup'   "<bold>"    (a tag: react-intl, react-i18next, HTML)
+ *   'brace'    "name"      (a single-brace "{name}" — an ICU/MessageFormat
+ *                           argument when the message parses as ICU, which
+ *                           lib/icu-structure.js checkICUStructure judges)
+ * printf conversions (%s, %(name)s) are not tokens here: checkICUStructure
+ * compares them and tags its findings 'printf'.
+ *
+ * @param {string} token
+ * @returns {'i18next'|'markup'|'brace'}
+ */
+function placeholderSyntax(token) {
+  if (token.startsWith('{{')) return 'i18next';
+  if (token.startsWith('<')) return 'markup';
+  return 'brace';
+}
+
+/**
+ * A token as the user wrote it: "name" was "{name}".
+ *
+ * @param {string} token
+ * @returns {string}
+ */
+function showPlaceholder(token) {
+  return placeholderSyntax(token) === 'brace' ? `{${token}}` : token;
+}
+
+/**
+ * Compare placeholders between source and target strings.
+ *
+ * @param {string} sourceValue - Source locale value
+ * @param {string} targetValue - Target locale value
+ * @returns {{ missing: string[], extra: string[] }} Placeholder differences
+ */
+function comparePlaceholders(sourceValue, targetValue) {
+  const sourcePH = extractPlaceholders(sourceValue);
+  const targetPH = extractPlaceholders(targetValue);
+
+  const missing = sourcePH.filter(p => !targetPH.includes(p));
+  const extra = targetPH.filter(p => !sourcePH.includes(p));
+
+  return { missing, extra };
+}
+
+/**
+ * Does `text` parse as an ICU message with arguments (either apostrophe
+ * reading — checkICUStructure's own test)? Its single-brace tokens are then
+ * the ICU check's to judge, argument by argument.
+ *
+ * @param {unknown} text
+ * @returns {boolean}
+ */
+function parsesAsICU(text) {
+  if (typeof text !== 'string' || !text.includes('{')) return false;
+  return ['icu', 'literal'].some((apostrophes) => {
+    const parsed = parseMessage(text, { apostrophes });
+    return parsed.ok && hasArguments(parsed.nodes);
+  });
+}
+
+/** What each syntax is called in a finding: "i18next {{…}} placeholder …". */
+const PLACEHOLDER_SYNTAX_NAMES = Object.freeze({
+  i18next: 'i18next {{…}}',
+  brace: '{…}',
+  markup: 'tag',
+});
+
+/**
+ * What a translation did to the placeholders its source carries, each change
+ * named by its syntax (placeholderSyntax) — the findings verify reports, and
+ * the ones the sync gate refuses:
+ *   { syntax: 'i18next', token: '{{name}}', issue: 'placeholder {{name}} was changed to {{nom}}' }
+ *
+ * A missing token and an extra one of the same name read as one change
+ * ("{{name}} was changed to {name}") and are named by the SOURCE's syntax;
+ * the remaining missing/extra tokens pair up in order. Left out:
+ *   - a brace token (single or double) when the source parses as an ICU
+ *     message — the ICU check compares its arguments, and the token scan
+ *     reads a one-word plural branch ("other {articles}") as a placeholder
+ *     and a branch holding only an argument ("other{{count}}") as i18next;
+ *   - a tag when the markup check already reported the value
+ *     (`markupReported`), said precisely there.
+ * printf conversions are not tokens here: checkICUStructure compares them.
+ *
+ * @param {string} sourceValue
+ * @param {string} targetValue
+ * @param {{ markupReported?: boolean }} [options]
+ * @returns {Array<{ syntax: 'i18next'|'brace'|'markup', token: string, issue: string }>}
+ */
+function placeholderChanges(sourceValue, targetValue, { markupReported = false } = {}) {
+  if (typeof sourceValue !== 'string' || typeof targetValue !== 'string') return [];
+  const { missing, extra } = comparePlaceholders(sourceValue, targetValue);
+  if (missing.length === 0 && extra.length === 0) return [];
+  const nameOf = (token) => (/[A-Za-z_$][\w.$]*/.exec(token) || [token])[0];
+  const changes = [];
+  for (const m of [...missing]) {
+    const e = extra.find(x => nameOf(x) === nameOf(m));
+    if (!e) continue;
+    changes.push([m, e]);
+    missing.splice(missing.indexOf(m), 1);
+    extra.splice(extra.indexOf(e), 1);
+  }
+  while (missing.length > 0 && extra.length > 0) changes.push([missing.shift(), extra.shift()]);
+  const items = [
+    ...changes.map(([m, e]) => [m, `placeholder ${showPlaceholder(m)} was changed to ${showPlaceholder(e)}`]),
+    ...missing.map(m => [m, `placeholder ${showPlaceholder(m)} is missing`]),
+    ...extra.map(e => [e, `placeholder ${showPlaceholder(e)} is not in the source`]),
+  ];
+  const icuSource = parsesAsICU(sourceValue);
+  const out = [];
+  for (const [token, issue] of items) {
+    const syntax = placeholderSyntax(token);
+    // In an ICU message, braces are ICU's: "other{{count}}" is a branch
+    // holding the argument, not i18next (an i18next "{{name}}" message never
+    // parses as ICU).
+    if ((syntax === 'brace' || syntax === 'i18next') && icuSource) continue;
+    if (syntax === 'markup' && markupReported) continue;
+    out.push({ syntax, token, issue });
+  }
+  return out;
+}
+
+// ── A sentence break inserted beside a placeholder ──────────────────────────
+//
+// "Take this medicine at {time}." came back as "… sina. {time}." — a sentence
+// end before the placeholder, so the app shows the time as a sentence of its
+// own — and the gate and verify both passed it: every placeholder was there
+// (Round 14, hospital persona; a medical string). The rule below is
+// structural and narrow: it does not judge fluency, only a sentence boundary
+// the TRANSLATION puts right beside a placeholder where the SOURCE has none.
+
+/**
+ * Sentence-final marks, across scripts: Latin/Cyrillic/Greek . ! ? (Greek's
+ * question mark ";" is left out: it is also the semicolon), CJK 。！？ and the
+ * halfwidth/fullwidth forms, Devanagari/Bengali danda । ॥, Arabic ؟ and the
+ * Urdu full stop ۔, Ethiopic ። ፧, Armenian ։, Myanmar ။, Khmer ។ ៕,
+ * Canadian Syllabics ᙮ (Cree, Inuktitut), and the doubled marks ‼ ⁇ ⁈ ⁉ ‽.
+ * Script-agnostic: a mark from any script is a boundary in any text.
+ */
+const SENTENCE_MARKS = new Set([...'.!?。！？｡．।॥؟۔።፧։။។៕᙮‼⁇⁈⁉‽']);
+
+/** Closing quotes and brackets that may stand between a mark and what follows it. */
+const CLOSERS = new Set([...'"\'”’»›」』）)］]']);
+
+/**
+ * Clause punctuation: where the SOURCE already sets a placeholder apart
+ * ("… corpora; {spec}.", "Error: {message}", "— {noncomp}, and …"), a
+ * translation may make that clause a sentence of its own.
+ */
+const CLAUSE_MARKS = new Set([...';:,—–·|(（[［、，；：']);
+
+/** Whitespace (JavaScript's \s includes no-break and narrow no-break space). */
+const SPACE = /\s/;
+
+/**
+ * Placeholder spans with their positions: i18next / Hugo / Handlebars
+ * `{{…}}`, a single-brace argument `{name}` / `{0}` / `{n, number}` (also
+ * Ruby's `%{name}`), and printf `%(name)s`, `%1$s`, `%s`, `%d`, `%@`.
+ * Nested ICU (plural/select) is not matched as one span — messages holding
+ * it are left out of this check (see placeholderSentenceBreaks).
+ */
+const PLACEHOLDER_SPAN = /\{\{[^{}]*\}\}|%?\{\s*(?:[A-Za-z_$][\w.$]*|\d+)\s*(?:,[^{}]*)?\}|%\(\w+\)[a-zA-Z]|%\d+\$[a-zA-Z@]|%[sd@]/g;
+
+/** An ICU plural/select/selectordinal argument — its branches are not placeholder spans. */
+const ICU_BRANCHING = /\{\s*[\w.$]+\s*,\s*(?:plural|select|selectordinal)\s*,/;
+
+function placeholderSpans(text) {
+  const spans = [];
+  for (const m of text.matchAll(PLACEHOLDER_SPAN)) {
+    spans.push({ token: m[0].replace(/\s+/g, ''), shown: m[0], start: m.index, end: m.index + m[0].length });
+  }
+  return spans;
+}
+
+/**
+ * Is text[i] a sentence boundary? Any mark in SENTENCE_MARKS but a full stop
+ * is one. A full stop "." is one unless it is:
+ *   - part of an ellipsis ("..." — a pause, not an end),
+ *   - followed directly by a letter, digit or "_" ("3.5", "{host}.com", "e.g"),
+ *   - the stop of a one-letter word — an initial or an abbreviation
+ *     ("M. {name}", "p. {page}", "z. B. {x}"),
+ *   - the stop of a short capitalised word (two or three letters) — an
+ *     abbreviation before a number or name ("Nr. {id}", "Dr. {name}",
+ *     "Tel. {phone}", "Ca. {count}").
+ * A stop right after a placeholder ("{time}.", "%(dose)s.") is never read as
+ * an abbreviation: `afterPlaceholder` says the stop follows one.
+ *
+ * @param {string} text
+ * @param {number} i
+ * @param {boolean} [afterPlaceholder]
+ * @returns {boolean}
+ */
+function isSentenceBoundaryAt(text, i, afterPlaceholder = false) {
+  const ch = text[i];
+  if (!SENTENCE_MARKS.has(ch)) return false;
+  if (ch !== '.') return true;
+  if (text[i - 1] === '.' || text[i + 1] === '.') return false;
+  const next = text[i + 1];
+  if (next !== undefined && /[\p{L}\p{N}_]/u.test(next)) return false;
+  if (afterPlaceholder) return true;
+  const word = /(\p{L}+)$/u.exec(text.slice(0, i))?.[1] || '';
+  if (word.length === 1) return false;
+  if (word.length <= 3 && /^\p{Lu}/u.test(word)) return false;
+  return true;
+}
+
+/**
+ * What stands on one side of a placeholder span, skipping spaces (and, on
+ * the left, closing quotes/brackets): the message's edge, a sentence
+ * boundary, clause punctuation, or a word (anything else, another
+ * placeholder included).
+ *
+ * @returns {{ kind: 'edge'|'boundary'|'clause'|'word', at: number }}
+ */
+function sideOf(text, span, side, placeholderEnds = new Set()) {
+  if (side === 'left') {
+    let j = span.start - 1;
+    while (j >= 0 && (SPACE.test(text[j]) || CLOSERS.has(text[j]))) j--;
+    if (j < 0) return { kind: 'edge', at: -1 };
+    if (isSentenceBoundaryAt(text, j, placeholderEnds.has(j))) return { kind: 'boundary', at: j };
+    if (SENTENCE_MARKS.has(text[j]) || CLAUSE_MARKS.has(text[j])) return { kind: 'clause', at: j };
+    return { kind: 'word', at: j };
+  }
+  let k = span.end;
+  while (k < text.length && SPACE.test(text[k])) k++;
+  if (k >= text.length) return { kind: 'edge', at: k };
+  if (isSentenceBoundaryAt(text, k, k === span.end || /^\s*$/.test(text.slice(span.end, k)))) return { kind: 'boundary', at: k };
+  if (SENTENCE_MARKS.has(text[k]) || CLAUSE_MARKS.has(text[k]) || CLOSERS.has(text[k])) return { kind: 'clause', at: k };
+  return { kind: 'word', at: k };
+}
+
+/**
+ * The sentence breaks a translation INSERTED beside a placeholder, leaving
+ * it standing as a sentence of its own — the one rule the sync gate refuses
+ * by (lib/validate.js) and verify flags by (lib/verify.js):
+ *   { token: '{time}', side: 'before', issue: 'a sentence break was inserted before {time} ("medisina. {time}.") — {time} now stands as a sentence of its own; in the source it is part of one' }
+ *
+ * A finding needs:
+ *   1. a placeholder the translation shares with the source, with a
+ *      sentence boundary right BEFORE or right AFTER it in the translation
+ *      ("medisina. {time}", "{time}. Kisik") — text before it, so a mark
+ *      that opens nothing is not one;
+ *   2. nothing of its sentence around it: on each side only a boundary or
+ *      the message's edge (". {time}." — the time shown as a sentence);
+ *   3. in the source, that placeholder inside a sentence: a word on one
+ *      side, not a boundary, clause punctuation or the edge
+ *      ("Take this medicine at {time}.").
+ * Narrow on purpose (Round 14 measurement, lib/placeholders.js tests): a
+ * placeholder that only starts or ends a sentence in the translation is
+ * word order, not damage — "Open an issue on {github} or …" is rightly
+ * "{github}에 이슈를 열거나 …" in Korean, and "Shipped by {carrier} on
+ * {date}." rightly "Expédié le {date} par {carrier}." Left out: ICU
+ * plural/select messages (the ICU check judges their branches).
+ *
+ * @param {string} sourceValue
+ * @param {string} targetValue
+ * @returns {Array<{ token: string, side: 'before'|'after', issue: string }>}
+ */
+function placeholderSentenceBreaks(sourceValue, targetValue) {
+  if (typeof sourceValue !== 'string' || typeof targetValue !== 'string') return [];
+  if (sourceValue === targetValue) return [];
+  if (ICU_BRANCHING.test(sourceValue) || ICU_BRANCHING.test(targetValue)) return [];
+  const targetSpans = placeholderSpans(targetValue);
+  if (targetSpans.length === 0) return [];
+  const sourceSpans = placeholderSpans(sourceValue);
+
+  // In the source the placeholder is part of a sentence: a word beside it.
+  const inSentence = (token) => sourceSpans.some(s => s.token === token
+    && (sideOf(sourceValue, s, 'left').kind === 'word' || sideOf(sourceValue, s, 'right').kind === 'word'));
+
+  const ends = new Set(targetSpans.map(t => t.end));
+  const found = [];
+  const seen = new Set();
+  for (const s of targetSpans) {
+    if (seen.has(s.token) || !inSentence(s.token)) continue; // a missing/extra placeholder is placeholderChanges' finding
+    const left = sideOf(targetValue, s, 'left', ends);
+    const right = sideOf(targetValue, s, 'right');
+    const alone = (left.kind === 'boundary' || left.kind === 'edge') && (right.kind === 'boundary' || right.kind === 'edge');
+    if (!alone || (left.kind !== 'boundary' && right.kind !== 'boundary')) continue;
+    // A boundary before it must close some text; one after it at the very
+    // end of the message, with nothing before, is the message's own end.
+    if (left.kind === 'boundary' && !/[\p{L}\p{N}]/u.test(targetValue.slice(0, left.at))) continue;
+    seen.add(s.token);
+    const side = left.kind === 'boundary' ? 'before' : 'after';
+    // The word before the mark, the placeholder, the mark after it: "medisina. {time}."
+    const from = left.kind === 'boundary' ? Math.max(0, targetValue.slice(0, left.at).search(/\S+\s*$/)) : s.start;
+    const to = right.kind === 'boundary' ? right.at + 1 : s.end;
+    const shown = targetValue.slice(from, to).trim();
+    found.push({
+      token: s.token,
+      side,
+      issue: `a sentence break was inserted ${side} ${s.shown} ("${shown}") — ${s.shown} now stands as a sentence of its own; in the source it is part of one`,
+    });
+  }
+  return found;
+}
+
+/** Sentence boundaries in a text: runs of marks ("?!", "。") count once. */
+function countSentenceBoundaries(text) {
+  const ends = new Set(placeholderSpans(text).map(s => s.end));
+  let n = 0;
+  for (let i = 0; i < text.length; i++) {
+    if (!SENTENCE_MARKS.has(text[i])) continue;
+    let j = i;
+    while (j + 1 < text.length && SENTENCE_MARKS.has(text[j + 1])) j++;
+    for (let k = i; k <= j; k++) {
+      if (isSentenceBoundaryAt(text, k, ends.has(k))) { n++; break; }
+    }
+    i = j;
+  }
+  return n;
+}
+
+export {
+  extractPlaceholders,
+  comparePlaceholders,
+  placeholderSyntax,
+  showPlaceholder,
+  parsesAsICU,
+  placeholderChanges,
+  placeholderSentenceBreaks,
+  countSentenceBoundaries,
+  PLACEHOLDER_SYNTAX_NAMES,
+};

package/lib/plugins.js

@@ -148,6 +148,9 @@ function validateManifest(manifest) {
   if (manifest.type === 'api' && !manifest.endpoint) {
     errors.push('API plugins must include an "endpoint" URL');
   }
+  if (manifest.acceptsInstructions !== undefined && typeof manifest.acceptsInstructions !== 'boolean') {
+    errors.push('"acceptsInstructions" must be true or false');
+  }
 
   // Benchmarks: validate structure if present
   if (manifest.benchmarks) {
@@ -375,6 +378,11 @@ function resolvePluginForPair(plugins, pairConfig) {
 
     // Plugin-specific fields
     endpoint: plugin.endpoint || pairConfig.endpoint || null,
+    // The endpoint's declared instruction capability: the pair's own
+    // statement wins, then the manifest's; null = unknown.
+    acceptsInstructions: typeof pairConfig.acceptsInstructions === 'boolean'
+      ? pairConfig.acceptsInstructions
+      : (typeof plugin.acceptsInstructions === 'boolean' ? plugin.acceptsInstructions : null),
     pluginName: plugin.name,
     pluginVersion: plugin.version,
     pluginDir: plugin._pluginDir,

package/lib/plural-gap-redo.js

@@ -0,0 +1,109 @@
+/**
+ * plural-gap-redo.js — a plural message left incomplete is asked again by a
+ * setup that has not tried it yet.
+ *
+ * THE FINDING (Round 13, Django persona): a local model left a Russian
+ * plural entry without its few/many forms. Sync wrote the entry with the
+ * "other" form standing in and marked it (`# champollion:` in a gettext
+ * catalog; a missing ICU branch in JSON), and every later sync treated it as
+ * done — including CI's, which runs a stronger hosted model. Every sync then
+ * exited 2 until a person re-ran with a key or wrote the forms by hand, and
+ * the CI workflow had no way to do either.
+ *
+ * THE RULE. A marked gap is not a translation. A sync asks for it again —
+ * from the model, not the cache, which holds the incomplete answer — when it
+ * runs a setup (method, model, register, coaching: the cache key) that has
+ * not answered this message's current text yet:
+ *   - the setup that wrote it is the one the lock records for the key
+ *     (`by`), while the record still matches the value on disk (a person who
+ *     edited the entry is never overruled);
+ *   - the setups that answered it without the forms before are kept in the
+ *     lock (`gaps`: source hash + method keys), so a local run and a CI run
+ *     do not take turns paying for the same incomplete answer;
+ *   - the pair's own fallback is part of the setup: a gap it left is not
+ *     re-asked by the same configuration.
+ * `--redo gaps` asks for every gap on disk, whoever left it (an explicit
+ * redo). A setup that also leaves the forms out leaves the gap marked, as
+ * before; the run exits 2.
+ *
+ * Shared by the sync and its cost estimate, so the estimate prices exactly
+ * what is asked again.
+ */
+
+import { pluralGapsInFile } from './verify.js';
+import { tmMethodKey } from './tm.js';
+import { decodeWritten, shortSourceHash, valueHash } from './locale-state.js';
+
+/**
+ * The setups a gap record says answered this text without the forms.
+ *
+ * @param {object|undefined} record - LockState.of(code).gaps[lockKey]
+ * @param {string} sourceValue
+ * @returns {string[]}
+ */
+export function gapTriedBy(record, sourceValue) {
+  if (!record || typeof record !== 'object' || typeof sourceValue !== 'string') return [];
+  if (record.source !== shortSourceHash(sourceValue) || !Array.isArray(record.methods)) return [];
+  return record.methods;
+}
+
+/**
+ * Remember that `methodKey` answered this key's current text without a form
+ * the language uses for ordinary counts.
+ *
+ * @param {object} localeState - LockState.of(code)
+ * @param {string} lockKey
+ * @param {string} sourceValue
+ * @param {string} methodKey
+ */
+export function recordGap(localeState, lockKey, sourceValue, methodKey) {
+  if (typeof sourceValue !== 'string' || !methodKey) return;
+  localeState.gaps = localeState.gaps || {};
+  const known = gapTriedBy(localeState.gaps[lockKey], sourceValue);
+  localeState.gaps[lockKey] = { source: shortSourceHash(sourceValue), methods: [...new Set([...known, methodKey])] };
+}
+
+/**
+ * Plural messages of one target file that this run asks for again.
+ *
+ * @param {object} p
+ * @param {object} p.file - layout file ({ path, format, rel })
+ * @param {object} p.expected - The target's expected map (source text per target key)
+ * @param {object} p.targetFlat - Values on disk
+ * @param {string} p.locale
+ * @param {object} p.localeState - LockState.peek(code)
+ * @param {(key: string) => string} p.lockKeyOf
+ * @param {object} p.pairConfig
+ * @param {boolean} [p.all] - `--redo gaps`: every gap on disk
+ * @returns {{ keys: string[], gaps: Array<{ key: string, missing: string[], marked: boolean }>,
+ *   by: Object<string, string|null>, missing: Object<string, string[]> }}
+ *   keys — target keys to ask again; gaps — every gap in the file; by — the
+ *   setup that left each asked key (null: no record); missing — its forms
+ */
+export function planGapRedo({ file, expected, targetFlat, locale, localeState, lockKeyOf, pairConfig, all = false }) {
+  const out = { keys: [], gaps: [], by: {}, missing: {} };
+  if (!file || !targetFlat) return out;
+  try { out.gaps = pluralGapsInFile({ file, expected, targetFlat, locale }); } catch { out.gaps = []; }
+  if (out.gaps.length === 0) return out;
+  const current = tmMethodKey(pairConfig);
+  const fallback = pairConfig.fallback ? tmMethodKey(pairConfig.fallback) : null;
+  for (const g of out.gaps) {
+    const src = expected[g.key];
+    const value = targetFlat[g.key];
+    if (typeof src !== 'string' || typeof value !== 'string') continue;
+    const lk = lockKeyOf(g.key);
+    // Who wrote what is on disk — only while the record still describes it.
+    const record = decodeWritten(localeState?.written?.[lk]);
+    const writer = record && record.source === shortSourceHash(src) && record.value === valueHash(value)
+      ? (localeState?.by?.[lk] || null) : null;
+    if (!all) {
+      if (!writer) continue;                                  // unknown, or a person's edit: never automatic
+      if (writer === current || writer === fallback) continue; // this setup left it
+      if (gapTriedBy(localeState?.gaps?.[lk], src).includes(current)) continue; // tried already
+    }
+    out.keys.push(g.key);
+    out.by[g.key] = writer;
+    out.missing[g.key] = g.missing;
+  }
+  return out;
+}