chaffjs 0.17.0 → 0.18.0

package/src/compare/match.ts

@@ -0,0 +1,105 @@
+import type { Atom } from "./atom.ts";
+
+/** The same fact written another way: 1,000円 and 1000円, 2026年4月1日 and 2026/4/1, a reworded heading. */
+export type Reformed = { readonly before: Atom; readonly after: Atom };
+
+export type Comparison = {
+  /** In the first document and not in the second. */
+  readonly dropped: readonly Atom[];
+  /** In the second document and not in the first. */
+  readonly added: readonly Atom[];
+  readonly reformed: readonly Reformed[];
+};
+
+const identity = (atom: Atom): string => `${atom.kind}\u0000${atom.key}`;
+
+const groupBy = <T>(items: readonly T[], keyOf: (item: T) => string): Map<string, T[]> => {
+  const groups = new Map<string, T[]>();
+  items.forEach((item) => {
+    const key = keyOf(item);
+    const group = groups.get(key);
+    if (group === undefined) groups.set(key, [item]);
+    else group.push(item);
+  });
+  return groups;
+};
+
+/** Takes and removes the first item the queue for key holds. */
+type Queues<T> = { readonly groups: Map<string, T[]>; readonly taken: Map<string, number> };
+
+const queuesOf = <T>(items: readonly T[], keyOf: (item: T) => string): Queues<T> => ({ groups: groupBy(items, keyOf), taken: new Map() });
+
+/** The next item not yet taken under key. A cursor, not shift(): shift moves every item left, which is quadratic over a long queue. */
+const takeFrom = <T>(queues: Queues<T>, key: string): T | undefined => {
+  const next = queues.taken.get(key) ?? 0;
+  const item = queues.groups.get(key)?.[next];
+  if (item !== undefined) queues.taken.set(key, next + 1);
+  return item;
+};
+
+/** Atoms of one fact: written the same way they need no report; written differently they pair as reformed; the rest are surplus. */
+const matchGroup = (before: readonly Atom[], after: readonly Atom[]): Comparison => {
+  const afterByText = queuesOf(after, (atom) => atom.text);
+  const sameText = new Set<Atom>();
+  const unmatchedBefore = before.filter((atom) => {
+    const twin = takeFrom(afterByText, atom.text);
+    if (twin !== undefined) sameText.add(twin);
+    return twin === undefined;
+  });
+  const unmatchedAfter = after.filter((atom) => !sameText.has(atom));
+  const pairs = Math.min(unmatchedBefore.length, unmatchedAfter.length);
+  return {
+    reformed: unmatchedBefore.slice(0, pairs).map((atom, index) => ({ before: atom, after: unmatchedAfter[index] ?? atom })),
+    dropped: unmatchedBefore.slice(pairs),
+    added: unmatchedAfter.slice(pairs),
+  };
+};
+
+const unitQueue = (value: string | undefined, unitless: boolean): string => `${value ?? ""}\u0000${unitless ? "bare" : "unit"}`;
+
+/**
+ * A number one side wrote with a unit the language package read (30 ＧＢ) and the other wrote so the unit was not read
+ * (30GB) is one number in another form. Two different units (5 件, 5 人) stay a dropped and an added fact.
+ */
+const pairUnits = (comparison: Comparison): Comparison => {
+  const addedNumbers = queuesOf(
+    comparison.added.filter((atom) => atom.kind === "number"),
+    (atom) => unitQueue(atom.value, atom.unitless === true),
+  );
+  const paired = new Set<Atom>();
+  const reformed = comparison.dropped.flatMap((atom): Reformed[] => {
+    // Two unitless numbers of one value share a key and were already paired; only the other side's kind can match.
+    const found = atom.kind === "number" ? takeFrom(addedNumbers, unitQueue(atom.value, atom.unitless !== true)) : undefined;
+    if (found === undefined) return [];
+    paired.add(atom).add(found);
+    return [{ before: atom, after: found }];
+  });
+  return {
+    dropped: comparison.dropped.filter((atom) => !paired.has(atom)),
+    added: comparison.added.filter((atom) => !paired.has(atom)),
+    reformed: [...comparison.reformed, ...reformed],
+  };
+};
+
+const byLine = (left: Atom, right: Atom): number => left.line - right.line;
+
+/**
+ * The facts of two documents compared as multisets: where a fact is written does not matter, how many times it is
+ * stated does. Pure: the same atoms always give the same comparison.
+ */
+export const compareAtoms = (before: readonly Atom[], after: readonly Atom[]): Comparison => {
+  const beforeGroups = groupBy(before, identity);
+  const afterGroups = groupBy(after, identity);
+  const keys = [...new Set([...beforeGroups.keys(), ...afterGroups.keys()])];
+  const groups = keys.map((key) => matchGroup(beforeGroups.get(key) ?? [], afterGroups.get(key) ?? []));
+  const paired = pairUnits({
+    dropped: groups.flatMap((group) => group.dropped),
+    added: groups.flatMap((group) => group.added),
+    reformed: groups.flatMap((group) => group.reformed),
+  });
+  return {
+    dropped: paired.dropped.toSorted(byLine),
+    added: paired.added.toSorted(byLine),
+    reformed: paired.reformed.toSorted((left, right) => byLine(left.before, right.before)),
+  };
+};

package/src/compare/outcome.ts

@@ -0,0 +1,70 @@
+import { type Atom, type AtomKind, type Extraction, type Unread } from "./atom.ts";
+import { compareAtoms, type Reformed } from "./match.ts";
+import { unwrittenNames } from "./written-names.ts";
+
+/** One document as compared: where it is, how many facts of each kind it states, and what could not be read. */
+export type Side = {
+  readonly path: string;
+  readonly counts: Readonly<Record<AtomKind, number>>;
+  readonly unread: readonly Unread[];
+};
+
+/** A dropped or added fact, and whether the command line allowed its kind to change. */
+export type Change = Atom & { readonly allowed: boolean };
+
+export type Outcome = {
+  readonly before: Side;
+  readonly after: Side;
+  readonly dropped: readonly Change[];
+  readonly added: readonly Change[];
+  readonly reformed: readonly Reformed[];
+  /** No fact dropped or added, other than of a kind allowed to be. */
+  readonly ok: boolean;
+};
+
+/** Kinds whose facts may be dropped or added on purpose (--allow-dropped, --allow-added). */
+export type Allowed = { readonly dropped: ReadonlySet<AtomKind>; readonly added: ReadonlySet<AtomKind> };
+
+const NOTHING_ALLOWED: Allowed = { dropped: new Set(), added: new Set() };
+
+const NO_FACTS: Readonly<Record<AtomKind, number>> = {
+  number: 0,
+  date: 0,
+  time: 0,
+  url: 0,
+  code: 0,
+  name: 0,
+  quote: 0,
+  heading: 0,
+  reference: 0,
+  footnote: 0,
+};
+
+const countsOf = (atoms: readonly Atom[]): Record<AtomKind, number> => {
+  const counts = { ...NO_FACTS };
+  atoms.forEach((atom) => {
+    counts[atom.kind] += 1;
+  });
+  return counts;
+};
+
+const sideOf = (path: string, extraction: Extraction): Side => ({ path, counts: countsOf(extraction.atoms), unread: extraction.unread });
+
+const marked = (atoms: readonly Atom[], allowed: ReadonlySet<AtomKind>): Change[] => atoms.map((atom) => ({ ...atom, allowed: allowed.has(atom.kind) }));
+
+export type Compared = { readonly path: string; readonly extraction: Extraction };
+
+/** What `chaff compare` reports: the comparison of two documents' facts, with the allowed kinds set aside. */
+export const outcomeOf = (before: Compared, after: Compared, allowed: Allowed = NOTHING_ALLOWED): Outcome => {
+  const comparison = compareAtoms(before.extraction.atoms, after.extraction.atoms);
+  const dropped = marked(unwrittenNames(comparison.dropped, before.extraction.atoms, after.extraction.nameText), allowed.dropped);
+  const added = marked(unwrittenNames(comparison.added, after.extraction.atoms, before.extraction.nameText), allowed.added);
+  return {
+    before: sideOf(before.path, before.extraction),
+    after: sideOf(after.path, after.extraction),
+    dropped,
+    added,
+    reformed: comparison.reformed,
+    ok: [...dropped, ...added].every((change) => change.allowed),
+  };
+};

package/src/compare/proper-nouns.ts

@@ -0,0 +1,65 @@
+import type { ProseDocument, Span, Token } from "../plugin.ts";
+import type { Atom } from "./atom.ts";
+import { overlapsAny, type SpanIndex } from "./spans.ts";
+
+const PROPER_NOUN = "PROPN";
+
+/** Two proper nouns with at most one space between them are one name: New York, 日本 銀行. */
+const MAX_GAP = 1;
+
+type Run = Span & { readonly words: number };
+
+const joins = (run: Run | undefined, token: Token, text: string): run is Run =>
+  run !== undefined && token.span.start - run.end <= MAX_GAP && text.slice(run.end, token.span.start).trim() === "";
+
+/** Adjacent proper nouns, each run one name. */
+const runsOf = (tokens: readonly Token[], text: string): Run[] =>
+  tokens.reduce<Run[]>((runs, token) => {
+    if (token.pos !== PROPER_NOUN) return runs;
+    const last = runs.at(-1);
+    if (joins(last, token, text)) runs[runs.length - 1] = { start: last.start, end: token.span.end, words: last.words + 1 };
+    else runs.push({ start: token.span.start, end: token.span.end, words: 1 });
+    return runs;
+  }, []);
+
+/** Words written in lower case. A capitalised word the same document also writes this way is a common word that began a sentence. */
+const LOWER_CASE_WORD = /(?<![\p{L}\p{M}])\p{Ll}[\p{L}\p{M}]*/gu;
+
+const lowerCaseWords = (source: string): ReadonlySet<string> => new Set(source.match(LOWER_CASE_WORD) ?? []);
+
+const withLowerFirst = (word: string): string => `${word.charAt(0).toLowerCase()}${word.slice(1)}`;
+
+/** A one-word name the document also writes in lower case ("Scammers" opening a sentence, "scammers" elsewhere). */
+const isCommonWord = (run: Run, written: string, lowerCase: ReadonlySet<string>): boolean => {
+  const lowered = withLowerFirst(written);
+  return run.words === 1 && lowered !== written && lowerCase.has(lowered);
+};
+
+export type NounInput = {
+  readonly doc: ProseDocument;
+  /** Facts other readers took and what is not prose: a proper noun inside one is read as that, not again here. */
+  readonly taken: SpanIndex;
+  readonly lineOf: (offset: number) => number;
+};
+
+/** A name as one spelling: full-width letters and line breaks inside a name do not make another name. */
+export const nameKey = (written: string): string => written.normalize("NFKC").replace(/\s+/gu, " ");
+
+/** Whether the language package tagged parts of speech: without them, a document with sentences has no proper nouns to read. */
+export const readsProperNouns = (doc: ProseDocument): boolean => doc.sentences.length === 0 || doc.sentences.some((sentence) => sentence.tokens !== undefined);
+
+/**
+ * Proper nouns as the part-of-speech tagger reads them in sentences. Headings are left out: they are compared as
+ * headings, may be reworded, and a title-case heading makes every word look like a name.
+ */
+export const properNouns = (input: NounInput): Atom[] => {
+  const { source } = input.doc;
+  const lowerCase = lowerCaseWords(source);
+  return input.doc.sentences.flatMap((sentence) =>
+    runsOf(sentence.tokens ?? [], source).flatMap((run): Atom[] => {
+      const text = source.slice(run.start, run.end);
+      if (overlapsAny(input.taken, run) || isCommonWord(run, text, lowerCase)) return [];
+      return [{ kind: "name", key: nameKey(text), text, line: input.lineOf(run.start) }];
+    }),
+  );
+};

package/src/compare/render.ts

@@ -0,0 +1,102 @@
+import { ATOM_KINDS, type Atom, type AtomKind } from "./atom.ts";
+import type { Reformed } from "./match.ts";
+import type { Change, Outcome, Side } from "./outcome.ts";
+import type { CompareText } from "./text.ts";
+
+/** How much of a fact is shown: a code block is shown by its first line. */
+const SHOWN_WIDTH = 60;
+
+/** One line of a fact as written: a code block or a wrapped quotation is shown on one line, cut at SHOWN_WIDTH. */
+const shown = (text: string): string => {
+  const oneLine = text.replace(/\s+/gu, " ").trim();
+  return oneLine.length > SHOWN_WIDTH ? `${oneLine.slice(0, SHOWN_WIDTH)}…` : oneLine;
+};
+
+const totalOf = (side: Side): number => ATOM_KINDS.reduce((sum, kind) => sum + side.counts[kind], 0);
+
+/** The count line. Every kind is listed, zeros too: "nothing dropped" reads as N facts checked, never as nothing looked at. */
+const checkedLine = (outcome: Outcome, text: CompareText): string => {
+  const perKind = ATOM_KINDS.map((kind) => `${text.kinds[kind]} ${String(outcome.before.counts[kind])}→${String(outcome.after.counts[kind])}`);
+  return text.checked(totalOf(outcome.before), totalOf(outcome.after), perKind.join(text.separator));
+};
+
+const failing = (changes: readonly Change[]): number => changes.filter((change) => !change.allowed).length;
+
+const verdict = (outcome: Outcome, text: CompareText): string => (outcome.ok ? text.clean : text.failed(failing(outcome.dropped), failing(outcome.added)));
+
+const kindColumn = (kind: AtomKind, text: CompareText): string => `${text.kind[kind]}: `;
+
+const changeLine = (change: Change, path: string, flag: string, text: CompareText): string =>
+  `  ${kindColumn(change.kind, text)}${shown(change.text)}  (${path}:${String(change.line)})${change.allowed ? text.allowed(flag) : ""}`;
+
+const reformedLine = (pair: Reformed, text: CompareText): string =>
+  `  ${kindColumn(pair.before.kind, text)}${shown(pair.before.text)} → ${shown(pair.after.text)}  (${text.lines(pair.before.line, pair.after.line)})`;
+
+const block = (heading: string, lines: readonly string[]): string[] => (lines.length === 0 ? [] : ["", heading, ...lines]);
+
+const unreadBlock = (outcome: Outcome, text: CompareText): string[] => {
+  const sides = [outcome.before, outcome.after];
+  const lines = sides.flatMap((side) => side.unread.map((unread) => `  ${side.path}: ${text.unread(unread.kind, unread.reason)}`));
+  return block(text.unreadHeading, lines);
+};
+
+/** For a person: what was dropped and added, with the line in each file, then what only changed its form. */
+export const renderFriendly = (outcome: Outcome, text: CompareText): string => {
+  const { before, after } = outcome;
+  return [
+    `${before.path} → ${after.path}`,
+    ...block(
+      text.droppedHeading(outcome.dropped.length, before.path, after.path),
+      outcome.dropped.map((change) => changeLine(change, before.path, "--allow-dropped", text)),
+    ),
+    ...block(
+      text.addedHeading(outcome.added.length, after.path),
+      outcome.added.map((change) => changeLine(change, after.path, "--allow-added", text)),
+    ),
+    ...block(
+      text.reformedHeading(outcome.reformed.length),
+      outcome.reformed.map((pair) => reformedLine(pair, text)),
+    ),
+    ...unreadBlock(outcome, text),
+    "",
+    checkedLine(outcome, text),
+    verdict(outcome, text),
+  ].join("\n");
+};
+
+const compactChange = (change: Change, path: string, word: string): string =>
+  `${path}:${String(change.line)}: ${word} ${change.kind} ${shown(change.text)}${change.allowed ? " (allowed)" : ""}`;
+
+/** For an engineer and grep: one line per fact, the words dropped / added / reformed and the kind in English. */
+export const renderCompact = (outcome: Outcome, text: CompareText): string => {
+  const { before, after } = outcome;
+  return [
+    ...outcome.dropped.map((change) => compactChange(change, before.path, "dropped")),
+    ...outcome.added.map((change) => compactChange(change, after.path, "added")),
+    ...outcome.reformed.map(
+      (pair) =>
+        `${before.path}:${String(pair.before.line)} → ${after.path}:${String(pair.after.line)}: reformed ${pair.before.kind} ${shown(pair.before.text)} → ${shown(pair.after.text)}`,
+    ),
+    ...before.unread.map((unread) => `${before.path}: unread ${unread.kind} (${unread.reason})`),
+    ...after.unread.map((unread) => `${after.path}: unread ${unread.kind} (${unread.reason})`),
+    checkedLine(outcome, text),
+    text.tally(failing(outcome.dropped), failing(outcome.added), outcome.reformed.length),
+  ].join("\n");
+};
+
+const atomJson = (atom: Atom): Readonly<Record<string, string | number>> => ({ kind: atom.kind, key: atom.key, text: atom.text, line: atom.line });
+
+/** For an AI to act on: every dropped and added fact with its line, the reformed pairs, and the counts per kind. */
+export const renderJson = (outcome: Outcome): string =>
+  JSON.stringify(
+    {
+      ok: outcome.ok,
+      before: outcome.before,
+      after: outcome.after,
+      dropped: outcome.dropped.map((change) => ({ ...atomJson(change), allowed: change.allowed })),
+      added: outcome.added.map((change) => ({ ...atomJson(change), allowed: change.allowed })),
+      reformed: outcome.reformed.map((pair) => ({ kind: pair.before.kind, key: pair.before.key, before: atomJson(pair.before), after: atomJson(pair.after) })),
+    },
+    null,
+    2,
+  );

package/src/compare/spans.ts

@@ -0,0 +1,17 @@
+import type { Span } from "../plugin.ts";
+import { firstEndingAfter } from "../soft-break.ts";
+import { mergeSpans } from "../span-merge.ts";
+
+/** Spans made ready for lookups: sorted and merged, so each lookup is a binary search, not a scan of every span. */
+export type SpanIndex = readonly Span[];
+
+export const spanIndex = (spans: readonly Span[]): SpanIndex => mergeSpans(spans, false);
+
+/** Whether span shares at least one character with a span in the index. */
+export const overlapsAny = (index: SpanIndex, span: Span): boolean => {
+  const region = index[firstEndingAfter(index, span.start)];
+  return region !== undefined && region.start < span.end;
+};
+
+/** Whether the character at offset is inside a span in the index. */
+export const coversOffset = (index: SpanIndex, offset: number): boolean => overlapsAny(index, { start: offset, end: offset + 1 });

package/src/compare/text-atoms.ts

@@ -0,0 +1,95 @@
+import { QUOTATION_MARKS, quotedSpans } from "../quoted-span.ts";
+import type { Span } from "../plugin.ts";
+import type { Atom } from "./atom.ts";
+import { coversOffset, overlapsAny, spanIndex, type SpanIndex } from "./spans.ts";
+
+/** Where a reader looks, and how it says where it found something. */
+export type TextInput = { readonly text: string; readonly source: string; readonly lineOf: (offset: number) => number };
+
+/** A fact found in the text: where it is and its key. */
+export type Found = Span & { readonly key: string };
+
+export const atomsOf = (found: readonly Found[], kind: Atom["kind"], input: TextInput): Atom[] =>
+  found.map((fact) => ({ kind, key: fact.key, text: input.source.slice(fact.start, fact.end), line: input.lineOf(fact.start) }));
+
+const DIGIT = "[0-9０-９]";
+const DATE_SEPARATOR = "[-/.／．]";
+
+/** 2026/4/1, 2026-04-01, 2026.4.1: a year of four digits first, so the order is never in doubt. */
+const NUMERIC_DATE = new RegExp(`(?<![0-9０-９])(${DIGIT}{4})${DATE_SEPARATOR}(${DIGIT}{1,2})${DATE_SEPARATOR}(${DIGIT}{1,2})(?![0-9０-９])`, "gu");
+
+const MONTHS = 12;
+const LAST_DAY = 31;
+const PADDED = 2;
+
+const dateKey = (year: string, month: string, day: string): string | undefined => {
+  const [m, d] = [Number(month.normalize("NFKC")), Number(day.normalize("NFKC"))];
+  if (m < 1 || m > MONTHS || d < 1 || d > LAST_DAY) return undefined;
+  return [year.normalize("NFKC"), String(m).padStart(PADDED, "0"), String(d).padStart(PADDED, "0")].join("-");
+};
+
+/** Dates written in figures only, where the language's date reader did not already read a date. */
+export const numericDates = (text: string, taken: SpanIndex): Found[] =>
+  [...text.matchAll(NUMERIC_DATE)].flatMap((match) => {
+    const key = dateKey(match[1] ?? "", match[2] ?? "", match[3] ?? "");
+    const span = { start: match.index, end: match.index + match[0].length };
+    return key === undefined || overlapsAny(taken, span) ? [] : [{ ...span, key }];
+  });
+
+/** A footnote mark `[^1]`, or the footnote itself `[^1]:`. */
+const FOOTNOTE = /\[\^[^\]\s]+\]:?/gu;
+
+export const footnotes = (text: string): Found[] =>
+  [...text.matchAll(FOOTNOTE)].map((match) => ({ start: match.index, end: match.index + match[0].length, key: match[0] }));
+
+/** 1,200 or 12.5 or ３: the digits, a thousands separator only between groups of three, and a decimal part. */
+const NUMBER = /[0-9０-９]+(?:[,，][0-9０-９]{3})*(?:[.．][0-9０-９]+)?/gu;
+
+const valueOf = (written: string): string => String(Number(written.normalize("NFKC").replace(/,/gu, "")));
+
+/** Numbers no other reader took: written with no unit the language package knows, or with none at all. */
+export const bareNumbers = (input: TextInput, taken: SpanIndex): Atom[] =>
+  [...input.text.matchAll(NUMBER)].flatMap((match): Atom[] => {
+    if (coversOffset(taken, match.index)) return [];
+    const value = valueOf(match[0]);
+    return [{ kind: "number", key: `${value} `, text: match[0], line: input.lineOf(match.index), unitless: true, value }];
+  });
+
+/** A paragraph break: a quotation does not run over one, and reading paragraph by paragraph keeps each read short. */
+const PARAGRAPH_BREAK = /\n[ \t]*\n/gu;
+
+const paragraphsOf = (text: string): Span[] => {
+  const breaks = [...text.matchAll(PARAGRAPH_BREAK)].map((match) => match.index);
+  return [-1, ...breaks].map((from, index) => ({ start: from + 1, end: breaks[index] ?? text.length }));
+};
+
+/** What is quoted, as one spelling: line breaks and runs of spaces are how it was wrapped, not what it says. */
+const quoteKey = (inner: string): string => inner.normalize("NFKC").replace(/\s+/gu, " ").trim();
+
+/** Quoted strings: 「」『』, “” and "". The key is what is inside; the marks are how it is written. */
+export const quotations = (input: TextInput): Atom[] =>
+  paragraphsOf(input.text).flatMap((paragraph) =>
+    quotedSpans(input.text.slice(paragraph.start, paragraph.end), QUOTATION_MARKS).flatMap((inner) => {
+      const key = quoteKey(input.text.slice(paragraph.start + inner.start, paragraph.start + inner.end));
+      const span = { start: paragraph.start + inner.start - 1, end: paragraph.start + inner.end + 1 };
+      return key === "" ? [] : atomsOf([{ ...span, key }], "quote", input);
+    }),
+  );
+
+const occurrences = (text: string, name: string): Span[] => {
+  const found: Span[] = [];
+  for (let at = text.indexOf(name); at !== -1; at = text.indexOf(name, at + name.length)) found.push({ start: at, end: at + name.length });
+  return found;
+};
+
+/** The team's own names (chaff.yaml's `names:`), wherever they are written. Longer names first: "Acme Cloud" before "Acme". */
+export const teamNames = (text: string, names: readonly string[]): Found[] => {
+  const longestFirst = [...new Set(names.filter((name) => name !== ""))].toSorted((left, right) => right.length - left.length);
+  return longestFirst
+    .reduce<Found[]>((taken, name) => {
+      const index = spanIndex(taken);
+      const fresh = occurrences(text, name).filter((span) => !overlapsAny(index, span));
+      return [...taken, ...fresh.map((span) => ({ ...span, key: name }))];
+    }, [])
+    .toSorted((left, right) => left.start - right.start);
+};

package/src/compare/text.ts

@@ -0,0 +1,120 @@
+import type { Texts } from "../ui.ts";
+import type { AtomKind, UnreadReason } from "./atom.ts";
+import { counted } from "../render/plural.ts";
+
+export type CompareText = {
+  readonly usage: string;
+  readonly unknownKind: (kind: string, kinds: string) => string;
+  /** A kind counted (numbers 3→3) and a kind of one fact (number: 1,000円). */
+  readonly kinds: Readonly<Record<AtomKind, string>>;
+  readonly kind: Readonly<Record<AtomKind, string>>;
+  readonly unread: (kind: AtomKind, reason: UnreadReason) => string;
+  readonly unreadHeading: string;
+  readonly droppedHeading: (count: number, before: string, after: string) => string;
+  readonly addedHeading: (count: number, after: string) => string;
+  readonly reformedHeading: (count: number) => string;
+  readonly allowed: (flag: string) => string;
+  readonly lines: (before: number, after: number) => string;
+  readonly checked: (before: number, after: number, perKind: string) => string;
+  readonly separator: string;
+  readonly clean: string;
+  readonly failed: (dropped: number, added: number) => string;
+  /** The tally of the compact output. */
+  readonly tally: (dropped: number, added: number, reformed: number) => string;
+};
+
+const KINDS_JA: Readonly<Record<AtomKind, string>> = {
+  number: "数",
+  date: "日付",
+  time: "時刻",
+  url: "URL",
+  code: "コード",
+  name: "固有名詞",
+  quote: "引用",
+  heading: "見出し",
+  reference: "条項の参照",
+  footnote: "脚注",
+};
+
+const KINDS_EN: Readonly<Record<AtomKind, string>> = {
+  number: "numbers",
+  date: "dates",
+  time: "times",
+  url: "URLs",
+  code: "code",
+  name: "names",
+  quote: "quotations",
+  heading: "headings",
+  reference: "references",
+  footnote: "footnotes",
+};
+
+const KIND_EN: Readonly<Record<AtomKind, string>> = {
+  number: "number",
+  date: "date",
+  time: "time",
+  url: "URL",
+  code: "code",
+  name: "name",
+  quote: "quotation",
+  heading: "heading",
+  reference: "reference",
+  footnote: "footnote",
+};
+
+const UNREAD_JA: Readonly<Record<UnreadReason, (kind: AtomKind) => string>> = {
+  "no-structure": (kind) =>
+    kind === "number" ? "言語パッケージが単位を読まないので、数字だけを比べました" : "言語パッケージが文書の構造を読まないので、読んでいません",
+  "no-dates": () => "言語パッケージが日付を読まないので、数字だけで書いた日付（2026-04-01）だけを比べました",
+  "no-pos": () => "品詞を読む解析器が無いので、chaff.yaml の names: だけを比べました",
+  "plain-text": () => "Markdown ではないので、コードの記法がありません",
+};
+
+const UNREAD_EN: Readonly<Record<UnreadReason, (kind: AtomKind) => string>> = {
+  "no-structure": (kind) =>
+    kind === "number"
+      ? "the language package reads no units, so only the figures were compared"
+      : "the language package reads no document structure, so none were read",
+  "no-dates": () => "the language package reads no dates, so only dates written in figures (2026-04-01) were compared",
+  "no-pos": () => "there is no part-of-speech tagger, so only chaff.yaml's names: were compared",
+  "plain-text": () => "the document is not Markdown, so it has no code markup",
+};
+
+export const COMPARE_TEXT: Texts<CompareText> = {
+  ja: {
+    usage: "使い方: chaff compare <書き換える前> <書き換えた後> [--compact | --json] [--allow-dropped <種類>] [--allow-added <種類>] [--language ja|en|…]",
+    unknownKind: (kind, kinds) => `知らない種類です: ${kind}（使えるのは ${kinds}）`,
+    kinds: KINDS_JA,
+    kind: KINDS_JA,
+    unread: (kind, reason) => `${KINDS_JA[kind]}: ${UNREAD_JA[reason](kind)}`,
+    unreadHeading: "読めなかったもの",
+    droppedHeading: (count, before, after) => `✗ 落ちた事実 ${String(count)} 件（${before} にあって ${after} に無い）`,
+    addedHeading: (count, after) => `✗ 足された事実 ${String(count)} 件（${after} にだけある）`,
+    reformedHeading: (count) => `i 書き方だけ変わった事実 ${String(count)} 件`,
+    allowed: (flag) => `（${flag} で許可）`,
+    lines: (before, after) => `${String(before)} 行目 → ${String(after)} 行目`,
+    checked: (before, after, perKind) => `照合した事実 ${String(before)} 件 → ${String(after)} 件: ${perKind}`,
+    separator: "、",
+    clean: "落ちた事実も足された事実もありません",
+    failed: (dropped, added) => `落ちた事実 ${String(dropped)} 件、足された事実 ${String(added)} 件`,
+    tally: (dropped, added, reformed) => `落ちた ${String(dropped)} 件、足された ${String(added)} 件、書き方だけ ${String(reformed)} 件`,
+  },
+  en: {
+    usage: "usage: chaff compare <before> <after> [--compact | --json] [--allow-dropped <kind>] [--allow-added <kind>] [--language ja|en|…]",
+    unknownKind: (kind, kinds) => `Unknown kind: ${kind} (the kinds are ${kinds})`,
+    kinds: KINDS_EN,
+    kind: KIND_EN,
+    unread: (kind, reason) => `${KINDS_EN[kind]}: ${UNREAD_EN[reason](kind)}`,
+    unreadHeading: "Not fully read",
+    droppedHeading: (count, before, after) => `✗ ${counted(count, "fact")} dropped (in ${before}, not in ${after})`,
+    addedHeading: (count, after) => `✗ ${counted(count, "fact")} added (only in ${after})`,
+    reformedHeading: (count) => `i ${counted(count, "fact")} written another way`,
+    allowed: (flag) => ` (allowed by ${flag})`,
+    lines: (before, after) => `line ${String(before)} → line ${String(after)}`,
+    checked: (before, after, perKind) => `Facts checked: ${String(before)} → ${String(after)}: ${perKind}`,
+    separator: ", ",
+    clean: "No fact dropped or added",
+    failed: (dropped, added) => `${counted(dropped, "fact")} dropped, ${counted(added, "fact")} added`,
+    tally: (dropped, added, reformed) => `${String(dropped)} dropped, ${String(added)} added, ${String(reformed)} written another way`,
+  },
+};

package/src/compare/tree-atoms.ts

@@ -0,0 +1,80 @@
+import { preOrder } from "../tree-walk.ts";
+import type { Span, StructureNode } from "../plugin.ts";
+import type { Atom, AtomKind } from "./atom.ts";
+import { coversOffset, overlapsAny, type SpanIndex } from "./spans.ts";
+
+/** Facts the structure reader found, and where they are written, so no other reader reads the same digits again. */
+export type TreeFacts = { readonly atoms: readonly Atom[]; readonly taken: readonly Span[] };
+
+export type TreeInput = {
+  readonly tree: StructureNode;
+  readonly source: string;
+  /** Code and URLs: what the reader found there belongs to them. */
+  readonly blanked: SpanIndex;
+  /** Times of day: 3時30分 is one time, not the quantities 3 時 and 30 分. */
+  readonly times: SpanIndex;
+  readonly lineOf: (offset: number) => number;
+};
+
+/** Numbered units that are the document's own structure. Items (a list's 1. 2. 3.) are left out: prose may replace a list. */
+const NUMBERED: ReadonlySet<string> = new Set(["chapter", "article"]);
+
+const attr = (node: StructureNode, name: string): string => String(node.attrs[name] ?? "");
+
+/** The unit as one spelling: ＧＢ and GB, ％ and % are one unit. */
+const unitKey = (unit: string): string => unit.normalize("NFKC");
+
+const quantityAtom = (node: StructureNode, text: string, line: number): Atom => {
+  const value = attr(node, "value");
+  return { kind: "number", key: `${value} ${unitKey(attr(node, "unit"))}`, text, line, value };
+};
+
+/** The same section in another wording: Section 4.2 and § 4.2 point at one place; Article 4.2 does not. */
+const referenceKey = (node: StructureNode): string => {
+  const numbering = attr(node, "numbering") || attr(node, "unitWord");
+  const document = attr(node, "document");
+  return [numbering, attr(node, "target"), document === "" ? "" : `@${document}`].filter((part) => part !== "").join(" ");
+};
+
+const leafAtom = (node: StructureNode, kind: AtomKind, text: string, line: number): Atom => {
+  if (kind === "number") return quantityAtom(node, text, line);
+  return { kind, key: kind === "date" ? attr(node, "value") : referenceKey(node), text, line };
+};
+
+const LEAF_KINDS: Readonly<Partial<Record<string, AtomKind>>> = { quantity: "number", date: "date", reference: "reference" };
+
+/** The label's place on its line (第3条 after "## "), or none when the line does not spell it. */
+const labelSpan = (node: StructureNode, source: string): Span[] => {
+  const label = attr(node, "label");
+  const lineEnd = source.indexOf("\n", node.span.start);
+  const at = source.indexOf(label, node.span.start);
+  return label === "" || at === -1 || (lineEnd !== -1 && at > lineEnd) ? [] : [{ start: at, end: at + label.length }];
+};
+
+const numberedFacts = (node: StructureNode, input: TreeInput): TreeFacts => {
+  const heading = attr(node, "heading");
+  const text = [attr(node, "label"), heading].filter((part) => part !== "").join(" ");
+  return { atoms: [{ kind: "heading", key: `${node.kind} ${node.address}`, text, line: node.line }], taken: labelSpan(node, input.source) };
+};
+
+const nodeFacts = (node: StructureNode, input: TreeInput): TreeFacts => {
+  if (NUMBERED.has(node.kind)) return numberedFacts(node, input);
+  const kind = LEAF_KINDS[node.kind];
+  if (kind === undefined || coversOffset(input.blanked, node.span.start)) return { atoms: [], taken: [] };
+  if (kind === "number" && overlapsAny(input.times, node.span)) return { atoms: [], taken: [] };
+  const span = withUnitAfter(node, input.source);
+  const text = input.source.slice(span.start, span.end);
+  return { atoms: [leafAtom(node, kind, text, input.lineOf(span.start))], taken: [span] };
+};
+
+/** A quantity's span may stop before the unit written right after it (25 of 25%); the fact is shown with its unit. */
+const withUnitAfter = (node: StructureNode, source: string): Span => {
+  const unit = attr(node, "unit");
+  return unit !== "" && source.startsWith(unit, node.span.end) ? { start: node.span.start, end: node.span.end + unit.length } : node.span;
+};
+
+/** Numbers with their unit, dates, references to articles and sections, and the numbered articles themselves. */
+export const treeFacts = (input: TreeInput): TreeFacts => {
+  const facts = preOrder(input.tree).map((node) => nodeFacts(node, input));
+  return { atoms: facts.flatMap((fact) => fact.atoms), taken: facts.flatMap((fact) => fact.taken) };
+};