chaffjs 0.18.0 → 0.20.0

package/src/date-format.ts

@@ -0,0 +1,123 @@
+// 日付の書き方（2026-10-02、2026年10月2日、Oct 2, 2026）を読む。純粋な関数だけを置く。
+// 月の名前と元号は言語パッケージの語彙表が渡す。年・月・日のそろった日付だけを読む（年だけ、月日だけは書き方を比べられない）。
+
+/** 見つけた日付 1 つ。offset は渡した文字列の中の位置、style は書き方の名前。 */
+export type WrittenDate = { readonly offset: number; readonly written: string; readonly style: string };
+
+/** 書き方を読むのに要る語。months は月の名前（January、Jan）、eras は元号（令和）。 */
+export type DateWords = { readonly months: readonly string[]; readonly eras: readonly string[] };
+
+const DIGIT = "[0-9０-９]";
+/** 日付の前に来ない字。URL や版（v1.2.2026.10）の中の数を日付と読まない。 */
+const NOT_AFTER = "(?<![\\w/.\\-])";
+
+const escaped = (word: string): string => word.replace(/[.*+?^${}()|[\]\\]/gu, "\\$&");
+
+type Shape = readonly [string, RegExp];
+
+/** 日付の後ろに来ない字。時刻の付いた日時（2026-10-02T12:00）とファイルの名前（2026-10-02.pdf）は日付の書き方ではない。 */
+const NOT_BEFORE = "(?![\\w/／\\-]|\\.\\w)";
+
+const Y = `(?<y>${DIGIT}{4})`;
+const M = `(?<m>${DIGIT}{1,2})`;
+const D = `(?<d>${DIGIT}{1,2})`;
+
+/** 書き方の名前と、それを読む形。数字の組は月と日の範囲を後で確かめる。全角の数字も読む。 */
+const numericShapes = (): Shape[] => [
+  ["iso", new RegExp(`${NOT_AFTER}${Y}-${M}-${D}${NOT_BEFORE}`, "gu")],
+  ["year-slash", new RegExp(`${NOT_AFTER}${Y}[/／]${M}[/／]${D}${NOT_BEFORE}`, "gu")],
+  ["year-dot", new RegExp(`${NOT_AFTER}${Y}\\.${M}\\.${D}(?!\\.?[\\d０-９])`, "gu")],
+  ["slash-year-last", new RegExp(`${NOT_AFTER}${M}[/／]${D}[/／]${Y}${NOT_BEFORE}`, "gu")],
+  ["nen-gappi", new RegExp(`${Y}年\\s?${M}月\\s?${D}日`, "gu")],
+];
+
+const monthShapes = (months: string): Shape[] =>
+  months === ""
+    ? []
+    : [
+        ["month-day-year", new RegExp(`\\b(?:${months})\\.?\\s+(?<d>\\d{1,2})(?:st|nd|rd|th)?,?\\s+\\d{4}\\b`, "gu")],
+        ["day-month-year", new RegExp(`\\b(?<d>\\d{1,2})(?:st|nd|rd|th)?\\s+(?:${months})\\.?,?\\s+\\d{4}\\b`, "gu")],
+      ];
+
+const eraShapes = (eras: string): Shape[] => (eras === "" ? [] : [["era", new RegExp(`(?:${eras})\\s?(?:${DIGIT}{1,2}|元)年\\s?${M}月\\s?${D}日`, "gu")]]);
+
+/** 語彙表の語を使う形。月の名前は書いたとおりに照らす（小文字の may は月ではない）。 */
+const namedShapes = (words: DateWords): Shape[] => [...monthShapes(words.months.map(escaped).join("|")), ...eraShapes(words.eras.map(escaped).join("|"))];
+
+const FULLWIDTH_OFFSET = 0xfee0;
+const halfWidth = (text: string): string => text.replace(/[０-９]/gu, (char) => String.fromCharCode(char.charCodeAt(0) - FULLWIDTH_OFFSET));
+
+const MONTHS_IN_YEAR = 12;
+const PERCENT = 100;
+const DAYS_IN_MONTH = 31;
+
+type Groups = Readonly<Record<string, string | undefined>> | undefined;
+
+const numberIn = (groups: Groups, name: string): number => Number(halfWidth(groups?.[name] ?? ""));
+
+const isMonthDay = (month: number, day: number): boolean => month >= 1 && month <= MONTHS_IN_YEAR && day >= 1 && day <= DAYS_IN_MONTH;
+
+/** 数字の組が月と日として読めるか。読めなければ版や番号（2026.10.45）。 */
+const isCalendar = (groups: Groups): boolean => isMonthDay(numberIn(groups, "m"), numberIn(groups, "d"));
+
+/** 月の名前で書いた日付の日。月は名前で決まる。 */
+const isDay = (groups: Groups): boolean => isMonthDay(1, numberIn(groups, "d"));
+
+/** 年が後ろの数字の組（10/2/2026、25/12/2026）。月日の順は国で違うので、どちらの順でも読めればよい。 */
+const isEitherOrder = (groups: Groups): boolean =>
+  isMonthDay(numberIn(groups, "m"), numberIn(groups, "d")) || isMonthDay(numberIn(groups, "d"), numberIn(groups, "m"));
+
+/** 書き方ごとの、数が月と日として読めるかの確かめ。 */
+const CHECKS: Readonly<Record<string, (groups: Groups) => boolean>> = {
+  iso: isCalendar,
+  "year-slash": isCalendar,
+  "year-dot": isCalendar,
+  "nen-gappi": isCalendar,
+  "slash-year-last": isEitherOrder,
+  era: isCalendar,
+  "month-day-year": isDay,
+  "day-month-year": isDay,
+};
+
+const found = (style: string, pattern: RegExp, text: string): WrittenDate[] => {
+  const check = CHECKS[style];
+  return [...text.matchAll(pattern)].flatMap((match) => (check?.(match.groups) === true ? [{ offset: match.index, written: match[0], style }] : []));
+};
+
+/** 文字列の中の日付と、その書き方。形どうしは同じ所に当たらない（区切りの字と、前後に来ない字が違う）。 */
+export const datesIn = (text: string, words: DateWords): WrittenDate[] =>
+  [
+    ...numericShapes().flatMap(([style, pattern]) => found(style, pattern, text)),
+    ...namedShapes(words).flatMap(([style, pattern]) => found(style, pattern, text)),
+  ].toSorted((left, right) => left.offset - right.offset);
+
+/** 少ないほうの書き方の日付と、多いほうの書き方の数と例。 */
+export type DateMinority = { readonly odd: readonly WrittenDate[]; readonly majority: WrittenDate; readonly count: number; readonly total: number };
+
+/**
+ * 書き方の多いほうと、それ以外の日付。ほかの書き方の日付の割合が limit パーセント以下のときだけ。
+ * それより多ければ、二つの書き方を使い分けている文書と読む（表は 2026-10-02、本文は 2026年10月2日）。いちばん多い書き方が同じ数で並べば言わない。
+ */
+export const dateMinority = (dates: readonly WrittenDate[], limit: number): DateMinority | undefined => {
+  const counts = new Map<string, number>();
+  dates.forEach((date) => counts.set(date.style, (counts.get(date.style) ?? 0) + 1));
+  const [top, second] = [...counts].toSorted((left, right) => right[1] - left[1]);
+  if (top === undefined || second === undefined || top[1] === second[1]) return undefined;
+  const odd = dates.filter((date) => date.style !== top[0]);
+  const majority = dates.find((date) => date.style === top[0]);
+  if (majority === undefined || odd.length * PERCENT > dates.length * limit) return undefined;
+  return { odd, majority, count: top[1], total: dates.length };
+};
+
+/** 行の頭から日付までが、箇条書きの印・表の区切り・強調の印と空白だけ。予定表の行の頭の日付（- 2026-10-16: …、| 2026-10-16 |）。 */
+const LINE_HEAD = /(?:^|\|)[\s>*_+-]*(?:\d+[.)]\s+)?$/u;
+
+/** 日付が行の頭（箇条書きの項目や表の升の頭）にあるか。 */
+export const opensLine = (text: string, offset: number): boolean => LINE_HEAD.test(text.slice(text.lastIndexOf("\n", offset - 1) + 1, offset));
+
+/**
+ * 行の頭の日付と、文の中の日付を分ける。予定表を 2026-10-16 で並べ、文では 14 October 2026 と書くのは使い分けで、混ざりではない。
+ * 書き方はそれぞれの組の中でだけ比べる。
+ */
+export const dateGroups = (text: string, dates: readonly WrittenDate[]): WrittenDate[][] =>
+  [true, false].map((head) => dates.filter((date) => opensLine(text, date.offset) === head)).filter((group) => group.length > 0);

package/src/derived/date-arithmetic.ts

@@ -0,0 +1,63 @@
+/**
+ * 木の日付の値（2026-04-01、04-01、2026-04、2026）に期間を足す。年の無い月日は、同じ年の中で数える（年をまたげば翌年）。
+ * 暦は先発グレゴリオ暦。月を足して月末を越えるとき（1月31日の1か月後）は、その月の末日にする。
+ */
+export type DurationUnit = "day" | "week" | "month" | "year";
+
+export type CalendarDate = { readonly year: number | undefined; readonly month: number; readonly day: number };
+
+const FULL = /^(\d{4})-(\d{2})-(\d{2})$/u;
+const MONTH_DAY = /^(\d{2})-(\d{2})$/u;
+const YEAR_ONLY = /^(\d{4})$/u;
+const FULL_YEAR = /^(\d{4})-\d{2}(?:-\d{2})?$/u;
+
+/** 年がわからない月日を数えるときの年。うるう年でない年にすると 2月29日 が読めないので、うるう年を使う。 */
+const LEAP_YEAR = 2000;
+const DAYS_IN_WEEK = 7;
+const MONTHS_IN_YEAR = 12;
+const PADDED = 2;
+
+/** 年月日か月日の値。それ以外（年月、年だけ）は undefined。 */
+export const calendarDateOf = (value: string): CalendarDate | undefined => {
+  const full = FULL.exec(value);
+  if (full !== null) return { year: Number(full[1]), month: Number(full[2]), day: Number(full[3]) };
+  const monthDay = MONTH_DAY.exec(value);
+  return monthDay === null ? undefined : { year: undefined, month: Number(monthDay[1]), day: Number(monthDay[2]) };
+};
+
+/** 年を書いた値（2015、2015-04、2015-04-01）の年。 */
+export const yearOf = (value: string): number | undefined => {
+  const year = YEAR_ONLY.exec(value) ?? FULL_YEAR.exec(value);
+  return year === null ? undefined : Number(year[1]);
+};
+
+const utcOf = (year: number, month: number, day: number): Date => new Date(Date.UTC(year, month - 1, day));
+
+const lastDayOf = (year: number, month: number): number => utcOf(year, month + 1, 0).getUTCDate();
+
+const addMonths = (date: Date, months: number): Date => {
+  const target = utcOf(date.getUTCFullYear(), date.getUTCMonth() + 1 + months, 1);
+  const day = Math.min(date.getUTCDate(), lastDayOf(target.getUTCFullYear(), target.getUTCMonth() + 1));
+  return utcOf(target.getUTCFullYear(), target.getUTCMonth() + 1, day);
+};
+
+const addDays = (date: Date, days: number): Date => utcOf(date.getUTCFullYear(), date.getUTCMonth() + 1, date.getUTCDate() + days);
+
+/** 日付に期間を足す（負なら引く）。 */
+export const shifted = (date: Date, amount: number, unit: DurationUnit): Date => {
+  if (unit === "day") return addDays(date, amount);
+  if (unit === "week") return addDays(date, amount * DAYS_IN_WEEK);
+  return addMonths(date, unit === "month" ? amount : amount * MONTHS_IN_YEAR);
+};
+
+/** 年の無い月日には仮の年を付ける。 */
+export const dateOf = (date: CalendarDate): Date => utcOf(date.year ?? LEAP_YEAR, date.month, date.day);
+
+export const sameDay = (left: Date, right: Date): boolean => left.getTime() === right.getTime();
+
+/** 日付を、元の値と同じ細かさの値（年が無ければ月日だけ）にする。 */
+export const valueOf = (date: Date, withYear: boolean): string => {
+  const month = String(date.getUTCMonth() + 1).padStart(PADDED, "0");
+  const day = String(date.getUTCDate()).padStart(PADDED, "0");
+  return withYear ? `${String(date.getUTCFullYear())}-${month}-${day}` : `${month}-${day}`;
+};

package/src/derived/durations.ts

@@ -0,0 +1,52 @@
+import type { Span } from "../plugin.ts";
+import { bySentence } from "./sentence-buckets.ts";
+import { calendarDateOf, dateOf, shifted, valueOf, type DurationUnit } from "./date-arithmetic.ts";
+
+/**
+ * 始まりと期間と終わりを一つの文に書いたもの（「4月1日から3か月間（6月30日まで）」"for 3 months from April 1, 2026 (until
+ * June 30, 2026)"）。始まりに期間を足した日が終わりと合わなければ言う。期間の数え方は三通りある。終わりの日を含めて数える
+ * （4月1日から3か月間は6月30日まで）か、含めずに数える（7月1日まで）か、初日を数えずに期間が過ぎた日（「到達日から2週間が経過した日
+ * （4月16日）」）。どれとも合わないときだけ言う。
+ * 文に日付がちょうど二つ、期間がちょうど一つのときだけ読む。それより多い文は、どの日付とどの期間が組なのか決まらない。
+ */
+export type DatedValue = Span & { readonly value: string };
+
+export type Duration = Span & { readonly amount: number; readonly unit: DurationUnit };
+
+export type DurationMismatch = { readonly start: DatedValue; readonly duration: Duration; readonly end: DatedValue; readonly expected: string };
+
+/** 期間を足した日と、その前の日（終わりを含めて数える）と、次の日（初日を数えずに期間が過ぎた日）。 */
+const expectedEnds = (start: Date, duration: Duration): Date[] => {
+  const exclusive = shifted(start, duration.amount, duration.unit);
+  return [shifted(exclusive, -1, "day"), exclusive, shifted(exclusive, 1, "day")];
+};
+
+/** 年まで書いた二つの日付は、早いほうを始まりにする（「7月1日まで、4月1日から3か月」）。月日だけなら書いた順（年をまたぐことがある）。 */
+const ordered = (first: DatedValue, second: DatedValue, withYear: boolean): [DatedValue, DatedValue] =>
+  withYear && second.value < first.value ? [second, first] : [first, second];
+
+const mismatchOf = (written: readonly [DatedValue, DatedValue], duration: Duration): DurationMismatch | undefined => {
+  const [firstDate, secondDate] = [calendarDateOf(written[0].value), calendarDateOf(written[1].value)];
+  if (firstDate === undefined || secondDate === undefined || (firstDate.year === undefined) !== (secondDate.year === undefined)) return undefined;
+  const withYear = firstDate.year !== undefined;
+  const [start, end] = ordered(written[0], written[1], withYear);
+  const startDate = calendarDateOf(start.value);
+  if (startDate === undefined) return undefined;
+  const ends = expectedEnds(dateOf(startDate), duration);
+  const [inclusive] = ends;
+  if (inclusive === undefined || ends.some((candidate) => valueOf(candidate, withYear) === end.value)) return undefined;
+  return { start, duration, end, expected: valueOf(inclusive, withYear) };
+};
+
+/** 一つの文の中の、始まり・期間・終わり。 */
+export const durationMismatches = (sentences: readonly Span[], dates: readonly DatedValue[], durations: readonly Duration[]): DurationMismatch[] => {
+  const datesIn = bySentence(sentences, dates);
+  return [...bySentence(sentences, durations).entries()].flatMap(([index, inDurations]) => {
+    const inDates = datesIn.get(index) ?? [];
+    const [first, second] = inDates;
+    const [duration] = inDurations;
+    if (inDates.length !== 2 || inDurations.length !== 1 || first === undefined || second === undefined || duration === undefined) return [];
+    const mismatch = mismatchOf([first, second], duration);
+    return mismatch === undefined ? [] : [mismatch];
+  });
+};

package/src/derived/elapsed.ts

@@ -0,0 +1,125 @@
+import type { Span } from "../plugin.ts";
+import { bySentence } from "./sentence-buckets.ts";
+
+/**
+ * 起点の年から数えた年数（「2015年創業」と「創業10年」、「1980年生まれ」と「45歳」）を、文書の日付の年と比べる。
+ * 年数は、文書の年から起点の年を引いた数と、前後一年まで合えばよい（誕生日の前後、「10年目」の数え方）。それより離れていれば言う。
+ * 創業の年数は、起点の語のすぐ後ろか前に、つなぎの語（から、以来、since）だけを挟んで書いたもの（「創業10年」「設立から10年」）。
+ * 起点の年は同じ文の中で起点の語に一番近い年、無ければ文書の中で起点の語に付いた年（一通りのときだけ）。
+ * 年齢は、生まれた年と同じ文の年齢だけを比べる（文に生まれた年と年齢が一つずつのとき）。別の人の年齢と組にしないため。
+ */
+export type Year = Span & { readonly year: number };
+
+export type Elapsed = Span & { readonly amount: number };
+
+export type OriginWord = Span & { readonly pattern: string };
+
+export type ElapsedInput = {
+  readonly source: string;
+  readonly sentences: readonly Span[];
+  readonly years: readonly Year[];
+  /** 創業・設立・founded など、年数を数え始める語。 */
+  readonly foundings: readonly OriginWord[];
+  /** 生まれ・born など。 */
+  readonly births: readonly OriginWord[];
+  /** 年数（10年、20周年、10 years）。 */
+  readonly elapsed: readonly Elapsed[];
+  /** 年齢（45歳、aged 45）。 */
+  readonly ages: readonly Elapsed[];
+  /** 起点の語と年数のあいだに書いてよい語（から、以来、since）。何も挟まないのもよい。 */
+  readonly links: readonly string[];
+  /** 文書の日付の年。 */
+  readonly reference: number;
+};
+
+export type ElapsedMismatch = { readonly elapsed: Elapsed; readonly origin: Year; readonly expected: number };
+
+const TOLERANCE = 1;
+
+const distance = (left: Span, right: Span): number => Math.max(0, left.start - right.end, right.start - left.end);
+
+const nearest = <T extends Span>(spans: readonly T[], to: Span): T | undefined =>
+  spans.reduce<T | undefined>((best, span) => (best === undefined || distance(span, to) < distance(best, to) ? span : best), undefined);
+
+/** 起点の語と年数のあいだが、空か、つなぎの語だけ。 */
+const attached = (input: ElapsedInput, word: Span, amount: Span): boolean => {
+  const between = (word.end <= amount.start ? input.source.slice(word.end, amount.start) : input.source.slice(amount.end, word.start)).trim().toLowerCase();
+  return between === "" || input.links.some((link) => link.toLowerCase() === between);
+};
+
+const mismatchOf = (input: ElapsedInput, elapsed: Elapsed, origin: Year): ElapsedMismatch[] => {
+  const expected = input.reference - origin.year;
+  return expected >= 0 && Math.abs(elapsed.amount - expected) > TOLERANCE ? [{ elapsed, origin, expected }] : [];
+};
+
+/** 起点の年は、起点の語のすぐ近く（「2015年に創業」「founded in 2015」）に書いたものだけ。同じ文の遠い数（Model 2015）は年と読まない。 */
+const MAX_YEAR_DISTANCE = 6;
+
+/** 一つの文の中の、読んだものすべて。 */
+type InSentence = {
+  readonly years: readonly Year[];
+  readonly foundings: readonly OriginWord[];
+  readonly births: readonly OriginWord[];
+  readonly elapsed: readonly Elapsed[];
+  readonly ages: readonly Elapsed[];
+};
+
+const yearAt = (years: readonly Year[], word: Span): Year | undefined => {
+  const year = nearest(years, word);
+  return year !== undefined && distance(year, word) <= MAX_YEAR_DISTANCE ? year : undefined;
+};
+
+/** 文書の中で、起点の語（創業、設立、founded）に付いた年。一通りに決まらなければ undefined。 */
+const documentYear = (sentences: readonly InSentence[]): Year | undefined => {
+  const found = sentences.flatMap((sentence) =>
+    sentence.foundings.flatMap((word) => {
+      const year = yearAt(sentence.years, word);
+      return year === undefined ? [] : [year];
+    }),
+  );
+  return new Set(found.map((year) => year.year)).size === 1 ? found[0] : undefined;
+};
+
+const foundingMismatches = (input: ElapsedInput, sentence: InSentence, fallback: Year | undefined): ElapsedMismatch[] =>
+  sentence.elapsed.flatMap((elapsed) => {
+    const word = sentence.foundings.find((candidate) => attached(input, candidate, elapsed));
+    if (word === undefined) return [];
+    const origin = yearAt(sentence.years, word) ?? fallback;
+    return origin === undefined ? [] : mismatchOf(input, elapsed, origin);
+  });
+
+const ageMismatches = (input: ElapsedInput, sentence: InSentence): ElapsedMismatch[] => {
+  const [word] = sentence.births;
+  const [age] = sentence.ages;
+  if (sentence.births.length !== 1 || sentence.ages.length !== 1 || word === undefined || age === undefined) return [];
+  const origin = sentence.years.length === 1 ? sentence.years[0] : undefined;
+  return origin === undefined ? [] : mismatchOf(input, age, origin);
+};
+
+/** 文書の年より前の年だけを起点にする。文書の日付そのもの（「2026年4月1日現在、創業10年」）は起点ではない。 */
+const sentencesOf = (input: ElapsedInput): InSentence[] => {
+  const groups = {
+    years: bySentence(
+      input.sentences,
+      input.years.filter((year) => year.year < input.reference),
+    ),
+    foundings: bySentence(input.sentences, input.foundings),
+    births: bySentence(input.sentences, input.births),
+    elapsed: bySentence(input.sentences, input.elapsed),
+    ages: bySentence(input.sentences, input.ages),
+  };
+  const indexes = [...new Set([...groups.foundings.keys(), ...groups.births.keys()])].toSorted((left, right) => left - right);
+  return indexes.map((index) => ({
+    years: groups.years.get(index) ?? [],
+    foundings: groups.foundings.get(index) ?? [],
+    births: groups.births.get(index) ?? [],
+    elapsed: groups.elapsed.get(index) ?? [],
+    ages: groups.ages.get(index) ?? [],
+  }));
+};
+
+export const elapsedMismatches = (input: ElapsedInput): ElapsedMismatch[] => {
+  const sentences = sentencesOf(input);
+  const fallback = documentYear(sentences);
+  return sentences.flatMap((sentence) => [...foundingMismatches(input, sentence, fallback), ...ageMismatches(input, sentence)]);
+};

package/src/derived/relative-dates.ts

@@ -0,0 +1,96 @@
+import type { Span } from "../plugin.ts";
+import { calendarDateOf, dateOf, shifted, valueOf, type CalendarDate, type DurationUnit } from "./date-arithmetic.ts";
+
+/**
+ * 基準の日付からの相対で書いた日付（「10月2日に申し込み、3日後の10月5日に通知」"Applications open on May 1, 2026. Two weeks
+ * later, on May 15, 2026, …"）。相対の期間のすぐ後ろに書いた日付を、基準の日付に期間を足した（前なら引いた）日と比べる。
+ * 基準は、同じ文で相対の期間より前にある一番近い日付。文が相対の期間で始まる（「Two weeks later, …」）ときだけ、前の文の日付も
+ * 基準にする。基準と書いた日付の片方にしか年が無ければ、月日だけを比べる。
+ */
+export type DatedValue = Span & { readonly value: string };
+
+/** 「3日後」「two weeks later」: 期間と向き（後なら 1、前なら -1）。 */
+export type Relative = Span & { readonly amount: number; readonly unit: DurationUnit; readonly direction: 1 | -1 };
+
+export type RelativeMismatch = { readonly base: DatedValue; readonly relative: Relative; readonly target: DatedValue; readonly expected: string };
+
+export type RelativeInput = {
+  readonly source: string;
+  /** 文の始まりの昇順。 */
+  readonly sentences: readonly Span[];
+  /** 書いた順の日付。 */
+  readonly dates: readonly DatedValue[];
+  readonly relatives: readonly Relative[];
+  /** 相対の語と日付のあいだに書いてよい語（の、on）。空白と区切りの記号（、, (）は、いつでもよい。 */
+  readonly links: readonly string[];
+};
+
+const SEPARATORS = /[\s,、，(（:：]/gu;
+
+/** start の昇順の列で、start が offset 以上の最初の位置。二分探索で、相対の期間ごとに文書の頭から数え直さない。 */
+const firstAtOrAfter = (spans: readonly Span[], offset: number): number => {
+  let low = 0;
+  let high = spans.length;
+  while (low < high) {
+    const middle = (low + high) >> 1;
+    if ((spans[middle]?.start ?? Infinity) < offset) low = middle + 1;
+    else high = middle;
+  }
+  return low;
+};
+
+const sentenceIndexOf = (sentences: readonly Span[], offset: number): number => {
+  const index = firstAtOrAfter(sentences, offset + 1) - 1;
+  const sentence = sentences[index];
+  return sentence !== undefined && offset < sentence.end ? index : -1;
+};
+
+/** 文の頭から相対の期間までに、空白と記号のほかに何も無い。 */
+const opensSentence = (source: string, sentence: Span, relative: Relative): boolean => !/[\p{L}\p{N}]/u.test(source.slice(sentence.start, relative.start));
+
+/** 相対の語のすぐ後ろの日付。あいだは区切りの記号と、つなぎの語だけ。 */
+const targetOf = (input: RelativeInput, relative: Relative): DatedValue | undefined => {
+  const next = input.dates[firstAtOrAfter(input.dates, relative.end)];
+  if (next === undefined) return undefined;
+  const between = input.source.slice(relative.end, next.start).replace(SEPARATORS, "").toLowerCase();
+  return between === "" || input.links.some((link) => link.replace(SEPARATORS, "").toLowerCase() === between) ? next : undefined;
+};
+
+/** 同じ文で、相対の期間より前にある一番近い日付。文の頭の相対の期間なら、前の文まで見る。 */
+const baseOf = (input: RelativeInput, relative: Relative): DatedValue | undefined => {
+  const index = sentenceIndexOf(input.sentences, relative.start);
+  const sentence = input.sentences[index];
+  if (sentence === undefined) return undefined;
+  const from = opensSentence(input.source, sentence, relative) ? (input.sentences[index - 1]?.start ?? sentence.start) : sentence.start;
+  const before = input.dates[firstAtOrAfter(input.dates, relative.start) - 1];
+  return before !== undefined && before.end <= relative.start && before.start >= from ? before : undefined;
+};
+
+/** 年の無い月日は、年を決めずに数えられない（2月28日の1日後は、うるう年なら2月29日）。考えられる年をすべて試す。 */
+const LEAP_YEAR = 2000;
+const COMMON_YEAR = 2001;
+
+const baseYears = (base: CalendarDate, target: CalendarDate): number[] => {
+  if (base.year !== undefined) return [base.year];
+  return target.year === undefined ? [LEAP_YEAR, COMMON_YEAR] : [target.year - 1, target.year, target.year + 1];
+};
+
+const mismatchOf = (base: DatedValue, relative: Relative, target: DatedValue): RelativeMismatch | undefined => {
+  const [baseDate, targetDate] = [calendarDateOf(base.value), calendarDateOf(target.value)];
+  if (baseDate === undefined || targetDate === undefined) return undefined;
+  const withYear = baseDate.year !== undefined && targetDate.year !== undefined;
+  const written = withYear ? target.value : valueOf(dateOf(targetDate), false);
+  const expected = baseYears(baseDate, targetDate).map((year) =>
+    valueOf(shifted(dateOf({ ...baseDate, year }), relative.amount * relative.direction, relative.unit), withYear),
+  );
+  const [shown] = expected;
+  return shown === undefined || expected.includes(written) ? undefined : { base, relative, target, expected: shown };
+};
+
+export const relativeMismatches = (input: RelativeInput): RelativeMismatch[] =>
+  input.relatives.flatMap((relative) => {
+    const target = targetOf(input, relative);
+    const base = baseOf(input, relative);
+    const mismatch = target === undefined || base === undefined ? undefined : mismatchOf(base, relative, target);
+    return mismatch === undefined ? [] : [mismatch];
+  });

package/src/derived/sentence-buckets.ts

@@ -0,0 +1,29 @@
+import type { Span } from "../plugin.ts";
+
+/** 文の始まりの昇順の列で、span をすっかり含む文の番号。無ければ -1。二分探索で引く。 */
+const sentenceIndexOf = (sentences: readonly Span[], span: Span): number => {
+  let low = 0;
+  let high = sentences.length - 1;
+  while (low <= high) {
+    const middle = (low + high) >> 1;
+    const sentence = sentences[middle];
+    if (sentence === undefined) return -1;
+    if (span.start < sentence.start) high = middle - 1;
+    else if (span.start >= sentence.end) low = middle + 1;
+    else return span.end <= sentence.end ? middle : -1;
+  }
+  return -1;
+};
+
+/** span を文ごとに分ける。文をまたぐ span はどの文にも入れない。文の数 × span の数を回らないため。 */
+export const bySentence = <T extends Span>(sentences: readonly Span[], spans: readonly T[]): ReadonlyMap<number, readonly T[]> => {
+  const buckets = new Map<number, T[]>();
+  spans.forEach((span) => {
+    const index = sentenceIndexOf(sentences, span);
+    if (index === -1) return;
+    const bucket = buckets.get(index);
+    if (bucket === undefined) buckets.set(index, [span]);
+    else bucket.push(span);
+  });
+  return buckets;
+};

package/src/detect-language.ts

@@ -6,7 +6,7 @@
  * 権威ある点数で、こちらは何も読み込んでいない状態での見当付け。
  * 似た計算を両方が持つのは、この順序から来る避けられない重複。
  *
- * 呼ぶ側が、コードブロック・インラインコード・URL・front matter を先に落としておくこと。
+ * 呼ぶ側が、コードブロック・インラインコード・URL・front matter を先に落としておくこと（guessLanguage は language-sample.ts で落とす）。
  * 技術文書は英語のコードを大量に含むため、落とさないと日本語文書が英語と判定される。
  */
 const JAPANESE = /[぀-ゟ゠-ヿ一-鿿]/gu;

package/src/detect.ts

@@ -5,6 +5,7 @@ import { parse } from "yaml";
 import { japaneseRatio, latinRatio } from "./detect-language.ts";
 import { referenceListSpans } from "./reference-lists.ts";
 import { withoutSpans } from "./soft-break.ts";
+import { languageSample } from "./language-sample.ts";
 
 export type LanguageGuess = { readonly language: string; readonly confidence: number; readonly from: string };
 
@@ -33,9 +34,15 @@ const bodyOf = (source: string): string => {
   return body.trim() === "" ? source : body;
 };
 
+/** The body without code, URLs and markup, or the whole body when nothing else is left (a document of only code). */
+const proseOf = (body: string): string => {
+  const sample = languageSample(body);
+  return sample.trim() === "" ? body : sample;
+};
+
 export const guessLanguage = (source: string): LanguageGuess => {
-  const body = bodyOf(source);
-  const japanese = japaneseRatio(body);
+  const prose = proseOf(bodyOf(source));
+  const japanese = japaneseRatio(prose);
   if (japanese >= JAPANESE_FLOOR) return { language: "ja", confidence: japanese, from: "content" };
-  return { language: "en", confidence: latinRatio(body), from: "content" };
+  return { language: "en", confidence: latinRatio(prose), from: "content" };
 };

package/src/detectors/absolute-exception.ts

@@ -0,0 +1,102 @@
+import type { Detector, Finding, Lexicon, ProseDocument, Sentence, Span, StructureNode } from "../plugin.ts";
+import { inDocumentOrder } from "../structure/issues.ts";
+import { contentKey, type Obligation } from "../structure/modal-conflict.ts";
+import { actOf, contradictedAbsolutes, type Statement } from "../structure/absolute-exception.ts";
+import { linesOf, lineNumberAt } from "../structure/lines.ts";
+import { isWithinAny, QUOTATION_MARKS, quotedIn } from "../quoted-span.ts";
+import { entryRanges } from "./lexicon-match.ts";
+import { obligationsOf } from "./modal-conflict.ts";
+import { quoteAt } from "./structure-tree.ts";
+
+const REQUIRING: ReadonlySet<string> = new Set(["must", "must-not"]);
+
+type Hit = { readonly word: string; readonly span: Span };
+
+/** Where the lexicon's words stand in the sentence, outside quotations, in reading order. Needs the sentence's tokens. */
+const hitsIn = (sentence: Sentence, lexicon: Lexicon, quoted: readonly Span[]): Hit[] => {
+  const tokens = sentence.tokens ?? [];
+  return lexicon
+    .flatMap((entry) =>
+      entryRanges(sentence, entry).flatMap((range) => {
+        const first = tokens[range.start];
+        const last = tokens[range.end - 1];
+        if (first === undefined || last === undefined) return [];
+        const span = { start: first.span.start, end: last.span.end };
+        return isWithinAny(quoted, span) ? [] : [{ word: entry.pattern, span }];
+      }),
+    )
+    .toSorted((left, right) => left.span.start - right.span.start);
+};
+
+const within = (span: Span, offset: number): boolean => offset >= span.start && offset < span.end;
+
+/** The units an exception may sit in and still be the statement's own proviso: articles, then lists, then paragraphs. */
+type Units = { readonly articles: readonly Span[]; readonly lists: readonly Span[]; readonly paragraphs: readonly Span[] };
+
+const unitsOf = (doc: ProseDocument, tree: StructureNode): Units => ({
+  articles: inDocumentOrder(tree).flatMap((node) => (node.kind === "article" ? [node.span] : [])),
+  lists: doc.lists.map((list) => list.span),
+  paragraphs: doc.paragraphs.map((paragraph) => paragraph.span),
+});
+
+/** The innermost span holding the offset: an article nested in another is a unit of its own. */
+const innermost = (spans: readonly Span[], offset: number): Span | undefined =>
+  spans
+    .filter((span) => within(span, offset))
+    .reduce<Span | undefined>((inner, span) => (inner === undefined || span.start > inner.start ? span : inner), undefined);
+
+const unitAt = (units: Units, offset: number): number =>
+  [units.articles, units.lists, units.paragraphs].map((spans) => innermost(spans, offset)).find((span) => span !== undefined)?.start ?? offset;
+
+type Words = { readonly absolute: Lexicon; readonly exception: Lexicon; readonly lightVerbs: ReadonlySet<string> };
+
+/** The sentence read for the rule, or nothing when it holds neither word. Only such sentences pay for the unit and the markers. */
+const statementOf = (sentence: Sentence, obligations: readonly Obligation[], words: Words, units: Units): Statement | undefined => {
+  const quoted = quotedIn(sentence, QUOTATION_MARKS);
+  const absolute = hitsIn(sentence, words.absolute, quoted);
+  const exception = hitsIn(sentence, words.exception, quoted);
+  const tokens = sentence.tokens;
+  if ((absolute.length === 0 && exception.length === 0) || tokens === undefined) return undefined;
+  const markers = obligations.filter((obligation) => within(sentence.span, obligation.start));
+  const read = [...absolute, ...exception].map((hit) => hit.span);
+  const key = contentKey(tokens, [...markers, ...read]);
+  const [marker] = markers;
+  return {
+    offset: sentence.span.start,
+    words: key === "" ? [] : key.split(" "),
+    absolute: absolute[0]?.word,
+    exception: exception[0]?.word,
+    obligation: markers.length === 1 && REQUIRING.has(marker?.type ?? ""),
+    act: marker === undefined ? undefined : actOf(tokens, marker, read, words.lightVerbs),
+    unit: unitAt(units, sentence.span.start),
+  };
+};
+
+/** An absolute obligation (すべての〜は…しなければならない, "must always") that a later sentence elsewhere makes an exception to. */
+export const absoluteException: Detector = (doc, options): Finding[] => {
+  if (doc.structure === undefined) return [];
+  const words = {
+    absolute: options.lexicon ?? [],
+    exception: doc.lexicons["exception-marker"] ?? [],
+    lightVerbs: new Set((doc.lexicons["light-verb"] ?? []).map((entry) => entry.pattern.toLowerCase())),
+  };
+  const obligations = obligationsOf(doc.structure);
+  const units = unitsOf(doc, doc.structure);
+  const statements = doc.sentences
+    .filter((sentence) => sentence.embeddedLanguage === undefined)
+    .flatMap((sentence) => statementOf(sentence, obligations, words, units) ?? []);
+  const lines = linesOf(doc.source);
+  return contradictedAbsolutes(statements).map(({ absolute, exception }) => ({
+    rule: "",
+    severity: "warning",
+    line: 0,
+    column: 0,
+    quote: quoteAt(doc.source, absolute.offset),
+    values: {
+      word: absolute.absolute ?? "",
+      exception: exception.exception ?? "",
+      otherLine: lineNumberAt(lines, exception.offset) ?? 0,
+      offset: absolute.offset,
+    },
+  }));
+};