disclosure-bureau/web/lib/bilingual.ts

/**
 * bilingual.ts — split a narrated case-file body into a single language.
 *
 * The case-writer emits each scene twice: an English `## N. <title>` section
 * followed by its Brazilian-Portuguese twin `## N. <título>` (CLAUDE.md §3).
 * The reader, however, must see ONE language at a time — otherwise an EN
 * visitor reads the EN scene immediately followed by the same scene in PT,
 * and vice-versa.
 *
 * We classify each `## ` section by language (PT-BR carries diacritics and
 * Portuguese function words; the English scene text does not) and keep only
 * the sections matching the active locale. This is more robust than relying
 * on EN-then-PT ordering: it survives a missing twin or an out-of-order
 * section. If a body can't be split (no `## ` sections, or none match the
 * locale), we return it unchanged — better to show everything than to hide
 * content.
 */

const PT_DIACRITICS = /[áàâãéêíóôõúüç]/gi;
const PT_WORDS =
  /\b(não|que|com|para|sobre|uma|um|também|após|às|aos|nos|nas|pelo|pela|seu|sua|foi|era|está|são|sem|entre|quando|onde|índice|registro|relatório|noite|céu)\b/gi;
const EN_WORDS =
  /\b(the|and|with|what|which|after|over|about|into|from|that|this|their|was|were|been|when|where|night|sky|report|record)\b/gi;

/**
 * Heuristic: does this markdown block read as Brazilian Portuguese?
 * Diacritics are the strongest signal — English narration here has none,
 * while PT-BR is accent-rich. Function-word counts break ties.
 */
export function isPortuguese(block: string): boolean {
  const diacritics = (block.match(PT_DIACRITICS) || []).length;
  const pt = (block.match(PT_WORDS) || []).length;
  const en = (block.match(EN_WORDS) || []).length;
  // Diacritics weigh heavily: a few accents reliably mark PT-BR even when an
  // English place name (e.g. "Áden") sneaks one into an EN block.
  return diacritics * 2 + pt > en;
}

/** Section number token ("I", "IV", "3") from an H2 like "## IV. Cables". */
function sectionKey(section: string): string | null {
  const m = section.match(/^##\s+([ivxlcdm\d]+)[.)]/i);
  return m ? m[1].toUpperCase() : null;
}

/**
 * Some legacy files tag the heading language explicitly:
 * "## §1 — The Case at Hand (EN)" / "## §1 — O Caso em Mãos (PT-BR)".
 * That tag is the strongest possible signal. Returns the language when a tag
 * is present, else null.
 */
function langTag(section: string): "en" | "pt-br" | null {
  const head = section.split("\n", 1)[0] ?? "";
  if (/\((?:pt-?br|pt|português|portugues)\)\s*$/i.test(head)) return "pt-br";
  if (/\((?:en|en-?us|english|inglês|ingles)\)\s*$/i.test(head)) return "en";
  return null;
}

/** Strip a trailing "(EN)" / "(PT-BR)" tag and dangling rule from a section. */
function cleanSection(section: string): string {
  return section
    .replace(/^(##\s+.*?)\s*\((?:en|en-?us|english|inglês|ingles|pt-?br|pt|português|portugues)\)\s*$/im, "$1")
    .replace(/\n+---\s*$/m, "")
    .trimEnd();
}

/**
 * Return only the portion of a bilingual case body that matches `locale`.
 * Operates on a body whose H1 title lines have already been stripped.
 *
 * Primary signal is structural, not linguistic: the narrator emits each
 * numbered scene as an English `## N.` section immediately followed by its
 * Brazilian-Portuguese twin `## N.`. So for a section number seen twice, the
 * first occurrence is English and the second is Portuguese — a far more
 * reliable split than language sniffing (an English scene quoting accented
 * names would otherwise be misread as Portuguese). Language detection is the
 * fallback only for unpaired or unnumbered sections.
 */
export function pickLocaleBody(body: string, locale: "en" | "pt-br"): string {
  const parts = body.split(/(?=^##\s)/m);
  // Content before the first H2 (rare intro/orphan) is shown in both languages.
  const head = parts.length > 0 && !/^##\s/.test(parts[0]) ? parts[0].trimEnd() : "";
  const sections = parts.filter((p) => /^##\s/.test(p));
  if (sections.length === 0) return body;

  const counts = new Map<string, number>();
  for (const s of sections) {
    const k = sectionKey(s);
    if (k) counts.set(k, (counts.get(k) ?? 0) + 1);
  }

  const wantPt = locale === "pt-br";
  const seen = new Map<string, number>();
  const picked: string[] = [];
  for (const s of sections) {
    const tag = langTag(s);
    const k = sectionKey(s);
    let isPt: boolean;
    if (tag) {
      isPt = tag === "pt-br"; // explicit heading tag — most reliable
    } else if (k && (counts.get(k) ?? 0) >= 2) {
      const order = seen.get(k) ?? 0;
      seen.set(k, order + 1);
      isPt = order >= 1; // first = EN, second = PT-BR
    } else {
      isPt = isPortuguese(s); // unpaired / unnumbered — sniff the language
    }
    if (isPt === wantPt) picked.push(cleanSection(s));
  }
  // Can't isolate the requested language — fall back to the full body so no
  // content silently disappears.
  if (picked.length === 0) return body;

  const out = picked.join("\n\n");
  return head ? `${head}\n\n${out}` : out;
}