docrev 0.10.1 → 0.10.2

package/lib/wordcomments.ts

@@ -1,840 +1,840 @@
-/**
- * Word comment injection with reply threading
- *
- * Flow:
- * 1. prepareMarkdownWithMarkers() - Parse comments, detect reply relationships
- *    - First comment in a cluster = parent (gets markers: ⟦CMS:n⟧anchor⟦CME:n⟧)
- *    - Subsequent adjacent comments = replies (no markers, attach to parent)
- * 2. Pandoc converts to DOCX
- * 3. injectCommentsAtMarkers() - Insert comment ranges for parents only
- *    - Replies go in comments.xml with parent reference in commentsExtended.xml
- */
-
-import * as fs from 'fs';
-import AdmZip from 'adm-zip';
-import { escapeXml } from './utils.js';
-
-const MARKER_START_PREFIX = '⟦CMS:';
-const MARKER_END_PREFIX = '⟦CME:';
-const MARKER_SUFFIX = '⟧';
-
-interface ParsedComment {
-  author: string;
-  text: string;
-  anchor: string | null;
-  start: number;
-  end: number;
-  fullMatch: string;
-}
-
-interface PreparedComment extends ParsedComment {
-  isReply: boolean;
-  parentIdx: number | null;
-  commentIdx: number;
-  anchorFromReply?: boolean;
-  placesParentMarkers?: boolean;
-}
-
-interface PrepareResult {
-  markedMarkdown: string;
-  comments: PreparedComment[];
-}
-
-interface CommentWithIds extends PreparedComment {
-  id: string;
-  paraId: string;
-  paraId2: string;
-  durableId: string;
-  parentParaId?: string;
-}
-
-interface InjectionResult {
-  success: boolean;
-  commentCount: number;
-  replyCount?: number;
-  skippedComments: number;
-  error?: string;
-}
-
-function generateParaId(commentIdx: number, paraNum: number): string {
-  // Generate 8-character uppercase hex ID matching Word format
-  // Word uses IDs like "3F25BC58", "0331C187"
-  // Must be deterministic - same inputs always produce same output
-  const id = 0x10000000 + (commentIdx * 0x00100000) + (paraNum * 0x00001000);
-  return id.toString(16).toUpperCase().padStart(8, '0');
-}
-
-/**
- * Parse comments and create markers
- *
- * Returns:
- * - markedMarkdown: markdown with markers for parent comments only
- * - comments: array with author, text, isReply, parentIdx
- */
-export function prepareMarkdownWithMarkers(markdown: string): PrepareResult {
-  // Match the comment block first; extend manually to capture an optional
-  // trailing `[anchor]{.mark}` span. A regex `[^\]]+` for the anchor would
-  // bail on the inner `]` of nested syntax (e.g. `[[0..9]]{.mark}` or
-  // `[*phrase*]{.mark}` after pandoc-rewriting), so we walk the brackets
-  // ourselves and verify a `{.mark}` suffix.
-  const commentPattern = /\{>>([\s\S]+?)<<\}/g;
-
-  function tryParseTrailingAnchor(
-    text: string,
-    fromIdx: number,
-  ): { anchor: string; endIdx: number } | null {
-    let i = fromIdx;
-    while (i < text.length && /\s/.test(text[i] ?? '')) i++;
-    if (text[i] !== '[') return null;
-    let depth = 1;
-    let j = i + 1;
-    while (j < text.length) {
-      const ch = text[j];
-      if (ch === '[') depth++;
-      else if (ch === ']') {
-        depth--;
-        if (depth === 0) break;
-      }
-      j++;
-    }
-    if (depth !== 0) return null;
-    if (text.slice(j + 1, j + 8) !== '{.mark}') return null;
-    return { anchor: text.slice(i + 1, j), endIdx: j + 8 };
-  }
-
-  const REPLY_PREFIX = '↪ ';
-  const rawMatches: (ParsedComment & { explicitReply: boolean })[] = [];
-  let match: RegExpExecArray | null;
-  while ((match = commentPattern.exec(markdown)) !== null) {
-    const content = match[1] ?? '';
-    let author = 'Unknown';
-    let text = content;
-    const colonIdx = content.indexOf(':');
-    if (colonIdx > 0 && colonIdx < 30) {
-      author = content.slice(0, colonIdx).trim();
-      text = content.slice(colonIdx + 1).trim();
-    }
-
-    // The `↪ ` prefix is the authoritative reply signal emitted by
-    // `insertCommentsIntoMarkdown`. Strip it from the author before injection
-    // so Word displays the real name.
-    let explicitReply = false;
-    if (author.startsWith(REPLY_PREFIX)) {
-      explicitReply = true;
-      author = author.slice(REPLY_PREFIX.length).trim();
-    }
-
-    const commentEnd = match.index + match[0].length;
-    const trailing = tryParseTrailingAnchor(markdown, commentEnd);
-
-    rawMatches.push({
-      author,
-      text,
-      anchor: trailing ? trailing.anchor : null,
-      start: match.index,
-      end: trailing ? trailing.endIdx : commentEnd,
-      fullMatch: markdown.slice(match.index, trailing ? trailing.endIdx : commentEnd),
-      explicitReply,
-    });
-
-    // Advance regex lastIndex past the consumed anchor so the next iteration
-    // doesn't re-scan inside it (e.g. `[*emphasis*]{.mark}` would otherwise
-    // tempt the matcher to look for another `{>>...<<}` in the body of the
-    // anchor span).
-    if (trailing) {
-      commentPattern.lastIndex = trailing.endIdx;
-    }
-  }
-
-  if (rawMatches.length === 0) {
-    return { markedMarkdown: markdown, comments: [] };
-  }
-
-  // Two-mode reply detection driven by the markdown itself:
-  //   - If any comment carries the `↪ ` author prefix, the markdown came
-  //     through `insertCommentsIntoMarkdown` and we use prefix-only mode.
-  //     Distinct clusters that happen to land at gap=0 (a real failure
-  //     mode on dense reviewer docs — 298-comment paper produced 9 such
-  //     collisions) are not misthreaded.
-  //   - If no comment carries the prefix, the markdown was hand-typed.
-  //     Fall back to gap < 10 adjacency for backward compat with users
-  //     who write CriticMarkup directly.
-  const ADJACENT_THRESHOLD = 10;
-  const useExplicitMode = rawMatches.some(m => m.explicitReply);
-  const comments: PreparedComment[] = [];
-  let clusterParentIdx = -1;  // Index of first comment in current cluster
-  let lastCommentEnd = -1;
-
-  for (let i = 0; i < rawMatches.length; i++) {
-    const m = rawMatches[i];
-    if (!m) continue;
-
-    const gap = lastCommentEnd >= 0 ? m.start - lastCommentEnd : Infinity;
-    const isAdjacent = useExplicitMode
-      ? m.explicitReply
-      : gap < ADJACENT_THRESHOLD;
-
-    // Reset cluster if there's a gap (comments not in same cluster)
-    if (!isAdjacent) {
-      clusterParentIdx = -1;
-    }
-
-    if (clusterParentIdx === -1) {
-      // First comment in cluster = parent (regardless of author)
-      comments.push({
-        author: m.author,
-        text: m.text,
-        anchor: m.anchor,
-        start: m.start,
-        end: m.end,
-        fullMatch: m.fullMatch,
-        isReply: false,
-        parentIdx: null,
-        commentIdx: comments.length
-      });
-      clusterParentIdx = comments.length - 1;
-    } else {
-      // Subsequent comment in cluster = reply to first comment
-      comments.push({
-        author: m.author,
-        text: m.text,
-        anchor: m.anchor,
-        start: m.start,
-        end: m.end,
-        fullMatch: m.fullMatch,
-        isReply: true,
-        parentIdx: clusterParentIdx,
-        commentIdx: comments.length
-      });
-    }
-
-    lastCommentEnd = m.end;
-  }
-
-  // Propagate anchors from replies to parents
-  // If a reply has an anchor but its parent doesn't, move the anchor to the parent
-  // Track flags for special handling during marker generation
-  for (const c of comments) {
-    if (c.isReply && c.anchor && c.parentIdx !== null) {
-      const parent = comments[c.parentIdx];
-      if (parent && !parent.anchor) {
-        parent.anchor = c.anchor;
-        parent.anchorFromReply = true;  // Parent's anchor came from a reply (markers placed by reply)
-        c.placesParentMarkers = true;   // This reply should place the parent's markers
-        c.anchor = null;
-      }
-    }
-  }
-
-  // Build marked markdown - only parent comments get markers
-  // Process from end to start to preserve positions
-  let markedMarkdown = markdown;
-
-  for (let i = comments.length - 1; i >= 0; i--) {
-    const c = comments[i];
-    if (!c) continue;
-
-    if (c.isReply) {
-      // Reply: remove from document entirely (will be in comments.xml only)
-      // Also consume one preceding whitespace char to avoid double spaces.
-      // We deliberately consume at most one — walking arbitrarily backwards
-      // would shift positions that lower-index comments still depend on.
-      let removeStart = c.start;
-      if (removeStart > 0 && /\s/.test(markedMarkdown[removeStart - 1] ?? '')) {
-        removeStart--;
-      }
-
-      // If this reply places parent's markers (anchor was propagated)
-      if (c.placesParentMarkers && c.parentIdx !== null) {
-        // Extract anchor text from the original match
-        const anchorMatch = c.fullMatch.match(/\[([^\]]+)\]\{\.mark\}$/);
-        if (anchorMatch) {
-          const anchorText = anchorMatch[1] ?? '';
-          // Output markers with PARENT's index around the anchor text
-          const parentIdx = c.parentIdx;
-          const replacement = `${MARKER_START_PREFIX}${parentIdx}${MARKER_SUFFIX}${anchorText}${MARKER_END_PREFIX}${parentIdx}${MARKER_SUFFIX}`;
-          markedMarkdown = markedMarkdown.slice(0, removeStart) + replacement + markedMarkdown.slice(c.end);
-        } else {
-          markedMarkdown = markedMarkdown.slice(0, removeStart) + markedMarkdown.slice(c.end);
-        }
-      } else {
-        markedMarkdown = markedMarkdown.slice(0, removeStart) + markedMarkdown.slice(c.end);
-      }
-    } else {
-      // Parent comment
-      if (c.anchorFromReply) {
-        // Anchor markers are placed by the reply, just remove this comment.
-        // Consume one preceding whitespace char only (see reply branch above).
-        let removeStart = c.start;
-        if (removeStart > 0 && /\s/.test(markedMarkdown[removeStart - 1] ?? '')) {
-          removeStart--;
-        }
-        markedMarkdown = markedMarkdown.slice(0, removeStart) + markedMarkdown.slice(c.end);
-      } else {
-        // Normal case: replace with markers
-        const anchor = c.anchor || '';
-        const replacement = `${MARKER_START_PREFIX}${i}${MARKER_SUFFIX}${anchor}${MARKER_END_PREFIX}${i}${MARKER_SUFFIX}`;
-        markedMarkdown = markedMarkdown.slice(0, c.start) + replacement + markedMarkdown.slice(c.end);
-      }
-    }
-  }
-
-  return { markedMarkdown, comments };
-}
-
-function createCommentsXml(comments: CommentWithIds[]): string {
-  // Word expects date without milliseconds: 2025-12-30T08:33:00Z
-  const now = new Date().toISOString().replace(/\.\d{3}Z$/, 'Z');
-
-  let xml = '<?xml version="1.0" encoding="UTF-8" standalone="yes"?>\n';
-  // Minimal namespaces matching golden file structure
-  xml += '<w:comments xmlns:mc="http://schemas.openxmlformats.org/markup-compatibility/2006" xmlns:w="http://schemas.openxmlformats.org/wordprocessingml/2006/main" xmlns:w14="http://schemas.microsoft.com/office/word/2010/wordml" xmlns:w15="http://schemas.microsoft.com/office/word/2012/wordml" mc:Ignorable="w14 w15">';
-
-  // Use a consistent rsid (8-char hex) for all comments in this batch
-  const rsid = '00' + (Date.now() % 0xFFFFFF).toString(16).toUpperCase().padStart(6, '0');
-
-  for (const comment of comments) {
-    xml += `<w:comment w:id="${comment.id}" w:author="${escapeXml(comment.author)}" w:date="${now}" w:initials="${comment.author.split(' ').map(n => n[0]).join('')}">`;
-    // First paragraph: rsidRDefault="00000000", annotationRef without rStyle wrapper
-    xml += `<w:p w14:paraId="${comment.paraId}" w14:textId="77777777" w:rsidR="${rsid}" w:rsidRDefault="00000000">`;
-    xml += `<w:r><w:annotationRef/></w:r>`;
-    xml += `<w:r><w:t>${escapeXml(comment.text)}</w:t></w:r>`;
-    xml += `</w:p>`;
-    if (comment.isReply) {
-      // Second empty paragraph: rsidRDefault matches rsidR
-      xml += `<w:p w14:paraId="${comment.paraId2}" w14:textId="77777777" w:rsidR="${rsid}" w:rsidRDefault="${rsid}"/>`;
-    }
-    xml += `</w:comment>`;
-  }
-
-  xml += '</w:comments>';
-  return xml;
-}
-
-function createCommentsExtendedXml(comments: CommentWithIds[]): string {
-  let xml = '<?xml version="1.0" encoding="UTF-8" standalone="yes"?>\n';
-  // Minimal namespaces matching golden file structure
-  xml += '<w15:commentsEx xmlns:mc="http://schemas.openxmlformats.org/markup-compatibility/2006" xmlns:w="http://schemas.openxmlformats.org/wordprocessingml/2006/main" xmlns:w14="http://schemas.microsoft.com/office/word/2010/wordml" xmlns:w15="http://schemas.microsoft.com/office/word/2012/wordml" mc:Ignorable="w14 w15">';
-
-  for (const comment of comments) {
-    if (comment.isReply && comment.parentParaId) {
-      // Reply: use paraId2 (the second/empty paragraph) and link to parent's paraId
-      xml += `<w15:commentEx w15:paraId="${comment.paraId2}" w15:paraIdParent="${comment.parentParaId}" w15:done="0"/>`;
-    } else {
-      // Parent comment: use paraId (first paragraph)
-      xml += `<w15:commentEx w15:paraId="${comment.paraId}" w15:done="0"/>`;
-    }
-  }
-
-  xml += '</w15:commentsEx>';
-  return xml;
-}
-
-function generateDurableId(index: number): string {
-  // Generate unique 8-char hex ID for durableId
-  // CRITICAL: Must stay within signed 32-bit range (< 0x7FFFFFFF = 2147483647)
-  // Word interprets durableIds as signed 32-bit integers
-  const base = 0x10000000 + (Date.now() % 0x40000000); // Base between 0x10000000 and 0x50000000
-  const id = (base + index * 0x01000000) % 0x7FFFFFFF; // Keep under signed 32-bit max
-  return id.toString(16).toUpperCase().padStart(8, '0');
-}
-
-function createCommentsIdsXml(comments: CommentWithIds[]): string {
-  let xml = '<?xml version="1.0" encoding="UTF-8" standalone="yes"?>\n';
-  // Minimal namespaces matching golden file structure
-  xml += '<w16cid:commentsIds ';
-  xml += 'xmlns:mc="http://schemas.openxmlformats.org/markup-compatibility/2006" ';
-  xml += 'xmlns:w16cid="http://schemas.microsoft.com/office/word/2016/wordml/cid" ';
-  xml += 'mc:Ignorable="w16cid">';
-
-  for (const comment of comments) {
-    // ONE entry per comment using the LAST paragraph's paraId:
-    // - Parent comments (1 paragraph): use paraId
-    // - Reply comments (2 paragraphs): use paraId2 (the second/empty paragraph)
-    const useParaId = comment.isReply ? comment.paraId2 : comment.paraId;
-    xml += `<w16cid:commentId w16cid:paraId="${useParaId}" w16cid:durableId="${comment.durableId}"/>`;
-  }
-
-  xml += '</w16cid:commentsIds>';
-  return xml;
-}
-
-function createCommentsExtensibleXml(comments: CommentWithIds[]): string {
-  const now = new Date().toISOString().replace(/\.\d{3}Z$/, 'Z');
-
-  let xml = '<?xml version="1.0" encoding="UTF-8" standalone="yes"?>\n';
-  // Minimal namespaces matching golden file structure
-  xml += '<w16cex:commentsExtensible ';
-  xml += 'xmlns:mc="http://schemas.openxmlformats.org/markup-compatibility/2006" ';
-  xml += 'xmlns:w16cex="http://schemas.microsoft.com/office/word/2018/wordml/cex" ';
-  xml += 'mc:Ignorable="w16cex">';
-
-  for (const comment of comments) {
-    // ONE entry per comment using the durableId
-    xml += `<w16cex:commentExtensible w16cex:durableId="${comment.durableId}" w16cex:dateUtc="${now}"/>`;
-  }
-
-  xml += '</w16cex:commentsExtensible>';
-  return xml;
-}
-
-// Generate deterministic user IDs for authors (no hardcoded personal data)
-
-function createPeopleXml(comments: CommentWithIds[]): string {
-  // Extract unique authors
-  const authors = [...new Set(comments.map(c => c.author))];
-
-  let xml = '<?xml version="1.0" encoding="UTF-8" standalone="yes"?>\n';
-  xml += '<w15:people ';
-  xml += 'xmlns:mc="http://schemas.openxmlformats.org/markup-compatibility/2006" ';
-  xml += 'xmlns:x="http://schemas.openxmlformats.org/spreadsheetml/2006/main" ';
-  xml += 'xmlns:r="http://schemas.openxmlformats.org/officeDocument/2006/relationships" ';
-  xml += 'xmlns:w="http://schemas.openxmlformats.org/wordprocessingml/2006/main" ';
-  xml += 'xmlns:w14="http://schemas.microsoft.com/office/word/2010/wordml" ';
-  xml += 'xmlns:w15="http://schemas.microsoft.com/office/word/2012/wordml" ';
-  xml += 'xmlns:w16cex="http://schemas.microsoft.com/office/word/2018/wordml/cex" ';
-  xml += 'xmlns:w16cid="http://schemas.microsoft.com/office/word/2016/wordml/cid" ';
-  xml += 'xmlns:w16="http://schemas.microsoft.com/office/word/2018/wordml" ';
-  xml += 'xmlns:w16sdtdh="http://schemas.microsoft.com/office/word/2020/wordml/sdtdatahash" ';
-  xml += 'xmlns:w16se="http://schemas.microsoft.com/office/word/2015/wordml/symex" ';
-  xml += 'mc:Ignorable="w14 w15 w16se w16cid w16 w16cex w16sdtdh">';
-
-  for (const author of authors) {
-    const userId = generateUserId(author);
-    xml += `<w15:person w15:author="${escapeXml(author)}">`;
-    xml += `<w15:presenceInfo w15:providerId="Windows Live" w15:userId="${userId}"/>`;
-    xml += `</w15:person>`;
-  }
-
-  xml += '</w15:people>';
-  return xml;
-}
-
-function generateUserId(author: string): string {
-  // Generate a deterministic 16-char hex ID from author name
-  let hash = 0;
-  for (let i = 0; i < author.length; i++) {
-    hash = ((hash << 5) - hash) + author.charCodeAt(i);
-    hash = hash & hash;
-  }
-  return Math.abs(hash).toString(16).padStart(16, '0').slice(0, 16);
-}
-
-/**
- * Inject comments at marker positions
- */
-export async function injectCommentsAtMarkers(
-  docxPath: string,
-  comments: PreparedComment[],
-  outputPath: string
-): Promise<InjectionResult> {
-  try {
-    if (!fs.existsSync(docxPath)) {
-      return { success: false, commentCount: 0, skippedComments: 0, error: `File not found: ${docxPath}` };
-    }
-
-    if (comments.length === 0) {
-      fs.copyFileSync(docxPath, outputPath);
-      return { success: true, commentCount: 0, skippedComments: 0 };
-    }
-
-    const zip = new AdmZip(docxPath);
-    const documentEntry = zip.getEntry('word/document.xml');
-    if (!documentEntry) {
-      return { success: false, commentCount: 0, skippedComments: 0, error: 'Invalid DOCX: no document.xml' };
-    }
-
-    let documentXml = zip.readAsText(documentEntry);
-
-    // Assign IDs and paraIds (IDs start at 1, not 0 - Word convention)
-    const commentsWithIds: CommentWithIds[] = comments.map((c, idx) => ({
-      ...c,
-      id: String(idx + 1),
-      paraId: generateParaId(idx, 1),       // First paragraph (e.g., 10000001)
-      paraId2: generateParaId(idx, 2),      // Second paragraph (e.g., 10000002)
-      durableId: generateDurableId(idx),    // Unique ID for commentsIds/commentsExtensible
-    }));
-
-    // Link replies to parent paraIds
-    for (const c of commentsWithIds) {
-      if (c.isReply && c.parentIdx !== null) {
-        const parent = commentsWithIds[c.parentIdx];
-        if (parent) {
-          c.parentParaId = parent.paraId;
-        }
-      }
-    }
-
-    const injectedIds = new Set<string>();
-
-    // Process only parent comments (non-replies) for document ranges
-    const parentComments = commentsWithIds.filter(c => !c.isReply);
-
-    for (let i = parentComments.length - 1; i >= 0; i--) {
-      const comment = parentComments[i];
-      if (!comment) continue;
-      const idx = comment.commentIdx;
-
-      const startMarker = `${MARKER_START_PREFIX}${idx}${MARKER_SUFFIX}`;
-      const endMarker = `${MARKER_END_PREFIX}${idx}${MARKER_SUFFIX}`;
-
-      // Pandoc duplicates inline image alt-text into <wp:docPr descr="...">
-      // metadata attributes AND into the visible caption paragraph. A naive
-      // indexOf hits the metadata-attribute occurrence first, where there is
-      // no <w:t> element so dissectRun fails. Skip occurrences whose position
-      // is inside an XML tag (last unbalanced '<' before position).
-      // See: https://github.com/gcol33/docrev/issues/4
-      function findInTextContent(haystack: string, needle: string, fromIdx = 0): number {
-        let i = fromIdx;
-        while (true) {
-          const p = haystack.indexOf(needle, i);
-          if (p < 0) return -1;
-          const lastLt = haystack.lastIndexOf('<', p);
-          const lastGt = haystack.lastIndexOf('>', p);
-          if (lastLt > lastGt) {
-            i = p + 1;
-            continue;
-          }
-          return p;
-        }
-      }
-
-      const startPos = findInTextContent(documentXml, startMarker);
-      const endPos = startPos === -1
-        ? -1
-        : findInTextContent(documentXml, endMarker, startPos + startMarker.length);
-
-      if (startPos === -1 || endPos === -1) continue;
-
-      // Find the runs containing each marker. Pandoc may split a single
-      // markdown anchor across multiple <w:r> blocks when it applies styling
-      // mid-anchor (smart-quote substitution, *italic*, `code`, **bold**).
-      // The same-run path (current happy path) collapses into the multi-run
-      // path when start and end runs coincide.
-      const startRunOpen = Math.max(
-        documentXml.lastIndexOf('<w:r>', startPos),
-        documentXml.lastIndexOf('<w:r ', startPos),
-      );
-      const startRunCloseIdx = documentXml.indexOf('</w:r>', startPos);
-      const endRunOpen = Math.max(
-        documentXml.lastIndexOf('<w:r>', endPos),
-        documentXml.lastIndexOf('<w:r ', endPos),
-      );
-      const endRunCloseIdx = documentXml.indexOf('</w:r>', endPos);
-
-      if (
-        startRunOpen === -1 || startRunCloseIdx === -1 ||
-        endRunOpen === -1 || endRunCloseIdx === -1
-      ) continue;
-
-      const startRunClose = startRunCloseIdx + '</w:r>'.length;
-      const endRunClose = endRunCloseIdx + '</w:r>'.length;
-
-      const startRunFull = documentXml.slice(startRunOpen, startRunClose);
-      const endRunFull = documentXml.slice(endRunOpen, endRunClose);
-
-      // Extract <w:rPr> and <w:t> element shape from each run. Both pieces
-      // are needed verbatim so a textBefore split keeps its original styling
-      // and so the post-anchor textAfter render keeps the end run's styling.
-      function dissectRun(runXml: string, marker: string): {
-        rPr: string;
-        tElement: string;
-        textBefore: string;
-        textAfter: string;
-      } | null {
-        const rPrMatch = runXml.match(/<w:rPr>[\s\S]*?<\/w:rPr>/);
-        const tMatch = runXml.match(/<w:t[^>]*>([\s\S]*?)<\/w:t>/);
-        if (!tMatch) return null;
-        const tOpenMatch = tMatch[0].match(/<w:t[^>]*>/);
-        if (!tOpenMatch) return null;
-        const tContent = tMatch[1] ?? '';
-        const markerInT = tContent.indexOf(marker);
-        if (markerInT === -1) return null;
-        return {
-          rPr: rPrMatch ? rPrMatch[0] : '',
-          tElement: tOpenMatch[0],
-          textBefore: tContent.slice(0, markerInT),
-          textAfter: tContent.slice(markerInT + marker.length),
-        };
-      }
-
-      let replacement = '';
-      const replies = commentsWithIds.filter(c => c.isReply && c.parentIdx === comment?.commentIdx);
-
-      const emitRangeStarts = () => {
-        replacement += `<w:commentRangeStart w:id="${comment.id}"/>`;
-        for (const reply of replies) {
-          replacement += `<w:commentRangeStart w:id="${reply.id}"/>`;
-        }
-      };
-
-      const emitRangeEnds = () => {
-        replacement += `<w:commentRangeEnd w:id="${comment.id}"/>`;
-        replacement += `<w:r><w:commentReference w:id="${comment.id}"/></w:r>`;
-        for (const reply of replies) {
-          replacement += `<w:commentRangeEnd w:id="${reply.id}"/>`;
-          replacement += `<w:r><w:commentReference w:id="${reply.id}"/></w:r>`;
-          injectedIds.add(reply.id);
-        }
-      };
-
-      if (startRunOpen === endRunOpen) {
-        // Same-run path: both markers live inside one <w:t>. Original logic.
-        const startInfo = dissectRun(startRunFull, startMarker);
-        if (!startInfo) continue;
-        const fullText = startInfo.textBefore + startMarker + startInfo.textAfter;
-        const endInTextRel = startInfo.textAfter.indexOf(endMarker);
-        if (endInTextRel === -1) continue;
-        const anchorTextSame = startInfo.textAfter.slice(0, endInTextRel);
-        let textAfter = startInfo.textAfter.slice(endInTextRel + endMarker.length);
-        let anchorText = anchorTextSame;
-        let textBefore = startInfo.textBefore;
-
-        // Empty anchor: borrow the next word so the comment has something
-        // to anchor on. Then normalize the trailing double space.
-        if (!anchorText && textAfter) {
-          const wordMatch = textAfter.match(/^\s*(\S+)/);
-          if (wordMatch) {
-            anchorText = wordMatch[1] ?? '';
-            textAfter = textAfter.slice(wordMatch[0].length);
-          }
-        }
-        if (!anchorText && textBefore.endsWith(' ') && textAfter.startsWith(' ')) {
-          textAfter = textAfter.slice(1);
-        }
-        // Suppress unused warning for pre-empty-anchor fullText var
-        void fullText;
-
-        if (textBefore) {
-          replacement += `<w:r>${startInfo.rPr}${startInfo.tElement}${textBefore}</w:t></w:r>`;
-        }
-        emitRangeStarts();
-        if (anchorText) {
-          replacement += `<w:r>${startInfo.rPr}${startInfo.tElement}${anchorText}</w:t></w:r>`;
-        }
-        emitRangeEnds();
-        if (textAfter) {
-          replacement += `<w:r>${startInfo.rPr}${startInfo.tElement}${textAfter}</w:t></w:r>`;
-        }
-        documentXml = documentXml.slice(0, startRunOpen) + replacement + documentXml.slice(startRunClose);
-        injectedIds.add(comment.id);
-        continue;
-      }
-
-      // Multi-run path: markers sit in different <w:r> blocks because pandoc
-      // applied mid-anchor styling. Split the start run at the start marker,
-      // keep all middle runs verbatim (they carry the styled anchor portions),
-      // split the end run at the end marker.
-      const startInfo = dissectRun(startRunFull, startMarker);
-      const endInfo = dissectRun(endRunFull, endMarker);
-      if (!startInfo || !endInfo) continue;
-
-      const middle = documentXml.slice(startRunClose, endRunOpen);
-
-      if (startInfo.textBefore) {
-        replacement += `<w:r>${startInfo.rPr}${startInfo.tElement}${startInfo.textBefore}</w:t></w:r>`;
-      }
-      emitRangeStarts();
-      if (startInfo.textAfter) {
-        replacement += `<w:r>${startInfo.rPr}${startInfo.tElement}${startInfo.textAfter}</w:t></w:r>`;
-      }
-      replacement += middle;
-      if (endInfo.textBefore) {
-        replacement += `<w:r>${endInfo.rPr}${endInfo.tElement}${endInfo.textBefore}</w:t></w:r>`;
-      }
-      emitRangeEnds();
-      if (endInfo.textAfter) {
-        replacement += `<w:r>${endInfo.rPr}${endInfo.tElement}${endInfo.textAfter}</w:t></w:r>`;
-      }
-
-      documentXml = documentXml.slice(0, startRunOpen) + replacement + documentXml.slice(endRunClose);
-      injectedIds.add(comment.id);
-    }
-
-    // Add required namespaces to document.xml for comment threading
-    const requiredNs: Record<string, string> = {
-      'xmlns:w14': 'http://schemas.microsoft.com/office/word/2010/wordml',
-      'xmlns:w15': 'http://schemas.microsoft.com/office/word/2012/wordml',
-      'xmlns:w16cid': 'http://schemas.microsoft.com/office/word/2016/wordml/cid',
-      'xmlns:w16cex': 'http://schemas.microsoft.com/office/word/2018/wordml/cex',
-      'xmlns:mc': 'http://schemas.openxmlformats.org/markup-compatibility/2006',
-    };
-
-    // Find <w:document and add namespaces
-    const docTagMatch = documentXml.match(/<w:document[^>]*>/);
-    if (docTagMatch) {
-      let docTag = docTagMatch[0];
-      let modified = false;
-      for (const [attr, val] of Object.entries(requiredNs)) {
-        if (!docTag.includes(attr)) {
-          docTag = docTag.replace('>', ` ${attr}="${val}">`);
-          modified = true;
-        }
-      }
-      // Add mc:Ignorable if mc namespace was added
-      if (modified && !docTag.includes('mc:Ignorable')) {
-        docTag = docTag.replace('>', ' mc:Ignorable="w14 w15 w16cid w16cex">');
-      }
-      documentXml = documentXml.replace(docTagMatch[0], docTag);
-    }
-
-    // Update document.xml
-    zip.updateFile('word/document.xml', Buffer.from(documentXml, 'utf-8'));
-
-    // All comments (parents + replies) go in comments.xml
-    // But only include if parent was injected
-    const includedComments = commentsWithIds.filter(c => {
-      if (!c.isReply) {
-        return injectedIds.has(c.id);
-      } else {
-        // Include reply if its parent was injected
-        const parent = c.parentIdx !== null ? commentsWithIds[c.parentIdx] : undefined;
-        return parent && injectedIds.has(parent.id);
-      }
-    });
-
-    // Create comments.xml
-    const commentsXml = createCommentsXml(includedComments);
-    if (zip.getEntry('word/comments.xml')) {
-      zip.updateFile('word/comments.xml', Buffer.from(commentsXml, 'utf-8'));
-    } else {
-      zip.addFile('word/comments.xml', Buffer.from(commentsXml, 'utf-8'));
-    }
-
-    // Create commentsExtended.xml with reply threading
-    const commentsExtXml = createCommentsExtendedXml(includedComments);
-    if (zip.getEntry('word/commentsExtended.xml')) {
-      zip.updateFile('word/commentsExtended.xml', Buffer.from(commentsExtXml, 'utf-8'));
-    } else {
-      zip.addFile('word/commentsExtended.xml', Buffer.from(commentsExtXml, 'utf-8'));
-    }
-
-    // Create commentsIds.xml (Word 2016+)
-    const commentsIdsXml = createCommentsIdsXml(includedComments);
-    if (zip.getEntry('word/commentsIds.xml')) {
-      zip.updateFile('word/commentsIds.xml', Buffer.from(commentsIdsXml, 'utf-8'));
-    } else {
-      zip.addFile('word/commentsIds.xml', Buffer.from(commentsIdsXml, 'utf-8'));
-    }
-
-    // Create commentsExtensible.xml (Word 2018+)
-    const commentsExtensibleXml = createCommentsExtensibleXml(includedComments);
-    if (zip.getEntry('word/commentsExtensible.xml')) {
-      zip.updateFile('word/commentsExtensible.xml', Buffer.from(commentsExtensibleXml, 'utf-8'));
-    } else {
-      zip.addFile('word/commentsExtensible.xml', Buffer.from(commentsExtensibleXml, 'utf-8'));
-    }
-
-    // Create people.xml (author definitions with Windows Live IDs)
-    const peopleXml = createPeopleXml(includedComments);
-    if (zip.getEntry('word/people.xml')) {
-      zip.updateFile('word/people.xml', Buffer.from(peopleXml, 'utf-8'));
-    } else {
-      zip.addFile('word/people.xml', Buffer.from(peopleXml, 'utf-8'));
-    }
-
-    // Update [Content_Types].xml
-    const contentTypesEntry = zip.getEntry('[Content_Types].xml');
-    if (contentTypesEntry) {
-      let contentTypes = zip.readAsText(contentTypesEntry);
-
-      if (!contentTypes.includes('comments.xml')) {
-        const insertPoint = contentTypes.lastIndexOf('</Types>');
-        contentTypes = contentTypes.slice(0, insertPoint) +
-          '<Override PartName="/word/comments.xml" ContentType="application/vnd.openxmlformats-officedocument.wordprocessingml.comments+xml"/>\n' +
-          contentTypes.slice(insertPoint);
-      }
-
-      if (!contentTypes.includes('commentsExtended.xml')) {
-        const insertPoint = contentTypes.lastIndexOf('</Types>');
-        contentTypes = contentTypes.slice(0, insertPoint) +
-          '<Override PartName="/word/commentsExtended.xml" ContentType="application/vnd.openxmlformats-officedocument.wordprocessingml.commentsExtended+xml"/>\n' +
-          contentTypes.slice(insertPoint);
-      }
-
-      if (!contentTypes.includes('commentsIds.xml')) {
-        const insertPoint = contentTypes.lastIndexOf('</Types>');
-        contentTypes = contentTypes.slice(0, insertPoint) +
-          '<Override PartName="/word/commentsIds.xml" ContentType="application/vnd.openxmlformats-officedocument.wordprocessingml.commentsIds+xml"/>\n' +
-          contentTypes.slice(insertPoint);
-      }
-
-      if (!contentTypes.includes('commentsExtensible.xml')) {
-        const insertPoint = contentTypes.lastIndexOf('</Types>');
-        contentTypes = contentTypes.slice(0, insertPoint) +
-          '<Override PartName="/word/commentsExtensible.xml" ContentType="application/vnd.openxmlformats-officedocument.wordprocessingml.commentsExtensible+xml"/>\n' +
-          contentTypes.slice(insertPoint);
-      }
-
-      if (!contentTypes.includes('people.xml')) {
-        const insertPoint = contentTypes.lastIndexOf('</Types>');
-        contentTypes = contentTypes.slice(0, insertPoint) +
-          '<Override PartName="/word/people.xml" ContentType="application/vnd.openxmlformats-officedocument.wordprocessingml.people+xml"/>\n' +
-          contentTypes.slice(insertPoint);
-      }
-
-      zip.updateFile('[Content_Types].xml', Buffer.from(contentTypes, 'utf-8'));
-    }
-
-    // Update relationships
-    const relsEntry = zip.getEntry('word/_rels/document.xml.rels');
-    if (relsEntry) {
-      let rels = zip.readAsText(relsEntry);
-
-      const rIdMatches = rels.match(/rId(\d+)/g) || [];
-      const maxId = rIdMatches.reduce((max, r) => Math.max(max, parseInt(r.replace('rId', ''))), 0);
-
-      if (!rels.includes('comments.xml')) {
-        const insertPoint = rels.lastIndexOf('</Relationships>');
-        rels = rels.slice(0, insertPoint) +
-          `<Relationship Id="rId${maxId + 1}" Type="http://schemas.openxmlformats.org/officeDocument/2006/relationships/comments" Target="comments.xml"/>\n` +
-          rels.slice(insertPoint);
-      }
-
-      if (!rels.includes('commentsExtended.xml')) {
-        const insertPoint = rels.lastIndexOf('</Relationships>');
-        rels = rels.slice(0, insertPoint) +
-          `<Relationship Id="rId${maxId + 2}" Type="http://schemas.microsoft.com/office/2011/relationships/commentsExtended" Target="commentsExtended.xml"/>\n` +
-          rels.slice(insertPoint);
-      }
-
-      if (!rels.includes('commentsIds.xml')) {
-        const insertPoint = rels.lastIndexOf('</Relationships>');
-        rels = rels.slice(0, insertPoint) +
-          `<Relationship Id="rId${maxId + 3}" Type="http://schemas.microsoft.com/office/2016/09/relationships/commentsIds" Target="commentsIds.xml"/>\n` +
-          rels.slice(insertPoint);
-      }
-
-      if (!rels.includes('commentsExtensible.xml')) {
-        const insertPoint = rels.lastIndexOf('</Relationships>');
-        rels = rels.slice(0, insertPoint) +
-          `<Relationship Id="rId${maxId + 4}" Type="http://schemas.microsoft.com/office/2018/08/relationships/commentsExtensible" Target="commentsExtensible.xml"/>\n` +
-          rels.slice(insertPoint);
-      }
-
-      if (!rels.includes('people.xml')) {
-        const insertPoint = rels.lastIndexOf('</Relationships>');
-        rels = rels.slice(0, insertPoint) +
-          `<Relationship Id="rId${maxId + 5}" Type="http://schemas.microsoft.com/office/2011/relationships/people" Target="people.xml"/>\n` +
-          rels.slice(insertPoint);
-      }
-
-      zip.updateFile('word/_rels/document.xml.rels', Buffer.from(rels, 'utf-8'));
-    }
-
-    zip.writeZip(outputPath);
-
-    const parentCount = includedComments.filter(c => !c.isReply).length;
-    const replyCount = includedComments.filter(c => c.isReply).length;
-
-    return {
-      success: true,
-      commentCount: parentCount,
-      replyCount: replyCount,
-      skippedComments: comments.length - includedComments.length,
-    };
-
-  } catch (err: any) {
-    return { success: false, commentCount: 0, skippedComments: 0, error: err.message };
-  }
-}
+/**
+ * Word comment injection with reply threading
+ *
+ * Flow:
+ * 1. prepareMarkdownWithMarkers() - Parse comments, detect reply relationships
+ *    - First comment in a cluster = parent (gets markers: ⟦CMS:n⟧anchor⟦CME:n⟧)
+ *    - Subsequent adjacent comments = replies (no markers, attach to parent)
+ * 2. Pandoc converts to DOCX
+ * 3. injectCommentsAtMarkers() - Insert comment ranges for parents only
+ *    - Replies go in comments.xml with parent reference in commentsExtended.xml
+ */
+
+import * as fs from 'fs';
+import AdmZip from 'adm-zip';
+import { escapeXml } from './utils.js';
+
+const MARKER_START_PREFIX = '⟦CMS:';
+const MARKER_END_PREFIX = '⟦CME:';
+const MARKER_SUFFIX = '⟧';
+
+interface ParsedComment {
+  author: string;
+  text: string;
+  anchor: string | null;
+  start: number;
+  end: number;
+  fullMatch: string;
+}
+
+interface PreparedComment extends ParsedComment {
+  isReply: boolean;
+  parentIdx: number | null;
+  commentIdx: number;
+  anchorFromReply?: boolean;
+  placesParentMarkers?: boolean;
+}
+
+interface PrepareResult {
+  markedMarkdown: string;
+  comments: PreparedComment[];
+}
+
+interface CommentWithIds extends PreparedComment {
+  id: string;
+  paraId: string;
+  paraId2: string;
+  durableId: string;
+  parentParaId?: string;
+}
+
+interface InjectionResult {
+  success: boolean;
+  commentCount: number;
+  replyCount?: number;
+  skippedComments: number;
+  error?: string;
+}
+
+function generateParaId(commentIdx: number, paraNum: number): string {
+  // Generate 8-character uppercase hex ID matching Word format
+  // Word uses IDs like "3F25BC58", "0331C187"
+  // Must be deterministic - same inputs always produce same output
+  const id = 0x10000000 + (commentIdx * 0x00100000) + (paraNum * 0x00001000);
+  return id.toString(16).toUpperCase().padStart(8, '0');
+}
+
+/**
+ * Parse comments and create markers
+ *
+ * Returns:
+ * - markedMarkdown: markdown with markers for parent comments only
+ * - comments: array with author, text, isReply, parentIdx
+ */
+export function prepareMarkdownWithMarkers(markdown: string): PrepareResult {
+  // Match the comment block first; extend manually to capture an optional
+  // trailing `[anchor]{.mark}` span. A regex `[^\]]+` for the anchor would
+  // bail on the inner `]` of nested syntax (e.g. `[[0..9]]{.mark}` or
+  // `[*phrase*]{.mark}` after pandoc-rewriting), so we walk the brackets
+  // ourselves and verify a `{.mark}` suffix.
+  const commentPattern = /\{>>([\s\S]+?)<<\}/g;
+
+  function tryParseTrailingAnchor(
+    text: string,
+    fromIdx: number,
+  ): { anchor: string; endIdx: number } | null {
+    let i = fromIdx;
+    while (i < text.length && /\s/.test(text[i] ?? '')) i++;
+    if (text[i] !== '[') return null;
+    let depth = 1;
+    let j = i + 1;
+    while (j < text.length) {
+      const ch = text[j];
+      if (ch === '[') depth++;
+      else if (ch === ']') {
+        depth--;
+        if (depth === 0) break;
+      }
+      j++;
+    }
+    if (depth !== 0) return null;
+    if (text.slice(j + 1, j + 8) !== '{.mark}') return null;
+    return { anchor: text.slice(i + 1, j), endIdx: j + 8 };
+  }
+
+  const REPLY_PREFIX = '↪ ';
+  const rawMatches: (ParsedComment & { explicitReply: boolean })[] = [];
+  let match: RegExpExecArray | null;
+  while ((match = commentPattern.exec(markdown)) !== null) {
+    const content = match[1] ?? '';
+    let author = 'Unknown';
+    let text = content;
+    const colonIdx = content.indexOf(':');
+    if (colonIdx > 0 && colonIdx < 30) {
+      author = content.slice(0, colonIdx).trim();
+      text = content.slice(colonIdx + 1).trim();
+    }
+
+    // The `↪ ` prefix is the authoritative reply signal emitted by
+    // `insertCommentsIntoMarkdown`. Strip it from the author before injection
+    // so Word displays the real name.
+    let explicitReply = false;
+    if (author.startsWith(REPLY_PREFIX)) {
+      explicitReply = true;
+      author = author.slice(REPLY_PREFIX.length).trim();
+    }
+
+    const commentEnd = match.index + match[0].length;
+    const trailing = tryParseTrailingAnchor(markdown, commentEnd);
+
+    rawMatches.push({
+      author,
+      text,
+      anchor: trailing ? trailing.anchor : null,
+      start: match.index,
+      end: trailing ? trailing.endIdx : commentEnd,
+      fullMatch: markdown.slice(match.index, trailing ? trailing.endIdx : commentEnd),
+      explicitReply,
+    });
+
+    // Advance regex lastIndex past the consumed anchor so the next iteration
+    // doesn't re-scan inside it (e.g. `[*emphasis*]{.mark}` would otherwise
+    // tempt the matcher to look for another `{>>...<<}` in the body of the
+    // anchor span).
+    if (trailing) {
+      commentPattern.lastIndex = trailing.endIdx;
+    }
+  }
+
+  if (rawMatches.length === 0) {
+    return { markedMarkdown: markdown, comments: [] };
+  }
+
+  // Two-mode reply detection driven by the markdown itself:
+  //   - If any comment carries the `↪ ` author prefix, the markdown came
+  //     through `insertCommentsIntoMarkdown` and we use prefix-only mode.
+  //     Distinct clusters that happen to land at gap=0 (a real failure
+  //     mode on dense reviewer docs — 298-comment paper produced 9 such
+  //     collisions) are not misthreaded.
+  //   - If no comment carries the prefix, the markdown was hand-typed.
+  //     Fall back to gap < 10 adjacency for backward compat with users
+  //     who write CriticMarkup directly.
+  const ADJACENT_THRESHOLD = 10;
+  const useExplicitMode = rawMatches.some(m => m.explicitReply);
+  const comments: PreparedComment[] = [];
+  let clusterParentIdx = -1;  // Index of first comment in current cluster
+  let lastCommentEnd = -1;
+
+  for (let i = 0; i < rawMatches.length; i++) {
+    const m = rawMatches[i];
+    if (!m) continue;
+
+    const gap = lastCommentEnd >= 0 ? m.start - lastCommentEnd : Infinity;
+    const isAdjacent = useExplicitMode
+      ? m.explicitReply
+      : gap < ADJACENT_THRESHOLD;
+
+    // Reset cluster if there's a gap (comments not in same cluster)
+    if (!isAdjacent) {
+      clusterParentIdx = -1;
+    }
+
+    if (clusterParentIdx === -1) {
+      // First comment in cluster = parent (regardless of author)
+      comments.push({
+        author: m.author,
+        text: m.text,
+        anchor: m.anchor,
+        start: m.start,
+        end: m.end,
+        fullMatch: m.fullMatch,
+        isReply: false,
+        parentIdx: null,
+        commentIdx: comments.length
+      });
+      clusterParentIdx = comments.length - 1;
+    } else {
+      // Subsequent comment in cluster = reply to first comment
+      comments.push({
+        author: m.author,
+        text: m.text,
+        anchor: m.anchor,
+        start: m.start,
+        end: m.end,
+        fullMatch: m.fullMatch,
+        isReply: true,
+        parentIdx: clusterParentIdx,
+        commentIdx: comments.length
+      });
+    }
+
+    lastCommentEnd = m.end;
+  }
+
+  // Propagate anchors from replies to parents
+  // If a reply has an anchor but its parent doesn't, move the anchor to the parent
+  // Track flags for special handling during marker generation
+  for (const c of comments) {
+    if (c.isReply && c.anchor && c.parentIdx !== null) {
+      const parent = comments[c.parentIdx];
+      if (parent && !parent.anchor) {
+        parent.anchor = c.anchor;
+        parent.anchorFromReply = true;  // Parent's anchor came from a reply (markers placed by reply)
+        c.placesParentMarkers = true;   // This reply should place the parent's markers
+        c.anchor = null;
+      }
+    }
+  }
+
+  // Build marked markdown - only parent comments get markers
+  // Process from end to start to preserve positions
+  let markedMarkdown = markdown;
+
+  for (let i = comments.length - 1; i >= 0; i--) {
+    const c = comments[i];
+    if (!c) continue;
+
+    if (c.isReply) {
+      // Reply: remove from document entirely (will be in comments.xml only)
+      // Also consume one preceding whitespace char to avoid double spaces.
+      // We deliberately consume at most one — walking arbitrarily backwards
+      // would shift positions that lower-index comments still depend on.
+      let removeStart = c.start;
+      if (removeStart > 0 && /\s/.test(markedMarkdown[removeStart - 1] ?? '')) {
+        removeStart--;
+      }
+
+      // If this reply places parent's markers (anchor was propagated)
+      if (c.placesParentMarkers && c.parentIdx !== null) {
+        // Extract anchor text from the original match
+        const anchorMatch = c.fullMatch.match(/\[([^\]]+)\]\{\.mark\}$/);
+        if (anchorMatch) {
+          const anchorText = anchorMatch[1] ?? '';
+          // Output markers with PARENT's index around the anchor text
+          const parentIdx = c.parentIdx;
+          const replacement = `${MARKER_START_PREFIX}${parentIdx}${MARKER_SUFFIX}${anchorText}${MARKER_END_PREFIX}${parentIdx}${MARKER_SUFFIX}`;
+          markedMarkdown = markedMarkdown.slice(0, removeStart) + replacement + markedMarkdown.slice(c.end);
+        } else {
+          markedMarkdown = markedMarkdown.slice(0, removeStart) + markedMarkdown.slice(c.end);
+        }
+      } else {
+        markedMarkdown = markedMarkdown.slice(0, removeStart) + markedMarkdown.slice(c.end);
+      }
+    } else {
+      // Parent comment
+      if (c.anchorFromReply) {
+        // Anchor markers are placed by the reply, just remove this comment.
+        // Consume one preceding whitespace char only (see reply branch above).
+        let removeStart = c.start;
+        if (removeStart > 0 && /\s/.test(markedMarkdown[removeStart - 1] ?? '')) {
+          removeStart--;
+        }
+        markedMarkdown = markedMarkdown.slice(0, removeStart) + markedMarkdown.slice(c.end);
+      } else {
+        // Normal case: replace with markers
+        const anchor = c.anchor || '';
+        const replacement = `${MARKER_START_PREFIX}${i}${MARKER_SUFFIX}${anchor}${MARKER_END_PREFIX}${i}${MARKER_SUFFIX}`;
+        markedMarkdown = markedMarkdown.slice(0, c.start) + replacement + markedMarkdown.slice(c.end);
+      }
+    }
+  }
+
+  return { markedMarkdown, comments };
+}
+
+function createCommentsXml(comments: CommentWithIds[]): string {
+  // Word expects date without milliseconds: 2025-12-30T08:33:00Z
+  const now = new Date().toISOString().replace(/\.\d{3}Z$/, 'Z');
+
+  let xml = '<?xml version="1.0" encoding="UTF-8" standalone="yes"?>\n';
+  // Minimal namespaces matching golden file structure
+  xml += '<w:comments xmlns:mc="http://schemas.openxmlformats.org/markup-compatibility/2006" xmlns:w="http://schemas.openxmlformats.org/wordprocessingml/2006/main" xmlns:w14="http://schemas.microsoft.com/office/word/2010/wordml" xmlns:w15="http://schemas.microsoft.com/office/word/2012/wordml" mc:Ignorable="w14 w15">';
+
+  // Use a consistent rsid (8-char hex) for all comments in this batch
+  const rsid = '00' + (Date.now() % 0xFFFFFF).toString(16).toUpperCase().padStart(6, '0');
+
+  for (const comment of comments) {
+    xml += `<w:comment w:id="${comment.id}" w:author="${escapeXml(comment.author)}" w:date="${now}" w:initials="${comment.author.split(' ').map(n => n[0]).join('')}">`;
+    // First paragraph: rsidRDefault="00000000", annotationRef without rStyle wrapper
+    xml += `<w:p w14:paraId="${comment.paraId}" w14:textId="77777777" w:rsidR="${rsid}" w:rsidRDefault="00000000">`;
+    xml += `<w:r><w:annotationRef/></w:r>`;
+    xml += `<w:r><w:t>${escapeXml(comment.text)}</w:t></w:r>`;
+    xml += `</w:p>`;
+    if (comment.isReply) {
+      // Second empty paragraph: rsidRDefault matches rsidR
+      xml += `<w:p w14:paraId="${comment.paraId2}" w14:textId="77777777" w:rsidR="${rsid}" w:rsidRDefault="${rsid}"/>`;
+    }
+    xml += `</w:comment>`;
+  }
+
+  xml += '</w:comments>';
+  return xml;
+}
+
+function createCommentsExtendedXml(comments: CommentWithIds[]): string {
+  let xml = '<?xml version="1.0" encoding="UTF-8" standalone="yes"?>\n';
+  // Minimal namespaces matching golden file structure
+  xml += '<w15:commentsEx xmlns:mc="http://schemas.openxmlformats.org/markup-compatibility/2006" xmlns:w="http://schemas.openxmlformats.org/wordprocessingml/2006/main" xmlns:w14="http://schemas.microsoft.com/office/word/2010/wordml" xmlns:w15="http://schemas.microsoft.com/office/word/2012/wordml" mc:Ignorable="w14 w15">';
+
+  for (const comment of comments) {
+    if (comment.isReply && comment.parentParaId) {
+      // Reply: use paraId2 (the second/empty paragraph) and link to parent's paraId
+      xml += `<w15:commentEx w15:paraId="${comment.paraId2}" w15:paraIdParent="${comment.parentParaId}" w15:done="0"/>`;
+    } else {
+      // Parent comment: use paraId (first paragraph)
+      xml += `<w15:commentEx w15:paraId="${comment.paraId}" w15:done="0"/>`;
+    }
+  }
+
+  xml += '</w15:commentsEx>';
+  return xml;
+}
+
+function generateDurableId(index: number): string {
+  // Generate unique 8-char hex ID for durableId
+  // CRITICAL: Must stay within signed 32-bit range (< 0x7FFFFFFF = 2147483647)
+  // Word interprets durableIds as signed 32-bit integers
+  const base = 0x10000000 + (Date.now() % 0x40000000); // Base between 0x10000000 and 0x50000000
+  const id = (base + index * 0x01000000) % 0x7FFFFFFF; // Keep under signed 32-bit max
+  return id.toString(16).toUpperCase().padStart(8, '0');
+}
+
+function createCommentsIdsXml(comments: CommentWithIds[]): string {
+  let xml = '<?xml version="1.0" encoding="UTF-8" standalone="yes"?>\n';
+  // Minimal namespaces matching golden file structure
+  xml += '<w16cid:commentsIds ';
+  xml += 'xmlns:mc="http://schemas.openxmlformats.org/markup-compatibility/2006" ';
+  xml += 'xmlns:w16cid="http://schemas.microsoft.com/office/word/2016/wordml/cid" ';
+  xml += 'mc:Ignorable="w16cid">';
+
+  for (const comment of comments) {
+    // ONE entry per comment using the LAST paragraph's paraId:
+    // - Parent comments (1 paragraph): use paraId
+    // - Reply comments (2 paragraphs): use paraId2 (the second/empty paragraph)
+    const useParaId = comment.isReply ? comment.paraId2 : comment.paraId;
+    xml += `<w16cid:commentId w16cid:paraId="${useParaId}" w16cid:durableId="${comment.durableId}"/>`;
+  }
+
+  xml += '</w16cid:commentsIds>';
+  return xml;
+}
+
+function createCommentsExtensibleXml(comments: CommentWithIds[]): string {
+  const now = new Date().toISOString().replace(/\.\d{3}Z$/, 'Z');
+
+  let xml = '<?xml version="1.0" encoding="UTF-8" standalone="yes"?>\n';
+  // Minimal namespaces matching golden file structure
+  xml += '<w16cex:commentsExtensible ';
+  xml += 'xmlns:mc="http://schemas.openxmlformats.org/markup-compatibility/2006" ';
+  xml += 'xmlns:w16cex="http://schemas.microsoft.com/office/word/2018/wordml/cex" ';
+  xml += 'mc:Ignorable="w16cex">';
+
+  for (const comment of comments) {
+    // ONE entry per comment using the durableId
+    xml += `<w16cex:commentExtensible w16cex:durableId="${comment.durableId}" w16cex:dateUtc="${now}"/>`;
+  }
+
+  xml += '</w16cex:commentsExtensible>';
+  return xml;
+}
+
+// Generate deterministic user IDs for authors (no hardcoded personal data)
+
+function createPeopleXml(comments: CommentWithIds[]): string {
+  // Extract unique authors
+  const authors = [...new Set(comments.map(c => c.author))];
+
+  let xml = '<?xml version="1.0" encoding="UTF-8" standalone="yes"?>\n';
+  xml += '<w15:people ';
+  xml += 'xmlns:mc="http://schemas.openxmlformats.org/markup-compatibility/2006" ';
+  xml += 'xmlns:x="http://schemas.openxmlformats.org/spreadsheetml/2006/main" ';
+  xml += 'xmlns:r="http://schemas.openxmlformats.org/officeDocument/2006/relationships" ';
+  xml += 'xmlns:w="http://schemas.openxmlformats.org/wordprocessingml/2006/main" ';
+  xml += 'xmlns:w14="http://schemas.microsoft.com/office/word/2010/wordml" ';
+  xml += 'xmlns:w15="http://schemas.microsoft.com/office/word/2012/wordml" ';
+  xml += 'xmlns:w16cex="http://schemas.microsoft.com/office/word/2018/wordml/cex" ';
+  xml += 'xmlns:w16cid="http://schemas.microsoft.com/office/word/2016/wordml/cid" ';
+  xml += 'xmlns:w16="http://schemas.microsoft.com/office/word/2018/wordml" ';
+  xml += 'xmlns:w16sdtdh="http://schemas.microsoft.com/office/word/2020/wordml/sdtdatahash" ';
+  xml += 'xmlns:w16se="http://schemas.microsoft.com/office/word/2015/wordml/symex" ';
+  xml += 'mc:Ignorable="w14 w15 w16se w16cid w16 w16cex w16sdtdh">';
+
+  for (const author of authors) {
+    const userId = generateUserId(author);
+    xml += `<w15:person w15:author="${escapeXml(author)}">`;
+    xml += `<w15:presenceInfo w15:providerId="Windows Live" w15:userId="${userId}"/>`;
+    xml += `</w15:person>`;
+  }
+
+  xml += '</w15:people>';
+  return xml;
+}
+
+function generateUserId(author: string): string {
+  // Generate a deterministic 16-char hex ID from author name
+  let hash = 0;
+  for (let i = 0; i < author.length; i++) {
+    hash = ((hash << 5) - hash) + author.charCodeAt(i);
+    hash = hash & hash;
+  }
+  return Math.abs(hash).toString(16).padStart(16, '0').slice(0, 16);
+}
+
+/**
+ * Inject comments at marker positions
+ */
+export async function injectCommentsAtMarkers(
+  docxPath: string,
+  comments: PreparedComment[],
+  outputPath: string
+): Promise<InjectionResult> {
+  try {
+    if (!fs.existsSync(docxPath)) {
+      return { success: false, commentCount: 0, skippedComments: 0, error: `File not found: ${docxPath}` };
+    }
+
+    if (comments.length === 0) {
+      fs.copyFileSync(docxPath, outputPath);
+      return { success: true, commentCount: 0, skippedComments: 0 };
+    }
+
+    const zip = new AdmZip(docxPath);
+    const documentEntry = zip.getEntry('word/document.xml');
+    if (!documentEntry) {
+      return { success: false, commentCount: 0, skippedComments: 0, error: 'Invalid DOCX: no document.xml' };
+    }
+
+    let documentXml = zip.readAsText(documentEntry);
+
+    // Assign IDs and paraIds (IDs start at 1, not 0 - Word convention)
+    const commentsWithIds: CommentWithIds[] = comments.map((c, idx) => ({
+      ...c,
+      id: String(idx + 1),
+      paraId: generateParaId(idx, 1),       // First paragraph (e.g., 10000001)
+      paraId2: generateParaId(idx, 2),      // Second paragraph (e.g., 10000002)
+      durableId: generateDurableId(idx),    // Unique ID for commentsIds/commentsExtensible
+    }));
+
+    // Link replies to parent paraIds
+    for (const c of commentsWithIds) {
+      if (c.isReply && c.parentIdx !== null) {
+        const parent = commentsWithIds[c.parentIdx];
+        if (parent) {
+          c.parentParaId = parent.paraId;
+        }
+      }
+    }
+
+    const injectedIds = new Set<string>();
+
+    // Process only parent comments (non-replies) for document ranges
+    const parentComments = commentsWithIds.filter(c => !c.isReply);
+
+    for (let i = parentComments.length - 1; i >= 0; i--) {
+      const comment = parentComments[i];
+      if (!comment) continue;
+      const idx = comment.commentIdx;
+
+      const startMarker = `${MARKER_START_PREFIX}${idx}${MARKER_SUFFIX}`;
+      const endMarker = `${MARKER_END_PREFIX}${idx}${MARKER_SUFFIX}`;
+
+      // Pandoc duplicates inline image alt-text into <wp:docPr descr="...">
+      // metadata attributes AND into the visible caption paragraph. A naive
+      // indexOf hits the metadata-attribute occurrence first, where there is
+      // no <w:t> element so dissectRun fails. Skip occurrences whose position
+      // is inside an XML tag (last unbalanced '<' before position).
+      // See: https://github.com/gcol33/docrev/issues/4
+      function findInTextContent(haystack: string, needle: string, fromIdx = 0): number {
+        let i = fromIdx;
+        while (true) {
+          const p = haystack.indexOf(needle, i);
+          if (p < 0) return -1;
+          const lastLt = haystack.lastIndexOf('<', p);
+          const lastGt = haystack.lastIndexOf('>', p);
+          if (lastLt > lastGt) {
+            i = p + 1;
+            continue;
+          }
+          return p;
+        }
+      }
+
+      const startPos = findInTextContent(documentXml, startMarker);
+      const endPos = startPos === -1
+        ? -1
+        : findInTextContent(documentXml, endMarker, startPos + startMarker.length);
+
+      if (startPos === -1 || endPos === -1) continue;
+
+      // Find the runs containing each marker. Pandoc may split a single
+      // markdown anchor across multiple <w:r> blocks when it applies styling
+      // mid-anchor (smart-quote substitution, *italic*, `code`, **bold**).
+      // The same-run path (current happy path) collapses into the multi-run
+      // path when start and end runs coincide.
+      const startRunOpen = Math.max(
+        documentXml.lastIndexOf('<w:r>', startPos),
+        documentXml.lastIndexOf('<w:r ', startPos),
+      );
+      const startRunCloseIdx = documentXml.indexOf('</w:r>', startPos);
+      const endRunOpen = Math.max(
+        documentXml.lastIndexOf('<w:r>', endPos),
+        documentXml.lastIndexOf('<w:r ', endPos),
+      );
+      const endRunCloseIdx = documentXml.indexOf('</w:r>', endPos);
+
+      if (
+        startRunOpen === -1 || startRunCloseIdx === -1 ||
+        endRunOpen === -1 || endRunCloseIdx === -1
+      ) continue;
+
+      const startRunClose = startRunCloseIdx + '</w:r>'.length;
+      const endRunClose = endRunCloseIdx + '</w:r>'.length;
+
+      const startRunFull = documentXml.slice(startRunOpen, startRunClose);
+      const endRunFull = documentXml.slice(endRunOpen, endRunClose);
+
+      // Extract <w:rPr> and <w:t> element shape from each run. Both pieces
+      // are needed verbatim so a textBefore split keeps its original styling
+      // and so the post-anchor textAfter render keeps the end run's styling.
+      function dissectRun(runXml: string, marker: string): {
+        rPr: string;
+        tElement: string;
+        textBefore: string;
+        textAfter: string;
+      } | null {
+        const rPrMatch = runXml.match(/<w:rPr>[\s\S]*?<\/w:rPr>/);
+        const tMatch = runXml.match(/<w:t[^>]*>([\s\S]*?)<\/w:t>/);
+        if (!tMatch) return null;
+        const tOpenMatch = tMatch[0].match(/<w:t[^>]*>/);
+        if (!tOpenMatch) return null;
+        const tContent = tMatch[1] ?? '';
+        const markerInT = tContent.indexOf(marker);
+        if (markerInT === -1) return null;
+        return {
+          rPr: rPrMatch ? rPrMatch[0] : '',
+          tElement: tOpenMatch[0],
+          textBefore: tContent.slice(0, markerInT),
+          textAfter: tContent.slice(markerInT + marker.length),
+        };
+      }
+
+      let replacement = '';
+      const replies = commentsWithIds.filter(c => c.isReply && c.parentIdx === comment?.commentIdx);
+
+      const emitRangeStarts = () => {
+        replacement += `<w:commentRangeStart w:id="${comment.id}"/>`;
+        for (const reply of replies) {
+          replacement += `<w:commentRangeStart w:id="${reply.id}"/>`;
+        }
+      };
+
+      const emitRangeEnds = () => {
+        replacement += `<w:commentRangeEnd w:id="${comment.id}"/>`;
+        replacement += `<w:r><w:commentReference w:id="${comment.id}"/></w:r>`;
+        for (const reply of replies) {
+          replacement += `<w:commentRangeEnd w:id="${reply.id}"/>`;
+          replacement += `<w:r><w:commentReference w:id="${reply.id}"/></w:r>`;
+          injectedIds.add(reply.id);
+        }
+      };
+
+      if (startRunOpen === endRunOpen) {
+        // Same-run path: both markers live inside one <w:t>. Original logic.
+        const startInfo = dissectRun(startRunFull, startMarker);
+        if (!startInfo) continue;
+        const fullText = startInfo.textBefore + startMarker + startInfo.textAfter;
+        const endInTextRel = startInfo.textAfter.indexOf(endMarker);
+        if (endInTextRel === -1) continue;
+        const anchorTextSame = startInfo.textAfter.slice(0, endInTextRel);
+        let textAfter = startInfo.textAfter.slice(endInTextRel + endMarker.length);
+        let anchorText = anchorTextSame;
+        let textBefore = startInfo.textBefore;
+
+        // Empty anchor: borrow the next word so the comment has something
+        // to anchor on. Then normalize the trailing double space.
+        if (!anchorText && textAfter) {
+          const wordMatch = textAfter.match(/^\s*(\S+)/);
+          if (wordMatch) {
+            anchorText = wordMatch[1] ?? '';
+            textAfter = textAfter.slice(wordMatch[0].length);
+          }
+        }
+        if (!anchorText && textBefore.endsWith(' ') && textAfter.startsWith(' ')) {
+          textAfter = textAfter.slice(1);
+        }
+        // Suppress unused warning for pre-empty-anchor fullText var
+        void fullText;
+
+        if (textBefore) {
+          replacement += `<w:r>${startInfo.rPr}${startInfo.tElement}${textBefore}</w:t></w:r>`;
+        }
+        emitRangeStarts();
+        if (anchorText) {
+          replacement += `<w:r>${startInfo.rPr}${startInfo.tElement}${anchorText}</w:t></w:r>`;
+        }
+        emitRangeEnds();
+        if (textAfter) {
+          replacement += `<w:r>${startInfo.rPr}${startInfo.tElement}${textAfter}</w:t></w:r>`;
+        }
+        documentXml = documentXml.slice(0, startRunOpen) + replacement + documentXml.slice(startRunClose);
+        injectedIds.add(comment.id);
+        continue;
+      }
+
+      // Multi-run path: markers sit in different <w:r> blocks because pandoc
+      // applied mid-anchor styling. Split the start run at the start marker,
+      // keep all middle runs verbatim (they carry the styled anchor portions),
+      // split the end run at the end marker.
+      const startInfo = dissectRun(startRunFull, startMarker);
+      const endInfo = dissectRun(endRunFull, endMarker);
+      if (!startInfo || !endInfo) continue;
+
+      const middle = documentXml.slice(startRunClose, endRunOpen);
+
+      if (startInfo.textBefore) {
+        replacement += `<w:r>${startInfo.rPr}${startInfo.tElement}${startInfo.textBefore}</w:t></w:r>`;
+      }
+      emitRangeStarts();
+      if (startInfo.textAfter) {
+        replacement += `<w:r>${startInfo.rPr}${startInfo.tElement}${startInfo.textAfter}</w:t></w:r>`;
+      }
+      replacement += middle;
+      if (endInfo.textBefore) {
+        replacement += `<w:r>${endInfo.rPr}${endInfo.tElement}${endInfo.textBefore}</w:t></w:r>`;
+      }
+      emitRangeEnds();
+      if (endInfo.textAfter) {
+        replacement += `<w:r>${endInfo.rPr}${endInfo.tElement}${endInfo.textAfter}</w:t></w:r>`;
+      }
+
+      documentXml = documentXml.slice(0, startRunOpen) + replacement + documentXml.slice(endRunClose);
+      injectedIds.add(comment.id);
+    }
+
+    // Add required namespaces to document.xml for comment threading
+    const requiredNs: Record<string, string> = {
+      'xmlns:w14': 'http://schemas.microsoft.com/office/word/2010/wordml',
+      'xmlns:w15': 'http://schemas.microsoft.com/office/word/2012/wordml',
+      'xmlns:w16cid': 'http://schemas.microsoft.com/office/word/2016/wordml/cid',
+      'xmlns:w16cex': 'http://schemas.microsoft.com/office/word/2018/wordml/cex',
+      'xmlns:mc': 'http://schemas.openxmlformats.org/markup-compatibility/2006',
+    };
+
+    // Find <w:document and add namespaces
+    const docTagMatch = documentXml.match(/<w:document[^>]*>/);
+    if (docTagMatch) {
+      let docTag = docTagMatch[0];
+      let modified = false;
+      for (const [attr, val] of Object.entries(requiredNs)) {
+        if (!docTag.includes(attr)) {
+          docTag = docTag.replace('>', ` ${attr}="${val}">`);
+          modified = true;
+        }
+      }
+      // Add mc:Ignorable if mc namespace was added
+      if (modified && !docTag.includes('mc:Ignorable')) {
+        docTag = docTag.replace('>', ' mc:Ignorable="w14 w15 w16cid w16cex">');
+      }
+      documentXml = documentXml.replace(docTagMatch[0], docTag);
+    }
+
+    // Update document.xml
+    zip.updateFile('word/document.xml', Buffer.from(documentXml, 'utf-8'));
+
+    // All comments (parents + replies) go in comments.xml
+    // But only include if parent was injected
+    const includedComments = commentsWithIds.filter(c => {
+      if (!c.isReply) {
+        return injectedIds.has(c.id);
+      } else {
+        // Include reply if its parent was injected
+        const parent = c.parentIdx !== null ? commentsWithIds[c.parentIdx] : undefined;
+        return parent && injectedIds.has(parent.id);
+      }
+    });
+
+    // Create comments.xml
+    const commentsXml = createCommentsXml(includedComments);
+    if (zip.getEntry('word/comments.xml')) {
+      zip.updateFile('word/comments.xml', Buffer.from(commentsXml, 'utf-8'));
+    } else {
+      zip.addFile('word/comments.xml', Buffer.from(commentsXml, 'utf-8'));
+    }
+
+    // Create commentsExtended.xml with reply threading
+    const commentsExtXml = createCommentsExtendedXml(includedComments);
+    if (zip.getEntry('word/commentsExtended.xml')) {
+      zip.updateFile('word/commentsExtended.xml', Buffer.from(commentsExtXml, 'utf-8'));
+    } else {
+      zip.addFile('word/commentsExtended.xml', Buffer.from(commentsExtXml, 'utf-8'));
+    }
+
+    // Create commentsIds.xml (Word 2016+)
+    const commentsIdsXml = createCommentsIdsXml(includedComments);
+    if (zip.getEntry('word/commentsIds.xml')) {
+      zip.updateFile('word/commentsIds.xml', Buffer.from(commentsIdsXml, 'utf-8'));
+    } else {
+      zip.addFile('word/commentsIds.xml', Buffer.from(commentsIdsXml, 'utf-8'));
+    }
+
+    // Create commentsExtensible.xml (Word 2018+)
+    const commentsExtensibleXml = createCommentsExtensibleXml(includedComments);
+    if (zip.getEntry('word/commentsExtensible.xml')) {
+      zip.updateFile('word/commentsExtensible.xml', Buffer.from(commentsExtensibleXml, 'utf-8'));
+    } else {
+      zip.addFile('word/commentsExtensible.xml', Buffer.from(commentsExtensibleXml, 'utf-8'));
+    }
+
+    // Create people.xml (author definitions with Windows Live IDs)
+    const peopleXml = createPeopleXml(includedComments);
+    if (zip.getEntry('word/people.xml')) {
+      zip.updateFile('word/people.xml', Buffer.from(peopleXml, 'utf-8'));
+    } else {
+      zip.addFile('word/people.xml', Buffer.from(peopleXml, 'utf-8'));
+    }
+
+    // Update [Content_Types].xml
+    const contentTypesEntry = zip.getEntry('[Content_Types].xml');
+    if (contentTypesEntry) {
+      let contentTypes = zip.readAsText(contentTypesEntry);
+
+      if (!contentTypes.includes('comments.xml')) {
+        const insertPoint = contentTypes.lastIndexOf('</Types>');
+        contentTypes = contentTypes.slice(0, insertPoint) +
+          '<Override PartName="/word/comments.xml" ContentType="application/vnd.openxmlformats-officedocument.wordprocessingml.comments+xml"/>\n' +
+          contentTypes.slice(insertPoint);
+      }
+
+      if (!contentTypes.includes('commentsExtended.xml')) {
+        const insertPoint = contentTypes.lastIndexOf('</Types>');
+        contentTypes = contentTypes.slice(0, insertPoint) +
+          '<Override PartName="/word/commentsExtended.xml" ContentType="application/vnd.openxmlformats-officedocument.wordprocessingml.commentsExtended+xml"/>\n' +
+          contentTypes.slice(insertPoint);
+      }
+
+      if (!contentTypes.includes('commentsIds.xml')) {
+        const insertPoint = contentTypes.lastIndexOf('</Types>');
+        contentTypes = contentTypes.slice(0, insertPoint) +
+          '<Override PartName="/word/commentsIds.xml" ContentType="application/vnd.openxmlformats-officedocument.wordprocessingml.commentsIds+xml"/>\n' +
+          contentTypes.slice(insertPoint);
+      }
+
+      if (!contentTypes.includes('commentsExtensible.xml')) {
+        const insertPoint = contentTypes.lastIndexOf('</Types>');
+        contentTypes = contentTypes.slice(0, insertPoint) +
+          '<Override PartName="/word/commentsExtensible.xml" ContentType="application/vnd.openxmlformats-officedocument.wordprocessingml.commentsExtensible+xml"/>\n' +
+          contentTypes.slice(insertPoint);
+      }
+
+      if (!contentTypes.includes('people.xml')) {
+        const insertPoint = contentTypes.lastIndexOf('</Types>');
+        contentTypes = contentTypes.slice(0, insertPoint) +
+          '<Override PartName="/word/people.xml" ContentType="application/vnd.openxmlformats-officedocument.wordprocessingml.people+xml"/>\n' +
+          contentTypes.slice(insertPoint);
+      }
+
+      zip.updateFile('[Content_Types].xml', Buffer.from(contentTypes, 'utf-8'));
+    }
+
+    // Update relationships
+    const relsEntry = zip.getEntry('word/_rels/document.xml.rels');
+    if (relsEntry) {
+      let rels = zip.readAsText(relsEntry);
+
+      const rIdMatches = rels.match(/rId(\d+)/g) || [];
+      const maxId = rIdMatches.reduce((max, r) => Math.max(max, parseInt(r.replace('rId', ''))), 0);
+
+      if (!rels.includes('comments.xml')) {
+        const insertPoint = rels.lastIndexOf('</Relationships>');
+        rels = rels.slice(0, insertPoint) +
+          `<Relationship Id="rId${maxId + 1}" Type="http://schemas.openxmlformats.org/officeDocument/2006/relationships/comments" Target="comments.xml"/>\n` +
+          rels.slice(insertPoint);
+      }
+
+      if (!rels.includes('commentsExtended.xml')) {
+        const insertPoint = rels.lastIndexOf('</Relationships>');
+        rels = rels.slice(0, insertPoint) +
+          `<Relationship Id="rId${maxId + 2}" Type="http://schemas.microsoft.com/office/2011/relationships/commentsExtended" Target="commentsExtended.xml"/>\n` +
+          rels.slice(insertPoint);
+      }
+
+      if (!rels.includes('commentsIds.xml')) {
+        const insertPoint = rels.lastIndexOf('</Relationships>');
+        rels = rels.slice(0, insertPoint) +
+          `<Relationship Id="rId${maxId + 3}" Type="http://schemas.microsoft.com/office/2016/09/relationships/commentsIds" Target="commentsIds.xml"/>\n` +
+          rels.slice(insertPoint);
+      }
+
+      if (!rels.includes('commentsExtensible.xml')) {
+        const insertPoint = rels.lastIndexOf('</Relationships>');
+        rels = rels.slice(0, insertPoint) +
+          `<Relationship Id="rId${maxId + 4}" Type="http://schemas.microsoft.com/office/2018/08/relationships/commentsExtensible" Target="commentsExtensible.xml"/>\n` +
+          rels.slice(insertPoint);
+      }
+
+      if (!rels.includes('people.xml')) {
+        const insertPoint = rels.lastIndexOf('</Relationships>');
+        rels = rels.slice(0, insertPoint) +
+          `<Relationship Id="rId${maxId + 5}" Type="http://schemas.microsoft.com/office/2011/relationships/people" Target="people.xml"/>\n` +
+          rels.slice(insertPoint);
+      }
+
+      zip.updateFile('word/_rels/document.xml.rels', Buffer.from(rels, 'utf-8'));
+    }
+
+    zip.writeZip(outputPath);
+
+    const parentCount = includedComments.filter(c => !c.isReply).length;
+    const replyCount = includedComments.filter(c => c.isReply).length;
+
+    return {
+      success: true,
+      commentCount: parentCount,
+      replyCount: replyCount,
+      skippedComments: comments.length - includedComments.length,
+    };
+
+  } catch (err: any) {
+    return { success: false, commentCount: 0, skippedComments: 0, error: err.message };
+  }
+}