docrev 0.9.13 → 0.9.15

package/dev_notes/stress2/build_adversarial.ts

@@ -1,186 +1,186 @@
-/**
- * Build an adversarial DOCX with hand-crafted commentRangeStart/End
- * markers that probe whether docrev locates comments by *real*
- * anchor (the highlighted run text, taken from the docx XML) or by
- * guesswork. Each comment is designed to expose a specific failure
- * mode if the matcher were not anchor-aware.
- */
-import AdmZip from 'adm-zip';
-import path from 'path';
-
-const out = path.resolve('dev_notes/stress2/adversarial.docx');
-
-// Sentence with: a triplicated word, unicode look-alikes, mid-word
-// hyphenation, and embedded XML-special chars in the *prose* (escaped).
-const paragraphs: { text: string; comments?: { id: string; on: string }[] }[] = [
-  // heading
-  { text: 'Methods', comments: [] },
-
-  // p0: single highlight, easy
-  { text: 'The methods section describes our approach in detail.',
-    comments: [{ id: '0', on: 'methods section' }] },
-
-  // p1: word "model" appears 3x — anchor must pin the SECOND occurrence
-  { text: 'The model is fitted, the model is checked, the model is reported.',
-    comments: [{ id: '1', on: 'model' /* second occurrence */ }] },
-
-  // p2: empty anchor (zero-width comment, classic Word "insert here")
-  { text: 'Reviewers often place a marker between two sentences. Like this. Then continue.',
-    comments: [{ id: '2', on: '' }] },
-
-  // p3: anchor spanning across multiple runs (formatting break)
-  { text: 'Our results show that p < 0.001 across all conditions.',
-    comments: [{ id: '3', on: 'p < 0.001' }] },
-
-  // p4: text containing literal angle brackets that would break a naive
-  // [^<]* regex — anchor includes "(<1825)"
-  { text: 'Trade volumes pre-industrial (<1825) were modest.',
-    comments: [{ id: '4', on: '(<1825)' }] },
-
-  // p5: very long anchor (full sentence) — matcher must still place it
-  { text: 'We used a hierarchical Bayesian model with weakly informative priors and Hamiltonian Monte Carlo sampling implemented in Stan.',
-    comments: [{ id: '5', on: 'We used a hierarchical Bayesian model with weakly informative priors and Hamiltonian Monte Carlo sampling implemented in Stan.' }] },
-
-  // p6: anchor on punctuation only — should fall back to context
-  { text: 'See Table 1 for details; numbers are rounded.',
-    comments: [{ id: '6', on: ';' }] },
-
-  // p7: two overlapping comment ranges share a word
-  { text: 'The overrepresented species are listed in Appendix A.',
-    comments: [
-      { id: '7', on: 'overrepresented species' },
-      { id: '8', on: 'overrepresented' },
-    ] },
-
-  // p8: anchor that exists VERBATIM elsewhere in the doc — context disambiguates
-  { text: 'The cohort was small. Limitations are discussed in Section 5.',
-    comments: [{ id: '9', on: 'small' }] },
-  { text: 'The effect was small but significant.',
-    comments: [{ id: '10', on: 'small' }] },
-
-  // p9: heading paragraph (<w:pStyle w:val="Heading1"/>)
-  { text: 'Discussion', comments: [], /* heading marker handled below */ },
-  { text: 'In this section we situate the findings in prior literature.',
-    comments: [{ id: '11', on: 'situate the findings' }] },
-];
-
-function escapeXml(s: string): string {
-  return s.replace(/&/g, '&amp;').replace(/</g, '&lt;').replace(/>/g, '&gt;')
-          .replace(/"/g, '&quot;').replace(/'/g, '&apos;');
-}
-
-function buildParagraphXml(p: typeof paragraphs[0], isHeading: boolean): string {
-  const text = p.text;
-  const comments = p.comments || [];
-
-  if (comments.length === 0) {
-    return `<w:p>${isHeading ? '<w:pPr><w:pStyle w:val="Heading1"/></w:pPr>' : ''}<w:r><w:t xml:space="preserve">${escapeXml(text)}</w:t></w:r></w:p>`;
-  }
-
-  // Plan ranges: each comment marks [start..end] character offsets into `text`.
-  // Empty anchors place start=end at first sentence boundary.
-  type Range = { id: string; start: number; end: number };
-  const ranges: Range[] = [];
-  for (const c of comments) {
-    if (c.on === '') {
-      const dotIdx = text.indexOf('.');
-      const pos = dotIdx >= 0 ? dotIdx + 1 : 0;
-      ranges.push({ id: c.id, start: pos, end: pos });
-      continue;
-    }
-    if (c.id === '1') {
-      // "model" — pick the SECOND occurrence
-      const first = text.indexOf('model');
-      const second = text.indexOf('model', first + 1);
-      ranges.push({ id: c.id, start: second, end: second + 'model'.length });
-      continue;
-    }
-    const start = text.indexOf(c.on);
-    if (start < 0) throw new Error(`anchor not found: ${c.on}`);
-    ranges.push({ id: c.id, start, end: start + c.on.length });
-  }
-
-  // Build event list: at each char boundary we may need to emit
-  // <w:commentRangeStart/> or <w:commentRangeEnd/>.
-  type Event = { pos: number; kind: 'start' | 'end'; id: string };
-  const events: Event[] = [];
-  for (const r of ranges) {
-    events.push({ pos: r.start, kind: 'start', id: r.id });
-    events.push({ pos: r.end, kind: 'end', id: r.id });
-  }
-  // Emit ends before starts at same position, so an empty anchor's start=end
-  // ordering keeps a zero-width range
-  events.sort((a, b) => a.pos - b.pos || (a.kind === 'end' ? -1 : 1));
-
-  let xml = `<w:p>${isHeading ? '<w:pPr><w:pStyle w:val="Heading1"/></w:pPr>' : ''}`;
-  let cursor = 0;
-  for (const ev of events) {
-    if (ev.pos > cursor) {
-      xml += `<w:r><w:t xml:space="preserve">${escapeXml(text.slice(cursor, ev.pos))}</w:t></w:r>`;
-      cursor = ev.pos;
-    }
-    xml += ev.kind === 'start'
-      ? `<w:commentRangeStart w:id="${ev.id}"/>`
-      : `<w:commentRangeEnd w:id="${ev.id}"/><w:r><w:commentReference w:id="${ev.id}"/></w:r>`;
-  }
-  if (cursor < text.length) {
-    xml += `<w:r><w:t xml:space="preserve">${escapeXml(text.slice(cursor))}</w:t></w:r>`;
-  }
-  xml += `</w:p>`;
-  return xml;
-}
-
-const headingTexts = new Set(['Methods', 'Discussion']);
-const documentBody = paragraphs.map(p => buildParagraphXml(p, headingTexts.has(p.text))).join('');
-
-const documentXml = `<?xml version="1.0" encoding="UTF-8" standalone="yes"?>
-<w:document xmlns:w="http://schemas.openxmlformats.org/wordprocessingml/2006/main">
-  <w:body>
-    ${documentBody}
-    <w:sectPr/>
-  </w:body>
-</w:document>`;
-
-const allComments = paragraphs.flatMap(p => p.comments || []);
-
-const commentsXml = `<?xml version="1.0" encoding="UTF-8" standalone="yes"?>
-<w:comments xmlns:w="http://schemas.openxmlformats.org/wordprocessingml/2006/main">
-${allComments.map(c => `  <w:comment w:id="${c.id}" w:author="Reviewer" w:initials="R" w:date="2026-01-01T00:00:00Z"><w:p><w:r><w:t>Comment ${c.id} on &quot;${escapeXml(c.on)}&quot; — text contains <embedded brackets and &amp;.</w:t></w:r></w:p></w:comment>`.replace('<embedded', '&lt;embedded')).join('\n')}
-</w:comments>`;
-
-const contentTypesXml = `<?xml version="1.0" encoding="UTF-8" standalone="yes"?>
-<Types xmlns="http://schemas.openxmlformats.org/package/2006/content-types">
-  <Default Extension="rels" ContentType="application/vnd.openxmlformats-package.relationships+xml"/>
-  <Default Extension="xml" ContentType="application/xml"/>
-  <Override PartName="/word/document.xml" ContentType="application/vnd.openxmlformats-officedocument.wordprocessingml.document.main+xml"/>
-  <Override PartName="/word/comments.xml" ContentType="application/vnd.openxmlformats-officedocument.wordprocessingml.comments+xml"/>
-  <Override PartName="/word/styles.xml" ContentType="application/vnd.openxmlformats-officedocument.wordprocessingml.styles+xml"/>
-</Types>`;
-
-const rootRelsXml = `<?xml version="1.0" encoding="UTF-8" standalone="yes"?>
-<Relationships xmlns="http://schemas.openxmlformats.org/package/2006/relationships">
-  <Relationship Id="rId1" Type="http://schemas.openxmlformats.org/officeDocument/2006/relationships/officeDocument" Target="word/document.xml"/>
-</Relationships>`;
-
-const docRelsXml = `<?xml version="1.0" encoding="UTF-8" standalone="yes"?>
-<Relationships xmlns="http://schemas.openxmlformats.org/package/2006/relationships">
-  <Relationship Id="rId1" Type="http://schemas.openxmlformats.org/officeDocument/2006/relationships/comments" Target="comments.xml"/>
-  <Relationship Id="rId2" Type="http://schemas.openxmlformats.org/officeDocument/2006/relationships/styles" Target="styles.xml"/>
-</Relationships>`;
-
-const stylesXml = `<?xml version="1.0" encoding="UTF-8" standalone="yes"?>
-<w:styles xmlns:w="http://schemas.openxmlformats.org/wordprocessingml/2006/main">
-  <w:style w:type="paragraph" w:styleId="Heading1"><w:name w:val="heading 1"/></w:style>
-</w:styles>`;
-
-const zip = new AdmZip();
-zip.addFile('[Content_Types].xml', Buffer.from(contentTypesXml, 'utf8'));
-zip.addFile('_rels/.rels', Buffer.from(rootRelsXml, 'utf8'));
-zip.addFile('word/document.xml', Buffer.from(documentXml, 'utf8'));
-zip.addFile('word/comments.xml', Buffer.from(commentsXml, 'utf8'));
-zip.addFile('word/styles.xml', Buffer.from(stylesXml, 'utf8'));
-zip.addFile('word/_rels/document.xml.rels', Buffer.from(docRelsXml, 'utf8'));
-zip.writeZip(out);
-
-console.log(`wrote ${out}`);
-console.log(`comments: ${allComments.length}`);
+/**
+ * Build an adversarial DOCX with hand-crafted commentRangeStart/End
+ * markers that probe whether docrev locates comments by *real*
+ * anchor (the highlighted run text, taken from the docx XML) or by
+ * guesswork. Each comment is designed to expose a specific failure
+ * mode if the matcher were not anchor-aware.
+ */
+import AdmZip from 'adm-zip';
+import path from 'path';
+
+const out = path.resolve('dev_notes/stress2/adversarial.docx');
+
+// Sentence with: a triplicated word, unicode look-alikes, mid-word
+// hyphenation, and embedded XML-special chars in the *prose* (escaped).
+const paragraphs: { text: string; comments?: { id: string; on: string }[] }[] = [
+  // heading
+  { text: 'Methods', comments: [] },
+
+  // p0: single highlight, easy
+  { text: 'The methods section describes our approach in detail.',
+    comments: [{ id: '0', on: 'methods section' }] },
+
+  // p1: word "model" appears 3x — anchor must pin the SECOND occurrence
+  { text: 'The model is fitted, the model is checked, the model is reported.',
+    comments: [{ id: '1', on: 'model' /* second occurrence */ }] },
+
+  // p2: empty anchor (zero-width comment, classic Word "insert here")
+  { text: 'Reviewers often place a marker between two sentences. Like this. Then continue.',
+    comments: [{ id: '2', on: '' }] },
+
+  // p3: anchor spanning across multiple runs (formatting break)
+  { text: 'Our results show that p < 0.001 across all conditions.',
+    comments: [{ id: '3', on: 'p < 0.001' }] },
+
+  // p4: text containing literal angle brackets that would break a naive
+  // [^<]* regex — anchor includes "(<1825)"
+  { text: 'Trade volumes pre-industrial (<1825) were modest.',
+    comments: [{ id: '4', on: '(<1825)' }] },
+
+  // p5: very long anchor (full sentence) — matcher must still place it
+  { text: 'We used a hierarchical Bayesian model with weakly informative priors and Hamiltonian Monte Carlo sampling implemented in Stan.',
+    comments: [{ id: '5', on: 'We used a hierarchical Bayesian model with weakly informative priors and Hamiltonian Monte Carlo sampling implemented in Stan.' }] },
+
+  // p6: anchor on punctuation only — should fall back to context
+  { text: 'See Table 1 for details; numbers are rounded.',
+    comments: [{ id: '6', on: ';' }] },
+
+  // p7: two overlapping comment ranges share a word
+  { text: 'The overrepresented species are listed in Appendix A.',
+    comments: [
+      { id: '7', on: 'overrepresented species' },
+      { id: '8', on: 'overrepresented' },
+    ] },
+
+  // p8: anchor that exists VERBATIM elsewhere in the doc — context disambiguates
+  { text: 'The cohort was small. Limitations are discussed in Section 5.',
+    comments: [{ id: '9', on: 'small' }] },
+  { text: 'The effect was small but significant.',
+    comments: [{ id: '10', on: 'small' }] },
+
+  // p9: heading paragraph (<w:pStyle w:val="Heading1"/>)
+  { text: 'Discussion', comments: [], /* heading marker handled below */ },
+  { text: 'In this section we situate the findings in prior literature.',
+    comments: [{ id: '11', on: 'situate the findings' }] },
+];
+
+function escapeXml(s: string): string {
+  return s.replace(/&/g, '&amp;').replace(/</g, '&lt;').replace(/>/g, '&gt;')
+          .replace(/"/g, '&quot;').replace(/'/g, '&apos;');
+}
+
+function buildParagraphXml(p: typeof paragraphs[0], isHeading: boolean): string {
+  const text = p.text;
+  const comments = p.comments || [];
+
+  if (comments.length === 0) {
+    return `<w:p>${isHeading ? '<w:pPr><w:pStyle w:val="Heading1"/></w:pPr>' : ''}<w:r><w:t xml:space="preserve">${escapeXml(text)}</w:t></w:r></w:p>`;
+  }
+
+  // Plan ranges: each comment marks [start..end] character offsets into `text`.
+  // Empty anchors place start=end at first sentence boundary.
+  type Range = { id: string; start: number; end: number };
+  const ranges: Range[] = [];
+  for (const c of comments) {
+    if (c.on === '') {
+      const dotIdx = text.indexOf('.');
+      const pos = dotIdx >= 0 ? dotIdx + 1 : 0;
+      ranges.push({ id: c.id, start: pos, end: pos });
+      continue;
+    }
+    if (c.id === '1') {
+      // "model" — pick the SECOND occurrence
+      const first = text.indexOf('model');
+      const second = text.indexOf('model', first + 1);
+      ranges.push({ id: c.id, start: second, end: second + 'model'.length });
+      continue;
+    }
+    const start = text.indexOf(c.on);
+    if (start < 0) throw new Error(`anchor not found: ${c.on}`);
+    ranges.push({ id: c.id, start, end: start + c.on.length });
+  }
+
+  // Build event list: at each char boundary we may need to emit
+  // <w:commentRangeStart/> or <w:commentRangeEnd/>.
+  type Event = { pos: number; kind: 'start' | 'end'; id: string };
+  const events: Event[] = [];
+  for (const r of ranges) {
+    events.push({ pos: r.start, kind: 'start', id: r.id });
+    events.push({ pos: r.end, kind: 'end', id: r.id });
+  }
+  // Emit ends before starts at same position, so an empty anchor's start=end
+  // ordering keeps a zero-width range
+  events.sort((a, b) => a.pos - b.pos || (a.kind === 'end' ? -1 : 1));
+
+  let xml = `<w:p>${isHeading ? '<w:pPr><w:pStyle w:val="Heading1"/></w:pPr>' : ''}`;
+  let cursor = 0;
+  for (const ev of events) {
+    if (ev.pos > cursor) {
+      xml += `<w:r><w:t xml:space="preserve">${escapeXml(text.slice(cursor, ev.pos))}</w:t></w:r>`;
+      cursor = ev.pos;
+    }
+    xml += ev.kind === 'start'
+      ? `<w:commentRangeStart w:id="${ev.id}"/>`
+      : `<w:commentRangeEnd w:id="${ev.id}"/><w:r><w:commentReference w:id="${ev.id}"/></w:r>`;
+  }
+  if (cursor < text.length) {
+    xml += `<w:r><w:t xml:space="preserve">${escapeXml(text.slice(cursor))}</w:t></w:r>`;
+  }
+  xml += `</w:p>`;
+  return xml;
+}
+
+const headingTexts = new Set(['Methods', 'Discussion']);
+const documentBody = paragraphs.map(p => buildParagraphXml(p, headingTexts.has(p.text))).join('');
+
+const documentXml = `<?xml version="1.0" encoding="UTF-8" standalone="yes"?>
+<w:document xmlns:w="http://schemas.openxmlformats.org/wordprocessingml/2006/main">
+  <w:body>
+    ${documentBody}
+    <w:sectPr/>
+  </w:body>
+</w:document>`;
+
+const allComments = paragraphs.flatMap(p => p.comments || []);
+
+const commentsXml = `<?xml version="1.0" encoding="UTF-8" standalone="yes"?>
+<w:comments xmlns:w="http://schemas.openxmlformats.org/wordprocessingml/2006/main">
+${allComments.map(c => `  <w:comment w:id="${c.id}" w:author="Reviewer" w:initials="R" w:date="2026-01-01T00:00:00Z"><w:p><w:r><w:t>Comment ${c.id} on &quot;${escapeXml(c.on)}&quot; — text contains <embedded brackets and &amp;.</w:t></w:r></w:p></w:comment>`.replace('<embedded', '&lt;embedded')).join('\n')}
+</w:comments>`;
+
+const contentTypesXml = `<?xml version="1.0" encoding="UTF-8" standalone="yes"?>
+<Types xmlns="http://schemas.openxmlformats.org/package/2006/content-types">
+  <Default Extension="rels" ContentType="application/vnd.openxmlformats-package.relationships+xml"/>
+  <Default Extension="xml" ContentType="application/xml"/>
+  <Override PartName="/word/document.xml" ContentType="application/vnd.openxmlformats-officedocument.wordprocessingml.document.main+xml"/>
+  <Override PartName="/word/comments.xml" ContentType="application/vnd.openxmlformats-officedocument.wordprocessingml.comments+xml"/>
+  <Override PartName="/word/styles.xml" ContentType="application/vnd.openxmlformats-officedocument.wordprocessingml.styles+xml"/>
+</Types>`;
+
+const rootRelsXml = `<?xml version="1.0" encoding="UTF-8" standalone="yes"?>
+<Relationships xmlns="http://schemas.openxmlformats.org/package/2006/relationships">
+  <Relationship Id="rId1" Type="http://schemas.openxmlformats.org/officeDocument/2006/relationships/officeDocument" Target="word/document.xml"/>
+</Relationships>`;
+
+const docRelsXml = `<?xml version="1.0" encoding="UTF-8" standalone="yes"?>
+<Relationships xmlns="http://schemas.openxmlformats.org/package/2006/relationships">
+  <Relationship Id="rId1" Type="http://schemas.openxmlformats.org/officeDocument/2006/relationships/comments" Target="comments.xml"/>
+  <Relationship Id="rId2" Type="http://schemas.openxmlformats.org/officeDocument/2006/relationships/styles" Target="styles.xml"/>
+</Relationships>`;
+
+const stylesXml = `<?xml version="1.0" encoding="UTF-8" standalone="yes"?>
+<w:styles xmlns:w="http://schemas.openxmlformats.org/wordprocessingml/2006/main">
+  <w:style w:type="paragraph" w:styleId="Heading1"><w:name w:val="heading 1"/></w:style>
+</w:styles>`;
+
+const zip = new AdmZip();
+zip.addFile('[Content_Types].xml', Buffer.from(contentTypesXml, 'utf8'));
+zip.addFile('_rels/.rels', Buffer.from(rootRelsXml, 'utf8'));
+zip.addFile('word/document.xml', Buffer.from(documentXml, 'utf8'));
+zip.addFile('word/comments.xml', Buffer.from(commentsXml, 'utf8'));
+zip.addFile('word/styles.xml', Buffer.from(stylesXml, 'utf8'));
+zip.addFile('word/_rels/document.xml.rels', Buffer.from(docRelsXml, 'utf8'));
+zip.writeZip(out);
+
+console.log(`wrote ${out}`);
+console.log(`comments: ${allComments.length}`);

package/dev_notes/stress2/drift_matcher.ts

@@ -1,62 +1,62 @@
-/**
- * Run anchors from the adversarial docx against a HEAVILY DRIFTED
- * markdown and report match quality per comment.
- */
-import { extractCommentAnchors } from '../../lib/word-extraction.js';
-import { findAnchorInText, classifyStrategy } from '../../lib/anchor-match.js';
-
-const docx = 'dev_notes/stress2/adversarial.docx';
-
-// Drifted markdown — same comments need to land on this.
-//  - p0 paraphrased
-//  - p1 reordered, only TWO occurrences of "model"
-//  - p2 prose lightly edited, sentence boundary preserved
-//  - p3 prose changed slightly: "p < 0.001 (n=412)"
-//  - p4 prose moved (<1825)
-//  - p5 mostly intact
-//  - p6 entire sentence rewritten ("Round numbers are reported in Table 1.")
-//  - p7 unchanged
-//  - p8/p9 unchanged
-//  - p10/p11 unchanged
-const drifted = [
-  '# Methods', '',
-  'Below we describe our methodology in detail.', '',
-  'The model is fitted; later, the model is reported.', '',
-  'Reviewers often place a marker between two sentences. Like this. Then continue.', '',
-  'Our results show that p < 0.001 (n=412) across all conditions.', '',
-  'Trade volumes were modest pre-industrial (<1825).', '',
-  'We used a hierarchical Bayesian model with weakly informative priors and Hamiltonian Monte Carlo sampling implemented in Stan.', '',
-  'Round numbers are reported in Table 1.', '',
-  'The overrepresented species are listed in Appendix A.', '',
-  'The cohort was small. Limitations are discussed in Section 5.',
-  'The effect was small but significant.', '',
-  '# Discussion', '',
-  'In this section we situate the findings in prior literature.',
-].join('\n');
-
-const { anchors } = await extractCommentAnchors(docx);
-
-const sorted = [...anchors.entries()].sort((a, b) => Number(a[0]) - Number(b[0]));
-
-const tally = { clean: 0, drift: 0, 'context-only': 0, ambiguous: 0, unmatched: 0 };
-
-console.log('id │ quality       │ strategy            │ #hits │ anchor');
-console.log('───┼───────────────┼─────────────────────┼───────┼──────────────────────');
-for (const [id, a] of sorted) {
-  const r = findAnchorInText(a.anchor, drifted, a.before, a.after);
-  let q: keyof typeof tally = classifyStrategy(r.strategy, r.occurrences.length) as any;
-  if (r.occurrences.length > 1) q = 'ambiguous';
-  tally[q]++;
-  const tag = a.isEmpty ? '∅' : a.anchor.length > 30 ? a.anchor.slice(0, 27) + '...' : a.anchor;
-  console.log(`${id.padStart(2)} │ ${q.padEnd(13)} │ ${r.strategy.padEnd(19)} │ ${String(r.occurrences.length).padStart(5)} │ ${JSON.stringify(tag)}`);
-}
-
-console.log('\nTally:', tally);
-
-// Also confirm: comment #6 (";") will fail anchor-direct match and force
-// context fallback (because the original sentence was rewritten).
-console.log('\nComment #6 (\";\") — sentence rewritten in drifted md.');
-const c6 = anchors.get('6')!;
-const r6 = findAnchorInText(c6.anchor, drifted, c6.before, c6.after);
-console.log(`  strategy=${r6.strategy}, hits=${r6.occurrences.length}`);
-console.log(`  ${r6.occurrences.length > 0 ? 'placed via fallback' : 'left for manual placement'}`);
+/**
+ * Run anchors from the adversarial docx against a HEAVILY DRIFTED
+ * markdown and report match quality per comment.
+ */
+import { extractCommentAnchors } from '../../lib/word-extraction.js';
+import { findAnchorInText, classifyStrategy } from '../../lib/anchor-match.js';
+
+const docx = 'dev_notes/stress2/adversarial.docx';
+
+// Drifted markdown — same comments need to land on this.
+//  - p0 paraphrased
+//  - p1 reordered, only TWO occurrences of "model"
+//  - p2 prose lightly edited, sentence boundary preserved
+//  - p3 prose changed slightly: "p < 0.001 (n=412)"
+//  - p4 prose moved (<1825)
+//  - p5 mostly intact
+//  - p6 entire sentence rewritten ("Round numbers are reported in Table 1.")
+//  - p7 unchanged
+//  - p8/p9 unchanged
+//  - p10/p11 unchanged
+const drifted = [
+  '# Methods', '',
+  'Below we describe our methodology in detail.', '',
+  'The model is fitted; later, the model is reported.', '',
+  'Reviewers often place a marker between two sentences. Like this. Then continue.', '',
+  'Our results show that p < 0.001 (n=412) across all conditions.', '',
+  'Trade volumes were modest pre-industrial (<1825).', '',
+  'We used a hierarchical Bayesian model with weakly informative priors and Hamiltonian Monte Carlo sampling implemented in Stan.', '',
+  'Round numbers are reported in Table 1.', '',
+  'The overrepresented species are listed in Appendix A.', '',
+  'The cohort was small. Limitations are discussed in Section 5.',
+  'The effect was small but significant.', '',
+  '# Discussion', '',
+  'In this section we situate the findings in prior literature.',
+].join('\n');
+
+const { anchors } = await extractCommentAnchors(docx);
+
+const sorted = [...anchors.entries()].sort((a, b) => Number(a[0]) - Number(b[0]));
+
+const tally = { clean: 0, drift: 0, 'context-only': 0, ambiguous: 0, unmatched: 0 };
+
+console.log('id │ quality       │ strategy            │ #hits │ anchor');
+console.log('───┼───────────────┼─────────────────────┼───────┼──────────────────────');
+for (const [id, a] of sorted) {
+  const r = findAnchorInText(a.anchor, drifted, a.before, a.after);
+  let q: keyof typeof tally = classifyStrategy(r.strategy, r.occurrences.length) as any;
+  if (r.occurrences.length > 1) q = 'ambiguous';
+  tally[q]++;
+  const tag = a.isEmpty ? '∅' : a.anchor.length > 30 ? a.anchor.slice(0, 27) + '...' : a.anchor;
+  console.log(`${id.padStart(2)} │ ${q.padEnd(13)} │ ${r.strategy.padEnd(19)} │ ${String(r.occurrences.length).padStart(5)} │ ${JSON.stringify(tag)}`);
+}
+
+console.log('\nTally:', tally);
+
+// Also confirm: comment #6 (";") will fail anchor-direct match and force
+// context fallback (because the original sentence was rewritten).
+console.log('\nComment #6 (\";\") — sentence rewritten in drifted md.');
+const c6 = anchors.get('6')!;
+const r6 = findAnchorInText(c6.anchor, drifted, c6.before, c6.after);
+console.log(`  strategy=${r6.strategy}, hits=${r6.occurrences.length}`);
+console.log(`  ${r6.occurrences.length > 0 ? 'placed via fallback' : 'left for manual placement'}`);

package/dev_notes/stress2/probe_anchors.ts

@@ -1,35 +1,35 @@
-/**
- * Probe: do extracted anchors come from `<w:commentRangeStart/End>`
- * markers (i.e. real anchors), and does each comment carry a unique
- * docPosition that disambiguates duplicate prose?
- */
-import { extractCommentAnchors } from '../../lib/word-extraction.js';
-
-const docx = 'dev_notes/stress2/adversarial.docx';
-const { anchors, fullDocText } = await extractCommentAnchors(docx);
-
-console.log(`Doc text length: ${fullDocText.length}`);
-console.log(`Anchors extracted: ${anchors.size}\n`);
-
-const sorted = [...anchors.entries()].sort((a, b) => Number(a[0]) - Number(b[0]));
-
-for (const [id, a] of sorted) {
-  const tag = a.isEmpty ? '∅(empty)' : JSON.stringify(a.anchor);
-  console.log(`#${id}  pos=${String(a.docPosition).padStart(4)}  anchor=${tag}`);
-  if (a.before) console.log(`       before=${JSON.stringify(a.before.slice(-40))}`);
-  if (a.after)  console.log(`       after =${JSON.stringify(a.after.slice(0, 40))}`);
-}
-
-// Sanity: comment 1 should pin the SECOND occurrence of "model"
-const c1 = anchors.get('1')!;
-const firstModel = fullDocText.indexOf('model');
-const secondModel = fullDocText.indexOf('model', firstModel + 1);
-console.log(`\nComment #1 expectation: pin SECOND "model" at pos ${secondModel}`);
-console.log(`Comment #1 actual  docPosition=${c1.docPosition}`);
-console.log(`Match: ${c1.docPosition === secondModel ? 'YES (real-anchor)' : 'NO'}`);
-
-// Comment 7 vs 8: overlapping ranges, both anchored on "overrepresented..."
-const c7 = anchors.get('7')!;
-const c8 = anchors.get('8')!;
-console.log(`\nComment #7 anchor: ${JSON.stringify(c7.anchor)} (expected "overrepresented species")`);
-console.log(`Comment #8 anchor: ${JSON.stringify(c8.anchor)} (expected "overrepresented")`);
+/**
+ * Probe: do extracted anchors come from `<w:commentRangeStart/End>`
+ * markers (i.e. real anchors), and does each comment carry a unique
+ * docPosition that disambiguates duplicate prose?
+ */
+import { extractCommentAnchors } from '../../lib/word-extraction.js';
+
+const docx = 'dev_notes/stress2/adversarial.docx';
+const { anchors, fullDocText } = await extractCommentAnchors(docx);
+
+console.log(`Doc text length: ${fullDocText.length}`);
+console.log(`Anchors extracted: ${anchors.size}\n`);
+
+const sorted = [...anchors.entries()].sort((a, b) => Number(a[0]) - Number(b[0]));
+
+for (const [id, a] of sorted) {
+  const tag = a.isEmpty ? '∅(empty)' : JSON.stringify(a.anchor);
+  console.log(`#${id}  pos=${String(a.docPosition).padStart(4)}  anchor=${tag}`);
+  if (a.before) console.log(`       before=${JSON.stringify(a.before.slice(-40))}`);
+  if (a.after)  console.log(`       after =${JSON.stringify(a.after.slice(0, 40))}`);
+}
+
+// Sanity: comment 1 should pin the SECOND occurrence of "model"
+const c1 = anchors.get('1')!;
+const firstModel = fullDocText.indexOf('model');
+const secondModel = fullDocText.indexOf('model', firstModel + 1);
+console.log(`\nComment #1 expectation: pin SECOND "model" at pos ${secondModel}`);
+console.log(`Comment #1 actual  docPosition=${c1.docPosition}`);
+console.log(`Match: ${c1.docPosition === secondModel ? 'YES (real-anchor)' : 'NO'}`);
+
+// Comment 7 vs 8: overlapping ranges, both anchored on "overrepresented..."
+const c7 = anchors.get('7')!;
+const c8 = anchors.get('8')!;
+console.log(`\nComment #7 anchor: ${JSON.stringify(c7.anchor)} (expected "overrepresented species")`);
+console.log(`Comment #8 anchor: ${JSON.stringify(c8.anchor)} (expected "overrepresented")`);

package/dev_notes/stress2/project/discussion.before.md

@@ -1,3 +1,3 @@
-# Discussion
-
-In this section we situate the findings in prior literature.
+# Discussion
+
+In this section we situate the findings in prior literature.

package/dev_notes/stress2/project/discussion.md

@@ -1,3 +1,3 @@
-# Discussion
-
-In this section we {>>Reviewer: Comment 11 on "situate the findings" — text contains <embedded brackets and &.<<}situate the findings in prior literature.
+# Discussion
+
+In this section we {>>Reviewer: Comment 11 on "situate the findings" — text contains <embedded brackets and &.<<}situate the findings in prior literature.

package/dev_notes/stress2/project/methods.before.md

@@ -1,20 +1,20 @@
-# Methods
-
-Below we describe our methodology in detail.
-
-The model is fitted; later, the model is reported.
-
-Reviewers often place a marker between two sentences. Like this. Then continue.
-
-Our results show that p < 0.001 (n=412) across all conditions.
-
-Trade volumes were modest pre-industrial (<1825).
-
-We used a hierarchical Bayesian model with weakly informative priors and Hamiltonian Monte Carlo sampling implemented in Stan.
-
-Round numbers are reported in Table 1.
-
-The overrepresented species are listed in Appendix A.
-
-The cohort was small. Limitations are discussed in Section 5.
-The effect was small but significant.
+# Methods
+
+Below we describe our methodology in detail.
+
+The model is fitted; later, the model is reported.
+
+Reviewers often place a marker between two sentences. Like this. Then continue.
+
+Our results show that p < 0.001 (n=412) across all conditions.
+
+Trade volumes were modest pre-industrial (<1825).
+
+We used a hierarchical Bayesian model with weakly informative priors and Hamiltonian Monte Carlo sampling implemented in Stan.
+
+Round numbers are reported in Table 1.
+
+The overrepresented species are listed in Appendix A.
+
+The cohort was small. Limitations are discussed in Section 5.
+The effect was small but significant.