Three Ways Multi-Column PDF Parsing Breaks (and the Fixes for Each)

TLDR

Histogram-based column split detection fails in three predictable ways: page-frame outer border rectangles claiming all text items as giant BOX regions, full-width section headers covering column gutters in X-coverage histograms, and pre-column BOX classifications claiming right-column text items before column detection runs. Resolving these defects requires page-frame dimensions guards, post-split full-width item rescue loops, and full-sample fallback re-detection passes.

Defect Pattern	Failure Mechanism	Structural Remediation
1. Page-Frame Claim	Outer border classified as giant BOX	Guard: (bx < 4%W && bw > 65%W) \
2. Header Gutter Mask	Wide headers block coverage zero-gaps	Rescue: Scan fullWidthIndices after split found
3. Pre-Column Box Claim	Side annotation boxes claim right col text	Fallback: Rerun _detectPageColumns on all textMeta

---

Technical defect remediation patterns

1. Page-Frame outer border guard

Filter out page outer border rectangles before BOX region assignment:

// Filter out full-page border rectangles from BOX classification
export function isPageFrameRectangle(rectX, rectWidth, viewportWidth) {
  return (rectX < viewportWidth  0.04 && rectWidth > viewportWidth  0.65) ||
         rectWidth > viewportWidth * 0.88;
}

---

2. Post-Split full-width index rescue loop

Re-evaluate items initially classified as full-width headers once a column split coordinate is identified:

export function rescueFalseFullWidthItems(fullWidthIndices, textMetaList, splitX) {
  for (let i = fullWidthIndices.length - 1; i >= 0; i--) {
    const item = textMetaList[fullWidthIndices[i]];
    const rightEdge = item.vx + (item.vWidth || 0);
if (rightEdge < splitX) {       fullWidthIndices.splice(i, 1);       item.columnIndex = 0; // Assigned to left column     } else if (item.vx > splitX) {       fullWidthIndices.splice(i, 1);       item.columnIndex = 1; // Assigned to right column     }   } }

---

3. Full-Sample fallback column re-detection

If unclaimed text items yield no column splits, rerun column detection over the complete text sample:

export function executeColumnDetectionFallback(unclaimedMeta, allTextMeta, viewport, scale) {
  let { splits } = _detectPageColumns(unclaimedMeta, viewport, scale);
// Fallback: If unclaimed items yield zero splits but full text set is significantly larger   if (splits.length === 0 && allTextMeta.length > unclaimedMeta.length + 4) {     const fallbackResult = _detectPageColumns(allTextMeta, viewport, scale);     splits = fallbackResult.splits;   }
return splits; }

Rule of thumb: Filter out page-frame rectangles, rescue single-column items from fullWidthIndices post-split, and fall back to full-sample text arrays if unclaimed items yield no column splits.