Post-Mortem: Diagnosing a PDF Pipeline That Works for Two PDFs and Fails on the Third

TLDR

Stress-testing a 2-column layout detector against a LaTeX paper revealed zero splits, initially blamed on math glyph baseline calibration errors (modeFont vs medianFont). Diagnostic instrumentation proved font baselines diverged by only $0.1\text{pt}$, proving that the failure was caused by single display-math text runs bridging the column gutter ($X \approx 310$). Adding a $vWidth > S \times 4$ filter isolates paragraph text, resolving the defect in two lines.

Diagnostic Test Case	Font Calibration Divergence	Fallback Gutter Evaluation	Split Detection Result
Unfiltered Fallback	$0.1\text{pt}$ (Negligible)	Blocked by $500\text{px}$ display math	0 Splits (Defective)
Math-Filtered Fallback	$0.1\text{pt}$ (Negligible)	Clear gutter at $X \approx 310$	1 Split (Correct)

---

Technical diagnostic protocol & findings

1. Verify Baseline Calibration: Compare modeFont vs medianFont across pages; reject calibration theories if divergence is $\le 0.5\text{pt}$.
2. Inspect Raw Item Extents: Log item width distributions (vWidth) to detect anomalously wide text blocks.
3. Isolate Gutter-Crossing Noise: Apply scale-relative width thresholds (vWidth <= S * 4) prior to running candidate coordinate evaluations.

// Scale-Relative Width Pre-Filter for Fallback Column Detection
export function isolateNarrowParagraphItems(textMeta, bodyFontPt) {
  const maxWidthPx = bodyFontPt  4  1.333; // Account for CSS scale factor
  return textMeta.filter(item => item.vWidth <= maxWidthPx);
}

Rule of thumb: Instrument diagnostic logging to measure font baseline divergence before refactoring calibration algorithms for multi-column documents.