fix: bullet count under-reports on short bullets and two-column layouts

src/lib/heuristics/pdf-extract.test.ts

@@ -68,4 +68,38 @@ describe("assembleTextFromLines", () => {
   it("returns empty string for empty input (scanned PDFs)", () => {
     expect(assembleTextFromLines([])).toBe("");
   });
+
+  it("splits two-column same-y items so right-column bullets keep line-start position (issue #9)", () => {
+    // Deedy-style two-column layout: left column has education text, right
+    // column has experience bullets, both at the same baseline y. Pre-fix
+    // groupIntoLines merged them into a single PdfLine like
+    //   "BS in Computer Science                            • Led migration"
+    // which dropped the bullet glyph out of line-start position and the
+    // score-side bullet counter missed it. The 50pt column-gap split keeps
+    // each column on its own line.
+    const items = [
+      // Left column: education text starting at x=35
+      item("BS in Computer Science", 35, 200, 1, 120),
+      // Right column: experience bullet starting at x=300 — 145pt gap
+      item("•", 300, 200, 1, 6),
+      item("Led migration of legacy auth system", 315, 200, 1, 200),
+    ];
+    const text = assembleTextFromLines(items);
+    const lines = text.split("\n").filter((l) => l.length > 0);
+    expect(lines).toHaveLength(2);
+    expect(lines[0]).toBe("BS in Computer Science");
+    expect(lines[1]).toMatch(/^•\s+Led migration/);
+  });
+
+  it("does NOT split a normal single-column line at small gaps", () => {
+    // Regression guard: a tightly-laid-out single-column line with normal
+    // word/run spacing (gaps well under 50pt) must stay as one PdfLine,
+    // otherwise we'd fragment every Awesome-CV / OpenResume-style PDF.
+    const items = [
+      item("Jane", 72, 100, 1, 30),
+      item("Smith", 110, 100, 1, 35),       // ~8pt gap
+      item("Senior Engineer", 160, 100, 1, 100),  // ~15pt gap
+    ];
+    expect(assembleTextFromLines(items)).toBe("Jane Smith Senior Engineer");
+  });
 });

src/lib/heuristics/sections.ts

@@ -42,6 +42,19 @@ export interface PdfSection {
 /** Items within this vertical distance (PDF points) are treated as same line. */
 const LINE_Y_EPS = 3.5;
 
+/**
+ * Horizontal gap inside a same-y cluster that flags a column boundary.
+ * Awesome-CV / single-column LaTeX exports produce essentially 0pt gaps
+ * between adjacent items even across `\hfill` alignment, so 50pt is well
+ * above any in-line word/run spacing while comfortably below the column
+ * gaps observed in real two-column resumes (Deedy's experience column
+ * jumps in at ~70pt past the education column edge). Splitting at this
+ * threshold rescues the bullet count on two-column layouts that don't
+ * trigger the `two_column` layout flag (asymmetric 0.33/0.66 splits
+ * like Deedy's slip past `probeTwoColumn`). Issue #9.
+ */
+const COLUMN_GAP_THRESHOLD = 50;
+
 // ── Line grouping ───────────────────────────────────────────────────────────
 
 export function groupIntoLines(items: PdfTextItem[]): PdfLine[] {
@@ -55,21 +68,39 @@ export function groupIntoLines(items: PdfTextItem[]): PdfLine[] {
   const lines: PdfLine[] = [];
   let current: PdfTextItem[] = [];
 
-  const flush = () => {
-    if (current.length === 0) return;
-    current.sort((a, b) => a.x - b.x);
-    const text = mergeItemText(current);
-    const ys = current.map((i) => i.y);
+  /** Build a PdfLine from a contiguous run of items (already x-sorted). */
+  const buildLine = (run: PdfTextItem[]): PdfLine => {
+    const text = mergeItemText(run);
+    const ys = run.map((i) => i.y);
     const avgY = ys.reduce((a, b) => a + b, 0) / ys.length;
-    lines.push({
-      page: current[0].page,
+    return {
+      page: run[0].page,
       y: avgY,
-      x: current[0].x,
-      items: [...current],
+      x: run[0].x,
+      items: [...run],
       text,
-      maxFontSize: Math.max(...current.map((i) => i.fontSize)),
+      maxFontSize: Math.max(...run.map((i) => i.fontSize)),
       allCaps: text.replace(/[^A-Za-z]/g, "").length > 0 && text === text.toUpperCase(),
-    });
+    };
+  };
+
+  const flush = () => {
+    if (current.length === 0) return;
+    current.sort((a, b) => a.x - b.x);
+    // Split the same-y cluster at column-sized horizontal gaps so two-column
+    // layouts that share a baseline don't get merged into one PdfLine — see
+    // COLUMN_GAP_THRESHOLD and issue #9.
+    let runStart = 0;
+    for (let i = 1; i < current.length; i++) {
+      const prev = current[i - 1];
+      const cur = current[i];
+      const gap = cur.x - (prev.x + prev.width);
+      if (gap > COLUMN_GAP_THRESHOLD) {
+        lines.push(buildLine(current.slice(runStart, i)));
+        runStart = i;
+      }
+    }
+    lines.push(buildLine(current.slice(runStart)));
     current = [];
   };
 

src/lib/score/score.test.ts

@@ -285,8 +285,36 @@ describe("computeAnonymousAtsScore", () => {
     expect(result.structure.score).toBe(0);
   });
 
+  it("counts short bullets in the visible total even when too short to score well (issue #9)", () => {
+    // Issue #9: previously ANON_BULLET_MIN_WORDS = 4 silently dropped short
+    // bullets like "• Everything that matters." (3 words after marker
+    // strip), so the displayed bullet count under-reported what the user
+    // could see in the PDF. The fix counts every marker-prefixed line and
+    // lets the well-formed length window flag short bullets in per-bullet
+    // feedback rather than hiding them.
+    const mixedLengths = [
+      "- Led migration of 3 microservices reducing latency by 40%",
+      "- Managed team of 5 engineers shipping weekly releases",
+      "- Reduced infrastructure cost by 35% through right-sizing efforts",
+      "- Increased conversion rate by 22% through experimentation",
+      "- Built CI pipeline cutting deploy time from 45 to 8 minutes",
+      "- Everything that matters.", // 3 words — pre-fix this was dropped
+    ].join("\n");
+    const result = computeAnonymousAtsScore(
+      makeAnonInput({ rawText: mixedLengths }),
+    );
+    expect(result.specificity.totalBullets).toBe(6);
+    expect(result.bullets?.length).toBe(6);
+    // The short bullet adds 1 to the denominator without contributing to
+    // metric/structure numerators, so quality scores drop slightly — that's
+    // the correct grading, previously masked by hiding the bullet.
+    expect(result.specificity.metricBullets).toBe(5);
+  });
+
   it("drops Specificity proportional to non-metric bullets", () => {
-    // Each line has ≥4 words after the marker so all six register as bullets.
+    // Each line has multiple words after the marker so all six register as
+    // bullets. (Min words to count as a bullet is 1; quality grading is
+    // handled by the length-window check, not the count filter.)
     const fewMetrics = [
       "- Reduced latency by 40%",
       "- Built a thing for users to enjoy",

src/lib/score/score.ts

@@ -469,11 +469,16 @@ export interface AnonymousAtsScoreInput {
 
 const ANON_CONTACT_CONFIDENCE_FLOOR = 0.5;
 const ANON_MIN_BULLETS_TO_GRADE = 3;
-/** Word-count floor for raw-text bullet extraction. The authed splitBullets
- *  uses a char floor instead because its input is already a curated
- *  description string; raw-text extraction needs to skip headers / one-line
- *  section labels that share a leading marker. */
-const ANON_BULLET_MIN_WORDS = 4;
+/** Word-count floor for raw-text bullet extraction. Set to 1 so the displayed
+ *  bullet count matches what the user can see in the PDF — every line that
+ *  begins with a recognised marker AND has at least one non-empty word is a
+ *  bullet. Quality grading (well-formed length window, action verb, metric)
+ *  is handled downstream by `scoreBulletPool` and `analyzeBullets`, which
+ *  naturally penalise short bullets without hiding them from the count.
+ *  Issue #9 — previously set to 4, which silently dropped legitimate short
+ *  bullets like "• Everything that matters." and caused bullet count to
+ *  under-report vs. the visible PDF. */
+const ANON_BULLET_MIN_WORDS = 1;
 
 const ANON_CONTACT_FIELDS: readonly {
   key: "full_name" | "email" | "phone" | "location" | "linkedin_url";
@@ -489,7 +494,11 @@ const ANON_CONTACT_FIELDS: readonly {
 /**
  * Pull bullet-like lines out of raw resume text. A line counts as a bullet
  * when it starts with a recognized bullet marker (`-`, `•`, etc.) or a
- * numbered list prefix and contains 4+ words after the marker is stripped.
+ * numbered list prefix and the stripped line contains at least one word.
+ * The displayed count is meant to match what a reader sees in the PDF —
+ * short or low-quality bullets are still bullets, they just get flagged
+ * by the downstream length / verb / metric checks in `scoreBulletPool`
+ * and `analyzeBullets`.
  *
  * We deliberately do NOT try to grade unmarked indented lines — most modern
  * resumes use markers, and grading paragraphs would either over-count

tests/fixtures/pdfs/latex/awesome-cv-cv.expected.json

@@ -34,21 +34,21 @@
     "sectionSource": "markdown"
   },
   "score": {
-    "overall": 59,
-    "preLayoutOverall": 59,
+    "overall": 58,
+    "preLayoutOverall": 58,
     "specificity": {
-      "score": 12,
+      "score": 11,
       "max": 40,
       "gradable": true,
       "metricBullets": 10,
-      "totalBullets": 58
+      "totalBullets": 59
     },
     "structure": {
       "score": 20,
       "max": 30,
       "gradable": true,
       "goodBullets": 39,
-      "totalBullets": 58
+      "totalBullets": 59
     },
     "completeness": {
       "score": 27,
@@ -63,7 +63,7 @@
       "multiplier": 1,
       "scanned": false
     },
-    "bulletCount": 58,
+    "bulletCount": 59,
     "algoVersion": "1.0"
   }
 }

tests/fixtures/pdfs/latex/awesome-cv-resume.expected.json

@@ -31,21 +31,21 @@
     "sectionSource": "markdown"
   },
   "score": {
-    "overall": 64,
-    "preLayoutOverall": 64,
+    "overall": 63,
+    "preLayoutOverall": 63,
     "specificity": {
-      "score": 18,
+      "score": 17,
       "max": 40,
       "gradable": true,
       "metricBullets": 8,
-      "totalBullets": 30
+      "totalBullets": 31
     },
     "structure": {
       "score": 23,
       "max": 30,
       "gradable": true,
       "goodBullets": 24,
-      "totalBullets": 30
+      "totalBullets": 31
     },
     "completeness": {
       "score": 23,
@@ -61,7 +61,7 @@
       "multiplier": 1,
       "scanned": false
     },
-    "bulletCount": 30,
+    "bulletCount": 31,
     "algoVersion": "1.0"
   }
 }

tests/fixtures/pdfs/latex/deedy-resume-macfonts.expected.json

@@ -9,18 +9,19 @@
     ],
     "suggestedEscalation": "ocr",
     "fieldsPopulated": [
+      "current_company",
+      "current_title",
       "email",
+      "experience",
       "family_name",
       "full_name",
-      "github_url",
       "given_name",
-      "linkedin_url",
-      "location",
       "phone",
+      "skills",
       "website_url"
     ],
-    "skillsCount": 0,
-    "experienceCount": 0,
+    "skillsCount": 25,
+    "experienceCount": 6,
     "educationCount": 0,
     "rawTextCharCount": 3304,
     "pageCount": 1,
@@ -29,39 +30,39 @@
     "sectionSource": "regex"
   },
   "score": {
-    "overall": 17,
-    "preLayoutOverall": 17,
+    "overall": 47,
+    "preLayoutOverall": 47,
     "specificity": {
-      "score": 0,
+      "score": 8,
       "max": 40,
-      "gradable": false,
-      "metricBullets": 0,
-      "totalBullets": 1
+      "gradable": true,
+      "metricBullets": 1,
+      "totalBullets": 8
     },
     "structure": {
-      "score": 0,
+      "score": 21,
       "max": 30,
-      "gradable": false,
-      "goodBullets": 1,
-      "totalBullets": 1
+      "gradable": true,
+      "goodBullets": 6,
+      "totalBullets": 8
     },
     "completeness": {
-      "score": 17,
+      "score": 18,
       "max": 30,
       "gradable": true,
       "missing": [
+        "LinkedIn",
         "education",
-        "skills",
-        "summary",
-        "work experience"
+        "location",
+        "summary"
       ]
     },
     "layout": {
       "triggers": [],
       "multiplier": 1,
       "scanned": false
     },
-    "bulletCount": 1,
+    "bulletCount": 8,
     "algoVersion": "1.0"
   }
 }