fix(heuristics): drop empty-title (date-only) entry blocks before scoring (#145)

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

@@ -888,3 +888,53 @@ describe("extractAchievements", () => {
     expect(extractAchievements(undefined)).toEqual({ value: [], confidence: 0 });
   });
 });
+
+// A header/anchor line whose entire text is a date leaves nothing after
+// stripDateRange, so the block carries no title. Such a phantom must be dropped
+// before it reaches the score, not emitted as a `{ title: "" }` entry (#145).
+describe("empty-title (date-only) block filtering (#145)", () => {
+  it("drops a date-only bullet in an achievements flat list", () => {
+    // Flat-list fallback path: a bullet whose only text is a year.
+    expect(extractAchievements(mkSection("achievements", [{ text: "• 2023" }]))).toEqual({
+      value: [],
+      confidence: 0,
+    });
+  });
+
+  it("drops a date-only header in achievements (buildEntryBlock path)", () => {
+    expect(extractAchievements(mkSection("achievements", [{ text: "2023" }]))).toEqual({
+      value: [],
+      confidence: 0,
+    });
+  });
+
+  it("drops a date-only project, leaving no name-less entry", () => {
+    expect(extractProjects(mkSection("projects", [{ text: "• 2023" }]))).toEqual({
+      value: [],
+      confidence: 0,
+    });
+  });
+
+  it("drops a date-only experience block with neither title nor company", () => {
+    expect(extractExperience(mkSection("experience", [{ text: "2020 - 2022" }]))).toEqual({
+      value: [],
+      confidence: 0,
+    });
+  });
+
+  it("keeps the real entry and confidence is not diluted by a dropped phantom", () => {
+    const section = mkSection("experience", [
+      { text: "Acme Corporation" },
+      { text: "Senior Software Engineer" },
+      { text: "Jan 2020 - Mar 2022" },
+      { text: "• Shipped the thing" },
+      { text: "2018 - 2019" }, // date-only phantom anchor, no header above
+    ]);
+    const { value, confidence } = extractExperience(section);
+    expect(value).toHaveLength(1);
+    expect(value[0].company).toBe("Acme Corporation");
+    expect(value[0].title).toBe("Senior Software Engineer");
+    // Phantom (score 0) is gone, so the average reflects only the real entry.
+    expect(confidence).toBeGreaterThan(0.8);
+  });
+});

src/lib/heuristics/extract/achievements.ts

@@ -6,7 +6,7 @@ import type { PdfSection } from "../sections.ts";
 import { parseEntryBlocks } from "../entry-blocks.ts";
 import type { EntryBlock } from "../entry-blocks.ts";
 import { YEAR_RE } from "../regex.ts";
-import { firstMatch, avgScore } from "./shared.ts";
+import { firstMatch, finalizeEntries } from "./shared.ts";
 import { liftHeaderLabel } from "./projects.ts";
 
 // ── Achievements ──────────────────────────────────────────────────────────────
@@ -39,12 +39,11 @@ export function extractAchievements(
     anchor: "first_line",
     collectBody: true,
   });
-  if (blocks.length === 0) return { value: [], confidence: 0 };
-  const built = blocks.map(achievementFromBlock);
-  return {
-    value: built.map((b) => b.entry),
-    confidence: avgScore(built.map((b) => b.score)),
-  };
+  // Drop any date-only / title-less block (#145) before scoring.
+  return finalizeEntries(
+    blocks.map(achievementFromBlock),
+    (e) => e.title !== "",
+  );
 }
 
 /** Map one entry block to a `HeuristicAchievement` and its confidence score.

src/lib/heuristics/extract/experience.ts

@@ -6,7 +6,7 @@ import type { PdfSection } from "../sections.ts";
 import { parseEntryBlocks } from "../entry-blocks.ts";
 import type { EntryBlock } from "../entry-blocks.ts";
 import { US_LOCATION_RE, INTL_LOCATION_RE } from "../regex.ts";
-import { looksLikeTitle, looksLikeCompany, avgScore } from "./shared.ts";
+import { looksLikeTitle, looksLikeCompany, finalizeEntries } from "./shared.ts";
 
 // ── Experience ──────────────────────────────────────────────────────────────
 
@@ -35,12 +35,12 @@ export function extractExperience(
     collectBody: true,
     headerLookback: 2,
   });
-  if (blocks.length === 0) return { value: [], confidence: 0 };
-  const built = blocks.map(experienceFromBlock);
-  return {
-    value: built.map((b) => b.entry),
-    confidence: avgScore(built.map((b) => b.score)),
-  };
+  // Drop a date-only phantom — a block with neither title nor company (#145).
+  // Experience has no single title axis, so we keep a role that has either.
+  return finalizeEntries(
+    blocks.map(experienceFromBlock),
+    (e) => e.title !== "" || e.company !== "",
+  );
 }
 
 /** Map one dated entry block to a `ResumeExperience` and its confidence score.

src/lib/heuristics/extract/projects.ts

@@ -6,7 +6,7 @@ import type { PdfSection } from "../sections.ts";
 import { parseEntryBlocks } from "../entry-blocks.ts";
 import type { EntryBlock } from "../entry-blocks.ts";
 import { URL_RE } from "../regex.ts";
-import { firstMatch, avgScore } from "./shared.ts";
+import { firstMatch, finalizeEntries } from "./shared.ts";
 
 // ── Projects ────────────────────────────────────────────────────────────────
 
@@ -36,12 +36,8 @@ export function extractProjects(
     anchor: "first_line",
     collectBody: true,
   });
-  if (blocks.length === 0) return { value: [], confidence: 0 };
-  const built = blocks.map(projectFromBlock);
-  return {
-    value: built.map((b) => b.entry),
-    confidence: avgScore(built.map((b) => b.score)),
-  };
+  // Drop any date-only / name-less block (#145) before scoring.
+  return finalizeEntries(blocks.map(projectFromBlock), (e) => e.name !== "");
 }
 
 /**

src/lib/heuristics/extract/shared.ts

@@ -56,3 +56,23 @@ export function looksLikeCompany(text: string): boolean {
 export function avgScore(scores: number[]): number {
   return scores.reduce((a, b) => a + b, 0) / Math.max(scores.length, 1);
 }
+
+/**
+ * Drop entries the parser couldn't label — a date-only / title-less block whose
+ * header reduced to "" (#145) — then package the survivors as the standard
+ * `{ value, confidence }` pair the three entry extractors return. Filtering on
+ * the built entry's label (not on empty `headerLines`) also catches a URL-only
+ * header, which `liftHeaderLabel` collapses to an empty label. Keeping the
+ * phantom out of the list also keeps its score 0 out of the `avgScore`
+ * denominator, so it no longer dilutes section confidence.
+ */
+export function finalizeEntries<T>(
+  built: { entry: T; score: number }[],
+  hasLabel: (entry: T) => boolean,
+): { value: T[]; confidence: number } {
+  const kept = built.filter((b) => hasLabel(b.entry));
+  return {
+    value: kept.map((b) => b.entry),
+    confidence: avgScore(kept.map((b) => b.score)),
+  };
+}