fix(heuristics): two-column flatten keeps flat-bullet awards as achievements (#131)

[rohithgollapalli]

What

Closes #131. A two-column résumé whose Awards/Achievements section is a flat bullet list (every item is itself a bullet, no name-led header line) was dropped entirely: collectAnchors found zero first_line anchors and parseEntryBlocks returned [].

This adds a bullet-anchored fallback: when a first_line section yields no anchors, each top-level bullet (one sitting at the bullet-marker margin) becomes its own entry. Marker-less lines below it (a bare year, a wrapped award name) fold into that entry's title; deeper-indented sub-bullets become its body.

The fallback is strictly additive — only first_line sections that previously returned [] are affected. Sections with a real header keep ≥1 anchor and are untouched.

Why this shape

The issue's original framing (two bullet margins interleaved in one section, global bulletMarkerX rejecting lines) was stale. detectColumnBoundaries already bands these fixtures into single-column reading order before sectioning, so each section is one column's clean list. The remaining bug was purely the zero-anchor flat list. This fix also re-seats cleanly on top of the #127 segmentation rework that landed on main (bulletMarkerX/isWrappedContinuation), reusing those helpers rather than duplicating them.

Acceptance criteria

• ✅ weasyprint-cairo-two-column Awards parse as two distinct achievements (Globex Engineering Excellence — 2021, Acme Innovation Prize — 2023) — achievementsCount 0 → 2. Same for the sibling google-docs-skia-proxy-two-column.
• ✅ No single-column regression: google-docs-skia-proxy-nonstandard projects stays 1; student-projects-activities-singlecol unchanged; main's wrapped-bullet-tail handling preserved.

Tests

• 2 new unit tests in entry-blocks.test.ts (flat-list split; sub-bullet-as-body).
• Re-baked corpus: only the 2 target two-column snapshots differ from main; all others byte-identical. No score changes.
• npm run typecheck, full vitest (628 passed), and eslint on changed files all clean.

🤖 Generated with Claude Code

[s-annam]

PR Review: two-column flatten keeps flat-bullet awards as achievements (#131)

Summary

Clean, strictly-additive fix. The zero-anchor flat bullet list (awards where every item is a bullet, no header line) was dropped entirely; now each top-level bullet becomes an entry, marker-less lines fold into the title, and deeper sub-bullets become the body. Well-scoped, well-tested, and re-seats correctly on the #127 bulletMarkerX/isWrappedContinuation helpers rather than duplicating them.

Gates (run on checked-out branch): npm run typecheck clean · npm run test 628/628 · PR diff touches only the 2 target snapshots (no masked corpus regression).

Spec alignment (#131)

AC	Status
Two-column Awards parse as 2 achievements (weasyprint-cairo, google-docs-skia-proxy)	✅ achievementsCount 0→2
No single-column regression	✅ only 2 snapshots differ, rest byte-identical
Reuse #127 helpers, no duplication	✅
Stay under cognitive-complexity bar (split buildBulletEntry)	✅

Highlights

• Trigger is tightly gated (anchors.length === 0 && cfg.anchor === "first_line") so other anchors keep returning [] and real-header sections are untouched — genuinely additive.
• Top-level vs sub-bullet split via x <= markerX + 2 matches the existing isWrappedContinuation tolerance — consistent margin logic.
• Two targeted unit tests (flat split + sub-bullet-as-body) plus a clean corpus rebake.

Key findings (non-blocking)

1. [Suggestion] buildBulletEntry routes every marker-less line to the title regardless of position relative to sub-bullets — a wrapped continuation of a body sub-bullet would be pulled up into the title. See inline. Unreachable for the current award fixtures; forward-looking hardening.
2. [Informational] The fallback fires for any zero-anchor first_line section, projects included — not just achievements. Almost certainly an improvement (no fixture regressed), but the docstring frames it as awards-specific; a one-line note that it's generic to first_line would match the behavior.

Verdict

APPROVE — no blocking items; correct, additive, gates green, snapshots clean.

[s-annam]

APPROVE — 0 blocking findings.

Gates (checked-out branch, fresh npm install)

• ✅ tsc -b --noEmit clean
• ✅ vitest 629/629; new entry-blocks.test.ts 15/15
• ✅ Only the 2 target snapshots change (weasyprint-cairo-two-column, google-docs-skia-proxy-two-column), achievementsCount 0→2 — matches acceptance criteria
• ✅ No PDF fixtures touched → PII-safe (count-only snapshot deltas)
• ✅ 150/3 LOC, single concern, reuses #127 helpers (bulletMarkerX/isBulletLine) instead of duplicating

Logic spot-check

buildBulletEntry title-vs-body state machine verified: marker-less line before the first sub-bullet folds into the title; after a sub-bullet it joins that bullet's tail. Windowing anchorIdx+1 .. nextIdx correct, no off-by-one. The 3 new unit tests cover title-fold, sub-bullet-as-body, and wrapped-tail.

Non-blocking follow-ups (not gating this PR)

1. Date-only bullet → empty-title achievement. A bullet that is only a year (• 2023, no name) yields {title:"", year:"2023"} with score 0, emitted unfiltered (achievementFromBlock). This is a pre-existing systemic behavior shared by all three extractors, not introduced here. Cheap hardening if wanted: filter empty-headerLines blocks in extractAchievements/extractProjects. Low priority (unrealistic input).
2. Greedy year-strip on joined titles. parseDateRange's YEAR_RE fallback consumes any bare 4-digit year mid-phrase (e.g. • ACM Fellow, Class of 2019). Pre-existing shared-parser limitation, now exercised at a new call site — worth a one-line JSDoc note.

The two parseBulletList [] guards (!isFinite markerX, zero top-level anchors) are unreachable given the caller gate + markerX = min(x) — safe defensive code, no action needed.

Ship it.

[s-annam]

All review comments addressed (wrapped sub-bullet tail fixed in c595f00, unit test added). Gates green. Approving for merge.