Refactor: split extract-fields.ts god-module (1060 LOC) into per-field modules

[s-annam]

Problem

src/lib/heuristics/extract-fields.ts is 1060 LOC — the largest module in the repo — and holds all 10 field extractors plus ~20 private helpers in one file:

extractName            (l.174)
extractContact         (l.292)   + ContactExtractionResult interface
extractSummary         (l.496)
extractSkills          (l.538)
extractExperience      (l.575)
extractProjects        (l.650)
extractAchievements    (l.744)
disambiguateCompanyTitle (l.932)
extractEducation       (l.988)

Each extractor is now independently complex (name scoring, contact band/annotation logic, education chunking, experience entry-blocking, column-skill splitting). The file has ~15 issue/template-specific heuristic anchors and keeps growing — PR #125 alone added +400 lines to it. This is a god-module: hard to navigate, hard to review a single extractor's change in isolation, and a magnet for unrelated churn.

This is a behavior-preserving refactor — no parsing logic changes, no score changes, no snapshot changes.

Why now

The deterministic Tier-1 extractor is where most template-fix PRs land, so the file's growth rate is high. Splitting now (before more accretion) keeps each extractor reviewable and makes the next round of fixes touch a ~150-LOC file instead of a 1060-LOC one.

Current structure facts (verified)

• Sole non-test importer: src/lib/heuristics/openresume.ts:44 imports extractName, extractContact, extractSummary, extractSkills, extractExperience, extractProjects, extractAchievements, extractEducation from ./extract-fields.ts. (disambiguateCompanyTitle is exported but consumed internally by extractExperience.)
• Cross-extractor private helpers:
  • looksLikeContactLink — used by extractSkills (token + cell rejection).
  • disambiguateCompanyTitle — used by extractExperience (and exported).
  • parseEducationDates / educationDateFields — education-only.
  • isBulletLine, stripBullet (imported from ./sections.ts), mergeItemText (imported from ./sections.ts), matchSectionHeader (from ./regex.ts) — already external, stay as imports.

Implementation plan

1. Create per-field modules under src/lib/heuristics/extract/ (new dir):
   • name.ts — extractName + isSingleNameWord, looksLikeMononymName, findContactClusterY, looksLikeDocTitleBoilerplate.
   • contact.ts — extractContact, ContactExtractionResult, isLinkedinProfileUrl, LINKEDIN_NONPROFILE_RE, normalizeUrl, annotation/band helpers.
   • summary.ts — extractSummary.
   • skills.ts — extractSkills + isSkillToken, splitColumnCells, looksLikeContactLink, PROFILE_LABEL_RE/PROFILE_HOST_RE/URLISH_RE.
   • experience.ts — extractExperience, disambiguateCompanyTitle, looksLikeTitle, looksLikeCompany, TITLE_KEYWORDS_RE.
   • education.ts — extractEducation, educationFromChunk, isDateOnlyLine, parseEducationDates, educationDateFields.
   • projects.ts — extractProjects.
   • achievements.ts — extractAchievements.
2. Shared helper placement: put a helper in the one module that uses it. If two modules need one (e.g. looksLikeContactLink if it ever spreads), promote it to a small extract/shared.ts — but only on real second use, not speculatively.
3. Keep src/lib/heuristics/extract-fields.ts as a barrel that re-exports the public surface, so openresume.ts:44 imports are unchanged:

   export { extractName } from "./extract/name.ts";
   export { extractContact, type ContactExtractionResult } from "./extract/contact.ts";
   // …etc
   export { disambiguateCompanyTitle } from "./extract/experience.ts";

   (Optionally update openresume.ts to import from the new paths and delete the barrel — but the barrel is the lower-risk default; decide during implementation.)
4. Move the co-located tests the same way, or keep extract-fields.test.ts importing through the barrel — either is fine as long as the suite stays green.
5. Preserve SPDX headers on every new file (the 3-line Apache-2.0 header).
6. Run npm run typecheck && npm run test && npm run lint.

Acceptance criteria

• extract-fields.ts is ≤ ~60 LOC (barrel re-exports only) or removed with openresume.ts importing the new modules directly.
• No single new extractor module exceeds ~250 LOC.
• npm run test green with zero changes to any *.expected.json snapshot (proves behavior-preserving).
• npm run typecheck and npm run lint clean.
• openresume.ts produces byte-identical HeuristicResult for the existing fixture corpus (snapshot stability is the proof).
• Every new .ts file carries the SPDX 3-line header.
• No new dependency added.

Out of scope

• Any change to extraction behavior (that's the heuristic PRs' lane).
• The scorer / segmentation rework (tracked separately — see the section-segmentation issue).