fix: recognize single-column label-rail section headers

[s-annam]

Summary

Single-column résumés that put the section name in a narrow left rail — sharing the row with the section's first entry (Experience Staff Engineer, Platform Aug 2024 - Present), or vertically stacked as the left cell of a skills grid (Technical / Skills) — left the header unrecognized. matchSectionHeader is exact-match / head-noun-last, so neither an inline leading-token header nor a stacked grid label opened its section: skills leaked into the experience band and the date anchor grabbed skill rows as role-1 title/company (bullets=0, education swallowed into a spurious role).

Fix is in section-header recognition (sections.ts), single-column only. The two-column detector and experience headerLookback are untouched (this résumé is genuinely single-column — triggers: []).

Closes #355

What changed

• Leading-token recognizer (classifyLine, single-column-gated): a line whose leading item exactly matches a section alias opens the section and retains the row remainder as its first content line — so the inline title/date reach the entry parser.
• Stacked rail-label join (splitIntoSections): two consecutive rows sharing a rail x whose leading tokens join to a canonical alias open that section, gated on each row being a value grid (≥2 short cells).
• FP guards (from adversarial review): trailing-colon reject on the label run; remainder must lead uppercase, no terminal sentence punctuation, and carry a strong date token (month/season+year, slash date, Present — not a bare YYYY-YYYY span) or a degree/institution tell in a leading window.
• Fixtures + tests under sections.test.ts.

Update (f77cf7f) — real-résumé repro was still failing; extended the fix

Dogfooding the canonical #355 résumé against the branch showed 923ad0a alone did not resolve it: score 21, skills(0), all roles bullets=0, role-1 title/company came out as skill tokens. Root cause was a compounding geometry the original synthetic fixture under-modeled: a genuine left-label-rail layout (section names at x≈26, body at x≥99) whose multi-column skills grid has irregular per-cell baselines, so groupIntoLines fragments it one-line-per-cell — the stacked Technical/Skills label rows are no longer adjacent clean grid rows, tryStackedRailLabel bails, and skills scatter into experience.

• splitByLabelRail — a single-column-only pre-pass: detect the narrow rail (min left-edge cluster ≥40pt from the body), recover each label (alone / stacked-join / inline leading-token), then assign every body line to the label whose y-band it falls in. Assignment is by absolute (page, y), so grid fragmentation and pdfjs emission order don't matter.
• Per-(page, y) keying throughout so a multi-page rail résumé can't scramble page-2 content into page-1 sections.
• RAIL_BAND_OVERLAP_TOL = 5pt (sub-line jitter, not a line-height) so a label ~one line below the prior section's last bullet can't steal that bullet.
• mergeRowsByBaseline (rail path, date-anchored sections only) rejoins lines the 50pt column-gap split had stranded, restoring glyphless-bullet detection.
• Standalone-alias gap guard (STANDALONE_ALIAS_MIN_GAP = 12pt, review feedback): a real title whose first word is its own item (Experience Designer) can't false-open a section — a genuine rail header has a rail→body x-jump, a compound title has ordinary word spacing.
• New generator gen-label-rail-grid.mjs + fragmented-grid fixture + snapshot; regression tests for the multi-page, tight-spacing, and standalone-alias hazards.

Result on the real repro: score 21 → 64, skills(0)→skills(10), roles bullets 2 / 10 / 7 / 3.

Adversarial review

• Base fix (923ad0a), 2 rounds: R1 caught a BLOCKING FP (styled Experience: / prose-year label lines false-opening) → fixed via the FP guards. R2 converged.
• Extension (f77cf7f), 2 rounds: R1 caught 2 BLOCKING confirmed bugs — multi-page y-band scramble (fixed with per-(page,y) keying) and a 15pt tolerance stealing the prior section's last bullet (dropped to 5pt); both regression-tested (fails@15/passes@5). R2 = reviewer (Rohith) approved; his one non-blocking residual (single-word alias vs. a real compound title) closed by the standalone-alias gap guard.

Test plan

• npm run test green (1641 passed; corpus 41/41, no snapshot regressions)
• npm run typecheck / npm run lint clean
• npm run verify green (passed on push)
• Real-résumé probe confirms recovery (score 64, skills 10, bullets 2/10/7/3) — real PDF is local-only, never committed
• Fixture personas verified synthetic — no real PII (JANE SMITH and JORDAN BLAKE, @example.com, (312) 555-0123)

[cloudflare-workers-and-pages]

Deploying resumelint with    Cloudflare Pages

Latest commit:	1098d23
Status:	✅  Deploy successful!
Preview URL:	https://f633151c.resumelint.pages.dev
Branch Preview URL:	https://gh-355.resumelint.pages.dev

View logs

[rohithgollapalli]

Reviewed the full diff plus the surrounding sections.ts machinery it depends on. Well-guarded and heavily tested — no blocking findings.

Verified clean:

• Regex statefulness — remainderLooksLikeEntry calls .test() on imported DATE_RANGE_RE (/i), DEGREE_RE (no flags), INSTITUTION_HINTS (/i); none global, so no lastIndex flakiness.
• Field consistency — isGridValueRow reads it.str, matching PdfTextItem/mergeItemText; the test helper maps text→str and models item widths the same way the recognizer reads them.
• Loop-var scoping — the new stacked-rail block reads openedRealSection/seenContactInProfile, both set before it fires (contact line → flag → next row un-gated), matching the fixture ordering.
• Item-boundary FP guard holds for the common case — a normal Education Coordinator title is one text run, so item[0] never equals a bare alias. The FP-rejection tests (colon lead, lowercase-connective lead, bare-year span, prose-that-joins-to-Technical Skills) are non-vacuous.
• Fixture PII — extracted the committed binary with pdftotext: JANE SMITH / jane.smith@example.com / (312) 555-0123, fully synthetic and phone-policy-compliant.

One low-severity, non-blocking residual left inline on matchLeadingTokenHeader (a real Experience Designer-type title with a split first item could false-open experience) — small blast radius, inherent to the item-boundary approach; a one-line acknowledgement comment would suffice. Nice work with the adversarial-review FP guards. LGTM 🎉

[s-annam]

Thanks for the review, @rohithgollapalli 🙏 — all outstanding items are resolved:

• Your review — the standalone-alias residual you floated is fixed (STANDALONE_ALIAS_MIN_GAP gap guard in matchLeadingTokenHeader), with the narrower same-x residual left deliberately + doc-commented.
• Fallow — both complexity findings cleared by a pure extract-function refactor (1098d23): splitByLabelRail → recoverRailLabels (+tryStackedLabel) + buildRailSections with bandFor/mergeRailEntryRows lifted to module scope; matchLeadingTokenHeader inner match → matchAliasPrefix. No behavior change.

Functionality verified end-to-end on the real résumé that motivated #355 — both via the probe-experience lane (score 64, skills 10, bullets 2/10/7/3) and on the deployed GitHub Pages preview.

Merging with admin bypass.