feat: section-recognition L2/L3 — anchor fallback + visual bounds

[s-annam]

Summary

Lands L2 + L3 of the section-recognition epic (#109, on top of merged L1 #110), plus the #113-review follow-up #114.

• Drift-guard comment in sections.config.ts overstates what it catches #114 — reword the sections.config.ts _drift guard comment to its true one-directional nature (TS catches missing SectionName keys; extra JSON keys slip through structurally) + add a runtime validate() extra-key check.
• L2 — Head-noun anchor fallback for qualified section headers (closes #108) #111 (L2) — head-noun anchor fallback in matchSectionHeader: a qualified header ("Relevant Experience", "Customer Service Experience") matches when its last token is a section anchor, behind 7 FP guardrails (length/word-count/head-noun-last/no-terminal-punct/not-bullet/anchorFallback flag/header-casing). Exports SECTION_ANCHORS + SECTION_ANCHOR_FALLBACKS; skills/other stay off the fallback so a flattened two-column sidebar label can't strand experience. Synthetic Chromium fixture added.
• L3 — Visual-primary boundary detection + name/contact disambiguation #112 (L3) — visual-primary boundary in splitIntoSections: a font-distinct line (ratio ≥ 1.2 over body baseline) opens a boundary even with no keyword, with name/contact disambiguation so the candidate name isn't promoted. First direct unit coverage for splitIntoSections (sections.test.ts). allCaps dropped as a visual signal after a corpus pass showed it was a net FP source (acronyms / skill tokens / sidebar labels); genuine all-caps headers are already caught by the keyword/anchor path.

Closes #108
Closes #111
Closes #112
Closes #114

Test plan

• npm run typecheck clean
• npm run lint (eslint) clean
• npm run test green — 438/438
• Fixture persona verified synthetic — no real PII (Step 3.5)
• Manually verified in npm run dev / npm run preview

[Samhit21]

PR Review: feat: section-recognition L2/L3 — anchor fallback + visual bounds

Summary

Clean L2 + L3 land. The 7-guardrail framing for the anchor fallback is the right shape, the name/contact disambiguation in splitIntoSections is a real fix (not a workaround), and dropping allCaps as
a visual signal after the corpus pass is the kind of decision the reasoning round-trips back into the docstrings — well-evidenced. New direct unit coverage for splitIntoSections plus the synthetic
Chromium fixture for the qualified-header path closes the right loops. validate()'s extra-key runtime check addresses the one-directional _drift gap from #114 cleanly.

Prior review activity: two fallow bot comments about SECTION_ANCHORS/SECTION_ANCHOR_FALLBACKS being unused outside the module — already addressed in the in-file comment at regex.ts:105-108 and not
re-flagging.

Findings

1. [Suggestion] Guard 7's casing check at regex.ts:172-176 lets digit-led tokens slip ("5 Years Experience", "10+ Years Experience"). /[a-z]/.test(first) is false for digits, so the negative
   check passes when the positive check wouldn't. Trivial swap to /[A-Z]/.test(first) plus a fixture row in the FP table.
2. [Suggestion] validate() in sections.config.ts:68-100 doesn't type-check the booleans or enforce anchor uniqueness across anchorFallback: true sections. Future-config hazard for the
   load-bearing L2 contract.
3. [Nit] hasContactShape in sections.ts:306-311 mutates exported /g regex state. Works today; pattern is fragile if a future caller reorders or runs after these regexes have an in-flight
   iterator.
4. [Question] sections.config.json routes the "qualifications" anchor to education. Common US-convention reading is closer to skills/profile; worth an inline comment pinning the intent
   (especially if the corpus you're tuning against is residency/CV-leaning).

Verdict

Action: COMMENT — none of the four are merge-blockers. #1 is the highest-value follow-up; the FP risk is real but contained to a few uncommon header shapes.

[s-annam]

@Samhit21 all three review threads addressed and resolved in e4ea6a5 — ready for final review.

• Guard 7 casing → require uppercase lead; 5 Years Experience / 10+ Years Experience / 3 Years Experience pinned in regex.test.ts.
• validate() → boolean-type checks + cross-section anchor uniqueness, split into helpers to stay under the fallow gate.
• hasContactShape → non-global regex clones, no shared lastIndex mutation.

Your Guard 7 catch also closed an incidental 20% Projects column-flatten match on chromium-two-column-sidebar (snapshot rebaked, projectsCount 2→0). Recovering that properly is its own concern — filed as #117 (lean toward the L3 visual-path attribution route), out of scope here.

Gates green: typecheck clean · 439/439 tests · fallow audit ✓ no issues in changed files. Requesting your approval when you get a chance. 🙏