test(corpus): reproducer fixtures for section-drop bugs (#95/#96/#97)

[s-annam]

What

Adds two PII-free synthetic PDF fixtures (+ baked *.expected.json snapshots) under tests/fixtures/pdfs/unknown/ that reproduce the parser regressions tracked by the children of epic #94. No parser changes — this lands the reproducers first so the fixes in #95/#96/#97 have a regression target.

Fixture	Layout	Reproduces
student-projects-activities-singlecol	1-col, Education → Projects → Experience → Activities	#95 projects dropped, #96 Activities dropped, #97 education end-date dropped
two-column-achievements-sidebar	2-col, ACHIEVEMENTS/STRENGTHS sidebar	#96 sidebar variant + two-column boundary

Evidence (today's buggy output, captured in the snapshots)

• Projects dropped as a section — not extracted, bullets leak to "Other" #95 — student snapshot has no projects in fieldsPopulated; experienceCount: 2 (only the real Experience entries — project entries vanish, not mis-bucketed).
• Achievements section dropped — demoted to "other" sink, never extracted #96 — no achievements field in either snapshot; two-col fires two_column with 4 clean Experience entries (sidebar not bleeding).
• Education end-date dropped — only first year captured, no date-range parse #97 — value-probe on student: education range Sep 2024 - July 2025 parses to year: "2024" only (end date dropped); the range text also bleeds unstripped into institution.

⚠️ The corpus snapshot is keys/counts-only (PII policy), so it cannot assert the #97 end-date drop directly — only educationCount. When implementing #97, add the value-level assertion (start_date/end_date on the Kwansei entry) to extract-fields.test.ts. This fixture supplies the input; the unit test is the guard.

Method & PII

Hand-authored HTML → headless Chrome --print-to-pdf (Skia/PDF), same path as the existing chromium-two-column-sidebar fixture. Synthetic personas (Maya Lin / Jordan Avery), @example.com, reserved (408) 555-0123/0145 phones (valid NANP area code + 555 exchange + 0100–0199 subscriber → pass libphonenumber-js, never ring). Verified clean across all four layers — body text, Info dict, full-binary strings, and the counts-only snapshots. Source numbers absent everywhere.

When #94/#95/#96/#97 land, npm run bake-fixtures re-bake diff proves each fix.

Refs #94
Refs #95
Refs #96
Refs #97