Fix 2025 PDF export: correct AcroForm mappings, print-layer emissions, conditional-form gating

[temujin9]

Summary

The 2025 PDF export produced returns that were not safe to file. Exercising it against a simple single-filer return (one W-2, 1099-DIV/INT/B, HSA, nondeductible IRA) surfaced four classes of defect:

1. Descriptor field mappings didn't match the IRS AcroForm layer. Values landed in wrong fields — the foreign tax credit printed inside Schedule 3's name field, Form 8606 line 1 printed in the home address field, capital-gain distributions landed in Schedule D row 8b's proceeds column, and the Form 1040 "Single" mapping actually checked the §301.9100-2 late-election header box.
2. Computed line values never reached the printer. Intermediate nodes (schedule3, schedule_b, schedule_d, form8606, form8880, form8889) route results downstream only, so their own pending dicts held raw inputs and the printed schedules came out blank.
3. Every $0 form was emitted. hasData treated any defined value as content, padding the export with 14 pages of inapplicable forms (Schedule A/EIC/SE, 6251, 8959, 8960, 8962, and an off-by-one-row Form 8880 for a filer 4x over the Saver's Credit phase-out).
4. Line 37 could disagree with the form's own arithmetic by $1 (computed from the cents-level difference instead of the rounded filed lines).

Changes

• Nodes self-emit print_* line values into their own pending (the pattern the f1040 output node already uses). Schedule D also splits covered/no-adjustment transactions onto direct-report lines 1a/8a per "Exception 1"; 8889 emits the full Part I ladder only when the form is required; 8880 emits only with a nonzero credit.
• Descriptors rebuilt from AcroForm field dumps (name/type/rect) of the official 2025 PDFs for f1040, Schedule 3, Schedule B, Schedule D, 8606, 8880, 8889 — including filing-status/digital-assets checkboxes, occupation, name/SSN headers on every schedule, and middle-initial support (general accepted taxpayer_middle_initial but dropped it).
• includeWhen gate on descriptors (with access to the full normalized pending) encodes each conditional form's statutory trigger: Schedule A only when itemizing, Schedule B only above $1,500, SE/6251/8959/8960 only when the corresponding Schedule 2 tax exists, 8962 only with 1095-A data.
• printZero field option for lines where an explicit 0 is meaningful (8606 line 2, 8889 lines 2/13).
• Year-pinned template URLs — pub/irs-pdf/*.pdf silently changes revision each season; the 2025 module now fetches pub/irs-prior/*--2025.pdf.
• Lines 34/37 computed from rounded operands so the printed return satisfies its own line 24 − line 33 arithmetic.

Also corrects three pre-existing test expectations that failed against correct engine output:

• two Form 8889 tests treated employer HSA contributions as deductible on line 13 (IRC §223(a): the deduction is min(line 2, line 12); employer contributions are already excluded from W-2 box 1);
• the EITC MFJ_vs_single_1_child test used a $30,323 phase-out start where Rev. Proc. 2024-40 §3.11 (and the engine config) has $30,470.

Test plan

• deno task test: 6,070 passed, 0 failed (baseline on main: 72 failing).
• End-to-end: exported the exercise return before/after, rendered every page (ghostscript) and verified field-by-field against hand-computed values — 8-page export containing exactly 1040, Schedule 3, Schedule D, 8606, 8889 with all lines, checkboxes and headers correct.

🤖 Generated with Claude Code