[developer0hye]

File submission policy

Before attaching or committing files, read the submission policy.
Complete a security review and obtain your organization's internal approval for
public sharing where applicable. You are responsible for lawful disclosure;
the project and maintainers disclaim related liability to the extent permitted
by applicable law.

• Any submitted sample files or attachments satisfy the submission policy, or none are submitted.

Summary

Second milestone of #2232: the spacing around an empty section mark that takes no line, before a continuous section.

Word lays the next section's first paragraph out after the previous paragraph's whole w:after, plus only the part of its own w:before beyond the mark's w:after: gap = previous after + max(0, next before − mark after). This is docxide-pdf's section_break_spacing (fe7b8bea, src/pdf/mod.rs ~5666), from Word probes. Typst collapsed the two neighbouring gaps to their maximum instead. Where the next section opened a #columns or #move frame, the first paragraph's w:before was lost entirely, because it fell at the start of the frame's region.

• Parser. When a mark is dropped before a continuous section, its resolved w:after goes on the first paragraph of the next section, as the new ParagraphStyle::section_mark_after. A plain paragraph, a caption and a list item all carry it. A mark right after a page or column break opens the section at a page or column top, so it hands nothing on.
• Renderer. At a continuous section's start, take_section_mark_space writes max(0, before − mark after) as fixed space (#v(…, weak: false)) before that paragraph, and sets the paragraph's own before-gap to zero. Typst's weak block spacing looks past fixed space, so the previous block's below survives whole and the residual adds to it. All three continuous-section paths get it: the plain flow (and its #move frame), the path for a section that breaks the page, and multi-column sections. In a multi-column section the space goes in the first column only. Word's export of covid_insomnia_mental_health starts its second column above the first column's numbered heading.
• The TODO(section mark spacing) is gone. With fix(docx): keep or drop an empty section mark's line by its place in its section #2371's keep/drop rule, this completes what DOCX: an empty section-mark paragraph's line is kept or dropped by whether its w:pPr is bare, not by its position in the section #2232 asks for. The new-page counterpart (Word keeps a new-page section's first w:before less the mark's w:after, and we drop it) is a separate code path, filed as DOCX: a new-page section's first paragraph loses its w:before, which Word keeps less the closing section mark's w:after #2372 from docxide-pdf case25.

Related issue

Related: #2232, #2372, #1874. Also #2371 (first milestone), #2256.

Testing

• cargo test --locked --workspace --profile ci: 4,508 passed, 0 failed.
• cargo fmt --all -- --check and cargo clippy --locked --workspace --all-targets -- -D warnings: clean.
• New failing-first tests in docx_section_mark_tests.rs, both seen against main's code before the fix:
  • a_dropped_section_mark_absorbs_only_its_own_after_of_the_next_before compiles the layout and measures the "Before"→"Next" baseline gap against a zero-spacing control. It uses eight (after, mark after, next before) triples: (12, 8, 10) → 14 (docxide's Word probe), (12, 0, 10) → 22 (covid's shape), (4, 0, 10) → 14, (6, 2, 4) → 8, (0, 0, 10) → 10, (10, 20, 15) → 10, (10, 10, 0) → 10 and (0, 4, 0) → 0. Each runs in one column, in two columns and in a section that later breaks the page, as a plain paragraph and as a list item. Main gives 12 for the first.
  • a_dropped_section_mark_after_a_page_break_leaves_the_next_page_top_alone checks that a 10pt w:before after a page break and a dropped mark still adds nothing at the new page's top. That is the shape of LibreOffice's tdf170119_bottomSpacing test. It passes on main and fails with the page-break guard removed.
• Reported file against its Word reference: docxide-pdf scraped/covid_insomnia_mental_health (fe7b8bea, Creator: Microsoft Word), page 1, mutool draw -F stext baselines, main → this PR (Word):
  • "1. Introduction" 642.68 → 652.68 (650.45). The 12pt after, 0pt mark and 10pt before give 12 + 10 = 22pt, where main gave 12.
  • Column 2's first line is unchanged at 641.54 (638.93). Both columns now start the same 2.2–2.6pt below Word's place; see Remaining.
• Corpus scan. Run-owned release binaries: main 19436ae1 (two byte-identical builds) and this branch. Inputs: 2,370 DOCX files (tracked fixtures, docxide-pdf fe7b8bea cases/ and scraped/, the fixtures-v1 bulk corpus). No exit status changed and only 2 PDFs differ:
  • covid_insomnia_mental_health, as above. A survey of every dropped mark before a continuous section in the corpus finds it is the only Word-referenced file where the rule differs from Typst's maximum.
  • LibreOffice's tdf90697_complexBreaksHeaders (no Word export). Three continuous starts open a #move or #columns frame for changed margins. At two of them (after 0, mark 0, before 3.1 / 7.4pt) main lost the w:before at the frame's top, and the paragraph now moves down by exactly 3.10 / 7.40pt. At the third (13.2, 13.2, 2.9pt) main added the 2.9pt on top of the 13.2pt, and the paragraph now moves up 2.90pt. All three now equal the rule's answer. The page count is unchanged at 7.

Visual impact

• No rendered PDF change
• Rendered PDF change or visual evidence added
• Reason: A continuous section's first paragraph after a dropped section mark now sits where Word puts it.

Visual audit

• Issue: DOCX: an empty section-mark paragraph's line is kept or dropped by whether its w:pPr is bare, not by its position in the section #2232
• Fixture: public pinned docxide-pdf tests/fixtures/scraped/covid_insomnia_mental_health/input.docx at fe7b8beafa7612cc89ad62a7d27b2c6141c4d6ed (not committed; SHA-256 13bb08d9074be0506444bbf84a7abd26f1b3dbfb79524ced44e6b07e8db407cd; Word reference SHA-256 0fa16efc0adca7dc73238ff396af75213ed1714aa83fe1fb5cecaa2e47943d39). Tracked regressions use synthetic packages built in docx_section_mark_tests.rs.
• Page(s): 1
• Renderer and DPI: pdftoppm via compare_render.py, 150 DPI
• Evidence mode: fix
• Baseline failure report: None
• Layout audit report: assets/bugfixes/issue-2232/layout-audit.json
• Render cluster reports: assets/bugfixes/issue-2232/render-clusters-page-1.json
• Reference exporter differences: None
• Fine-detail threshold: 0.5pt
• Layout audit page count: Pass
• Layout audit text flow: DOCX: synthesized superscript placement differs from native Word across fonts and sizes #2141, DOCX: justified Latin text wraps differently from Word with hyphenation disabled #2063
• Layout audit visible fills: Pass
• Layout audit rectangle geometry: DOCX: consecutive paragraphs with the same border each draw it when one is empty, instead of one border around the group #2379, Tracker: sub-materiality visual deviations (revisit in bulk) #1874
• Layout audit large shifts: DOCX: a proportional w:line below single does not tighten body lines #2374, DOCX: a left or center w:ptab in a header or footer is drawn as a 1em space #2378, DOCX: empty paragraphs and paragraph spacing are dropped when a header is drawn #2062
• Layout audit fine shifts: DOCX: a proportional w:line below single does not tighten body lines #2374, DOCX: a left or center w:ptab in a header or footer is drawn as a 1em space #2378, DOCX: empty paragraphs and paragraph spacing are dropped when a header is drawn #2062, DOCX: synthesized superscript placement differs from native Word across fonts and sizes #2141, Tracker: sub-materiality visual deviations (revisit in bulk) #1874
• New follow-up issues found in this audit: DOCX: a continuous section's margins, header and footer never take effect on the pages after the one it begins on #2373, DOCX: a proportional w:line below single does not tighten body lines #2374, DOCX: a table-cell line compressed by w:line below 240 seats a half-advance ascent, so rows grow and text crosses the top rule #2375, DOCX: a table style's band1Horz top and bottom borders are not painted on banded rows without direct tcBorders #2376, DOCX: a token wider than a narrow table cell is never broken, so it overflows into neighbouring cells #2377, DOCX: a left or center w:ptab in a header or footer is drawn as a 1em space #2378, DOCX: consecutive paragraphs with the same border each draw it when one is empty, instead of one border around the group #2379, DOCX: a numbering level's hanging indent loses to the paragraph style's w:ind firstLine, so the list is drawn at the column edge #2380
• Model vision findings: I inspected the full 150-DPI GT, before and after pages for page 1, its 5% diff with every cluster boxed, and full-resolution crops of the column tops, the author and affiliation lines and the Keywords rules; all five pages were viewed side by side for the whole-document audit.
  • Before: "1. Introduction" sits level with column 2's first line, about one line too high. After: it sits one line below column 2's first line, as in Word, and column 1's body follows it.
  • Above the two-column body, the diff is glyph outlines only. Word embeds Calibri and Times New Roman, this host paints the metric-compatible Carlito and Liberation Serif. The exception is the four author superscripts: Word synthesizes 7pt digits, and ours are the face's superscript glyphs (DOCX: synthesized superscript placement differs from native Word across fonts and sizes #2141).
  • Remaining on page 1, all present on main:
    • An extra rule under "Keywords" (DOCX: consecutive paragraphs with the same border each draw it when one is empty, instead of one border around the group #2379).
    • Column 1's justified lines wrap one word earlier (DOCX: justified Latin text wraps differently from Word with hyphenation disabled #2063).
    • Both columns advance 11.5pt per line against Word's 10.93 (DOCX: a proportional w:line below single does not tighten body lines #2374).
    • The footer's left item sits 177pt right (DOCX: a left or center w:ptab in a header or footer is drawn as a 1em space #2378), and the whole footer 13.5pt low (DOCX: empty paragraphs and paragraph spacing are dropped when a header is drawn #2062).
  • Pages 2–5 keep section 1's header, footer and margins (DOCX: a continuous section's margins, header and footer never take effect on the pages after the one it begins on #2373), and their tables sit loose and miss banded rules (DOCX: a table-cell line compressed by w:line below 240 seats a half-advance ascent, so rows grow and text crosses the top rule #2375, DOCX: a table style's band1Horz top and bottom borders are not painted on banded rows without direct tcBorders #2376, DOCX: a token wider than a narrow table cell is never broken, so it overflows into neighbouring cells #2377). See the deviation table.
  • Rules are single solid hairlines on both sides, with no dashes. The title and headings are bold and the affiliations italic in both; there is no underline.
• GT: assets/bugfixes/issue-2232/gt.jpg
• Before: assets/bugfixes/issue-2232/before.jpg
• After: assets/bugfixes/issue-2232/after.jpg
• Native: None
• Compare: None

Visual comparison

GT	Before	After

Required inspection

• Rendered all evidence at 150 DPI or higher
• Stored progressive JPEG quality 86 assets with metadata stripped
• Used Codex/Claude vision to inspect the full GT/output pages, diff, and matched crops
• Inspected matched region crops at full resolution
• Ran compare_layout.py --audit --fine-shift PT and dispositioned every fine/large text-instance shift, rectangle geometry deviation, painted-text visibility mismatch, and visible-fill occlusion
• Ran compare_render.py --cluster-report PATH --strict-clusters and dispositioned every material 5% fuzz diff cluster by explicit ID
• Inventoried hairlines and border dash styles
• Inventoried font weight, italic, and underline emphasis

Deviation audit

Check	Result
Page count/order	Remaining: #2374, #2373, #2375, #2377, #2063. Page 1, the compared page, matches. The whole document runs to 6 pages against Word's 5: page 6 holds only the three trailing empty paragraphs, pushed over by long body lines (#2374), pages that hold less text (#2373), taller tables (#2375), an unbroken cell token (#2377) and wraps (#2063).
Element presence	Remaining: #2379, #2373, #2376. All text is present. Page 1 draws an extra rule under "Keywords" (#2379). Pages 2–5 show section 1's two-line header and three-item footer instead of section 2's one-line header and page number (#2373). Tables 2 and 4 miss their banded-row rules (#2376).
Position/size	Remaining: #2374, #2378, #2062, #2141, #2375, #2380, #1874. This PR fixes "1. Introduction": 652.68pt against Word's 650.45 (main 642.68). Column tops sit 2.2–2.8pt low (#1874). Column 2's lines drift with the line pitch (#2374). The footer sits 13.47pt low (#2062), with its left item 177pt right (#2378). The affiliation superscripts sit 1.3pt high (#2141). Table rows are taller (#2375). The bivariate list items start at the column edge (#2380).
Rotation/flip	No deviation observed.
Fill	No deviation observed: no page has fills.
Stroke/border	Remaining: #2379, #2376, #2375, #1874. Page 1 has an extra Keywords rule, and the group's rule sits 1.04pt low (#2379, #1874). Banded table rules are missing (#2376). Table header text crosses the top rule (#2375). The header rule sits 0.17pt low (#1874). All rules are solid hairlines on both sides.
Shape outline geometry	No deviation observed: no page has shapes.
Text content	Matches GT: page 1's non-space characters are identical to Word's; one space differs, as on main.
Font family/weight/style	Remaining: #2141, #2337. Calibri and Times New Roman are painted as metric-compatible Carlito and Liberation Serif, and bold and italic match, with two exceptions. The author superscripts use the face's superscript glyphs (#2141). Reference [18]'s label is upright where Word italicises it (#2337).
Text color	Matches GT: black text on white on every page.
Alignment	Remaining: #2378, #2380, #2063, #2297, #1874. The footer's left item is drawn 177pt right (#2378). The list items start at the column edge (#2380). Justified lines wrap differently (#2063). A reference URL breaks after "https://" (#2297). "Introduction" starts 2.43pt left of Word's (#1874).
Line/paragraph spacing	Remaining: #2374, #2375. This PR fixes the gap after the dropped section mark: 12 + 10 = 22pt, as in Word. Body lines advance 11.5pt against 10.93 (#2374), and table-cell rows grow (#2375).
Clipping/overflow	Remaining: #2377. "OR value" runs 11.6pt past the right margin, and "2.317" runs over the next cell.

Final performance pass

• Status: Complete
• Measurements: elapsed time and peak RSS, medians of 10 interleaved runs per input and binary (/usr/bin/time -f "%e %M").
  • Release CLI (cargo build --locked --release -p office2pdf-cli) on an Intel i5-10400, 12 threads, 32 GB, Linux 7.0.0, with Liberation, Carlito and Caladea and no Microsoft core fonts. Warm page cache.
  • Binaries: two separately built, byte-identical baselines from main 19436ae1 (base, base2), and the candidate. The candidate is byte-identical to a fresh release build of the PR head's sources.
  • Inputs: covid_insomnia_mental_health (evidence, 5 pages) and feminist_voice_dissertation (105 pages), both docxide scraped/; tdf90697_complexBreaksHeaders (bulk, 7 pages, the other mover); tracked spaced_section_mark.docx and section_alignment_continuous.docx.
  • Median elapsed / peak RSS for base, base2, cand:
    • covid: 0.520s / 67.8 MiB, 0.520s / 67.9 MiB, 0.520s / 65.9 MiB
    • tdf90697: 0.580s / 79.9 MiB, 0.580s / 79.8 MiB, 0.580s / 78.5 MiB
    • spaced mark: 0.270s / 42.7 MiB, 0.270s / 42.6 MiB, 0.270s / 40.4 MiB
    • continuous alignment: 0.420s / 43.6 MiB, 0.420s / 43.5 MiB, 0.420s / 41.9 MiB
    • 105-page thesis: 1.000s / 140.1 MiB, 1.000s / 137.5 MiB, 1.010s / 137.3 MiB
• Outcome and revalidation: no measurable change, so no optimization is warranted.
  • Elapsed time is equal within 10ms.
  • The candidate's peak RSS is up to 2.3 MiB lower on the small inputs. The two identical baselines differ by up to 2.6 MiB, so that is within build-to-build variation, not a gain.
  • The change adds one Option<f64> per paragraph style and, per continuous section, a scan to the first paragraph block.
  • No optimization code was changed, so the visual comparison and test suite above stand as the final revalidation.

Checklist

• Commits include a Signed-off-by line
• Final memory/speed pass completed after visual acceptance, or no conversion runtime impact is explained
• PR scope contains one root cause
• Remaining converter or harness deviations each reference an open issue

🤖 Generated with Claude Code