Fix five layout defects found against Word: blank pages, nested tables, list counters, table widths, column balancing

[ikashapov]

Fix five layout defects found against Word: blank pages, nested tables, list counters, table widths, column balancing

Five issues, one PR, because they were found together — a user converted a batch of real documents on Windows, compared every page against Word 2024 LTSC, and sent the diffs. Each fix is its own commit and can be read on its own.

Where the spec left a question open, it was settled by measuring Word 16.0 (Office 2024 LTSC) rather than by choosing a reading: eleven probe documents, each isolating one decision, rendered in Word and read back through its object model. The numbers are in the commit messages and in the test files.

#228 — a bookmark-only paragraph was charged two lines

Fragment::occupies_line already said a bookmark draws nothing; nothing acted on it. The paragraph mark's synthetic LineBreak is injected at the front of the fragment vector, so a paragraph of pure annotations reached the fitter as [LineBreak, Bookmark], and fit_lines emitted the stranded bookmark as a second line that line_height_for then charged a full line box.

Word writes the automatic _GoBack bookmark into almost every document it saves, usually into the trailing empty paragraph, so a last page full to within one line gained a blank page. fit_lines now widens the previous line's range instead of opening a new one; the bookmark still reaches line_emit and still becomes a PDF named destination.

#229 — a nested table was measured against the sheet

clamp_auto_grid_to_page took the caller's width only as a floor under a paper-edge limit. For a top-level table that slack is deliberate — 40 corpus tables declare a grid a few points past the text column and Word draws them into the margin. For a nested table the caller's width is the host cell's content box, computed correctly and then discarded: 019_nested_tables.docx drew its nested rows 12 pt past the outer border and over the neighbouring cell's text.

build_table now takes a containment ceiling — None at the two top-level call sites, the cell's inner width at the nested one. §17.4.57 floating tables keep the page limit until a Word render says whether Word confines one.

#230 — list counters were keyed by the instance, not the definition

§17.9.2's w:abstractNum is the list and carries the sequence; w:num is an instance of it, and declaring a second instance is how a document applies w:startOverride. Counting per instance split one list in two: a specification document numbered its last two headings 6 and 7 where Word numbers 9 and 10.

Measured: two instances over one abstract, alternating, give 1. 2. 3. 4. 5. in Word — one sequence. An instance carrying startOverride=8, used twice, gives 1. 2. 8. 9. 10. 11. — the restart fires once, at first use. So the counter moved to the abstract definition and w:startOverride became the one-shot §17.9.28 describes, carried beside the level's own w:start rather than written over it. A second commit records what numbering-direct-indent.docx settles: instantiating a level as an ancestor also spends the restart.

#231 — table widths

Two independent defects in the same resolution code.

The cell-margin extension that models Word's autofit-to-window fired for every table at or past full width. Word does not extend one that asks for more than the window: for a 104% table it saves a grid summing to 1.04 × the text extents with no margin term. The gate is now an equality; a real document's table moves from 549.7 pt back onto the 539.9 pt margin.

An autofit table also kept its declared w:tblGrid and never read w:tcW. Word resolves the §17.4.63/§17.4.71 conflict toward the preferences: with a 2000/4000/4080 grid and tcW 1000/3000/6080 — where the grid, the preferences and an equal division each predict something different — Word draws 50 / 150 / 304 pt. Preferences under the text column are not stretched; over it they are scaled down. tblLayout fixed, a missing preference and a spanned cell all keep the grid.

Word's content-minimum rule (a column widened past its preference to fit an unbreakable token) is measured in the test file and deliberately not implemented here.

#232 — columns were never balanced, and w:sep was never drawn

The flow had one height budget per page and no second pass, so the last page of a column run was filled greedily and the following section pushed onto a new page. Word balances at a continuous break and only there: six short paragraphs come out 3 / 3 under <w:type w:val="continuous"/>, and all six in column 1 under nextPage or at the end of the document.

PageLayoutState grows column_bottom beside bottom — content asks the column budget, footnotes and the §17.6.22 handoff keep the page's own bottom — and layout_document estimates the balanced height, then bisects up toward the greedy bottom if a line does not fit.

w:sep was parsed and read by nothing. Word draws a hairline down the gutter centre from the body top to the bottom of the deepest column, and draws none at all when the second column took no content.

Verification

• cargo test --all — 2745 passed, 0 failed; cargo fmt --check, clippy --all-targets --all-features -D warnings, cargo doc -D warnings, --no-default-features and --features capi all clean.
• Every fix is spec-first: the tests were written against Word's measured behaviour and watched fail before the change.
• The reported documents were re-rendered and compared page by page: the numbering document now reads 9 and 10, the nested table stays inside its cell, the 104% table ends on the margin, the column run balances onto one page, and the trailing _GoBack paragraph no longer costs a page.

Closes #228, closes #229, closes #230, closes #231, closes #232