fix(pagination): lay the NVCA charter out on LibreOffice's 48 pages (#1151)

[JSv4]

Closes #1151

Why

TestFiles/NVCA-Model-COI.docx (the NVCA model charter) paginates to 51 pages in the paginated view on main (which now includes #1152's 6pt spacing between notes). LibreOffice 25.8.7.3 lays it out on 48. The issue had located only part of the gap. A page-by-page comparison against LibreOffice found five contributors that interact with each other. With all of them fixed, the paginated view lays the charter out on 49 pages, against LibreOffice's 48 (it was exactly 48 before #1152 landed). No footnote now starts on a later page than its citation; on main 23 do. No page holds only footnotes; on main pages 6 and 7 do. The remaining page has not been located: the layouts first diverge around pages 18–21, where Docxodus carries about five lines of a note continuation onto the next page that LibreOffice finishes. This is recorded as open in docs/ooxml_corner_cases.md.

What was wrong and how it works now

1. Line pitch. Chromium rounds line-height: normal to whole pixels. That gave 12.75pt against LibreOffice's 12.65pt per body line, and 12.0 against 11.5pt per note line. The export already corrected this with applyUnroundedNormalLineHeights. The paginator now runs the same pass on its staging tree, and the paginated editor runs it on every block it re-renders after an edit, so the view, the editor and the export share one line pitch. The pass now reads all styles before it probes fonts and caches per-character fallback probes. On the charter it takes about 57ms instead of 67–155ms, with identical results.
2. Notes piled up onto later pages. A paragraph that cited a footnote stayed on its page even when the note could not start there, so the note was deferred, sometimes into note-only pages. Placement now checks whether the paragraph's new notes can begin in the space left on the page. If they can't, the paragraph is split above the citing line, which always moves whole, or moved whole if it keeps its lines together. The note is still deferred in these cases:
   • a block that starts its page;
   • a table;
   • a paragraph kept with the block before it;
   • the paginated editor, which never splits paragraphs.
3. Paragraphs that could not be split. Headings (h1–h6, most of the charter's clauses), paragraphs that cite notes, paragraphs with list markers, and notes whose first paragraph opens with a tab after the number could not be split. Now they can. Their leading box (list marker, note number or empty tab box) stays whole on the first piece.
4. A block that cites no note was refused when existing notes filled the capped note area. That left an empty body page, and it happens on main too. Room is now judged by the space still free.
5. Missing spacing between numbered paragraphs. The converter dropped the spacing after between consecutive numbered paragraphs of the same list. On page 4 that removed the gap between FIRST, SECOND and THIRD. The only markup that removes that spacing is w:contextualSpacing, which these paragraphs don't set, and LibreOffice keeps the gap (24.7pt between FIRST and SECOND; this PR gives 24.64pt; main gave 12.5pt). The converter's grouping (ApplyContextualSpacing) now groups only by w:contextualSpacing and style, and treats w:val="0" as off. On its own this fix adds two pages, pages LibreOffice lays out too.

Validation

• Tests.
  • New .NET tests in NumberedParagraphSpacingTests.
  • New specs in pagination-note-citation-placement.spec.ts (13 tests), plus pagination-line-pitch.spec.ts and editor-paginated-line-pitch.spec.ts.
  • Reverting each rule fails its spec. That covers the line-start cut, the keep-with-next guard, table deferral, the heading and leading-box gates, the free-space comparison, the note-cap gate and the editor pass.
  • Playwright (pagination, page-map, editor, standalone-export and visual-parity unit specs): 545 passed. npm run typecheck is clean. The .NET suite passes.
• Charter checks. All 94 citations appear exactly once, all 94 notes are present, and the body text length is identical to main.
• Existing expectations changed (approved).
  • editor-paginated-charter.spec.ts PAGES 53 → 54, as measured on the combined code. The editor never splits paragraphs; the extra page is not attributed.
  • One case dropped from the "keeps paragraphs intact" spec in pagination-paragraph-fragments.spec.ts: a paragraph that cites a footnote, which may now split on purpose.
  • pagination.bundle.js size budget 20 → 23 KiB. It is now 20.5 KiB, up from 19.45 KiB.
• Visual-parity ratchet, refreshed as approved. The browser-page record (npm/tests/visual-parity/ratchet.json) was rebuilt with the repo's own buildRecord from the summary of CI run https://github.com/JSv4/Docxodus/actions/runs/38988464367. That run measured this branch's tip on the pinned reference environment from a clean tree.
  • Improved: two-column-section (severe → major), wrapped-image-tight (severe → major), legal-contract, endnote, footnote, merged-table and fields-and-tabs.
  • Accepted drops: wrapped-image-square, inline-image and tracked-deletion (close → minor). None of the three has a Word reference. On the seven corpus cases that do, the page's ink now ends within a pixel of Word's, where main was up to 36px off.
  • Generated-PDF ratchet: it passes unchanged, with improvements only.
• Cost. On the charter, paginate() takes about 1.2–1.4s against about 0.59s on main in the paginated view, and about 0.6–0.7s against 0.5s in editor mode (noisy machine). The extra work is the additional splitting, about 10 more measurements per page, and it grows linearly with document length.

Overlap with open PR #1152

#1152 changes paginated footnotes too. The two apply together with conflicts only in CHANGELOG.md, docs/ooxml_corner_cases.md and the charter spec's PAGES line. #1152 adds 6pt after each note, so whichever merges second should re-measure the charter's page count rather than pick either number. This PR already handles #1152's note number sitting inside the first paragraph.

Adversarial review

One round of hostile review ran before opening. It confirmed that no content is lost or duplicated, that no hangs or blank pages appear, that the memoization is sound, and that the export path stays idempotent. It found:

1. A citing line cut in mid-line on the charter itself (pages 9→10 and 44→45). Note-aware splits now break only at a visual line start.
2. A citing paragraph moved away from a keep-with-next heading. The move no longer fires in that case.
3. A note-citing table moved whole, half-emptying a page. Tables keep their place again.
4. The line pass caused the whole editor slowdown. It is now faster.
5. Edited editor blocks reverted to the rounded pitch. They are now re-resolved.
6. fix(render): keep a paginated footnote's first-paragraph indentation (#1130) #1152's in-paragraph note numbers weren't recognised as leading boxes. They now are.
7. Docs overclaimed Word behaviour. They now say only what the LibreOffice evidence shows.

A related keep-with-next stranding with a pending note continuation predates this change and is pinned by an existing spec. It is filed as #1180, with #1178 (the 60% note-area cap against LibreOffice's ~94%) and #1179 (a whole note placed after a continued one).

Final CI round. After the rebase, CI's Playwright job showed two toc-line-geometry tests expecting the base line box Chromium rounds to a whole pixel. With CI's fonts, that rounding is 0.9px. The paginated view now lays lines at the font's unrounded height, which is Word's measured single spacing, so the spec's probe now measures the unrounded box (test-only change; TOC geometry unchanged). Reproducing CI's fonts locally also turned up a real regression from the line-pass speed-up: it probed non-Latin characters one at a time, and Chromium chose a taller fallback font, so exported CJK lines came out too tall. It probes them together again; the CJK export spec fails without the fix. All 17 checks pass, and the visual-parity ratchet holds for all 21 cases: https://github.com/JSv4/Docxodus/actions/runs/38994865996