fix(docx): hide the empty paragraph closing a cell that ends in a nested table

[DemchaAV]

Why

Word ends a table cell on a paragraph, so the export closes every table nested in a cell with a hairline one, 2 twips tall (newTable).

• Word lays these hairlines out at no height that shows. Hiding them or thinning them to 1 twip changed nothing measurable in CobaltRota.
• LibreOffice gives each one its tenth of a point.
  • CobaltRota has a chip in a table in a table in each cell, so every row stood up to 0.3pt taller there than on the page.
  • Its last row stood 5.2pt low.

What changed

• hideTheCellClosingMarks. After the whole document is written, every table cell is checked, nested tables included. Where a cell ends in a paragraph right after a table, and that paragraph is still the bare hairline, its mark is hidden (w:vanish), so it is not laid out. The bare hairline means:

  • structure alone: no run, field, anchor, bookmark, border, shading or section properties;
  • an exact 2-twip line;
  • no space above or below.

  A hairline holding the space under the table keeps its mark, since that space would go with it.

• structureAlone / hide. Extracted from hideTheClosingMark, which hides the paragraph closing a document that ends in a table, so both passes share the rule.

• Docs. CHANGELOG.md, the composed-cells row of docs/recipes/docx-export.md, and the table row of docs/architecture/backend-capability-matrix.md.

Verification

• Full reactor gate: ./mvnw -B -ntp clean verify -pl :graph-compose-core,:graph-compose-render-pdf,:graph-compose-render-docx,:graph-compose-render-pptx,:graph-compose-templates,:graph-compose-testing,:graph-compose-qa,:graph-compose-coverage -am gives BUILD SUCCESS (1791 + 127 tests).
  • render-docx runs 742 tests.
  • After install, examples run 93 green.
  • The knowledge checks pass.
• DocxPanelHeightTest (+5).
  • A card ending in a nested table hides that table's closing mark.
  • A table nested in a table in a card hides the inner closing mark too.
  • A card with text after its table writes that text into the paragraph that closed the table, and hides no mark.
  • A closing paragraph holding an anchored table's bookmark keeps its mark.
  • A closing paragraph holding the space below the table keeps its mark.
• Each rule fails its own test when reverted:
  • Not hiding the marks fails the first test.
  • Not recursing into nested tables fails the second.
  • Not checking the spacing fails the space-below test.
  • Not checking that the paragraph holds nothing fails the bookmark test.
• Template corpus (62 documents), against 2.5-dev: 33 documents change. Word renders all 33 as before. The page count stays the same in both editors, and no median drift in LibreOffice moves by more than 0.2pt.
  • CobaltRota in LibreOffice: p90 drift 3.86pt → 2.91. Its last row stands 3.5pt low instead of 5.2.
  • Better medians in LibreOffice: InvoiceMetered, InvoiceWorkspace, InvoicePayments, InvoiceSubscription, ReceiptModern.
  • A few medians in LibreOffice move by 0.04–0.08pt the other way: EngineeringResume, TealPulse, MerchantInvoice, ProposalIndigo, ObsidianInvoice.

Known limits

• A closing paragraph holding anything stays laid out at its tenth of a point in LibreOffice: space below the table, a bookmark, or the anchor of a drawing carried there.
• A reader who types into a hidden closing paragraph in Word types hidden text. The same holds for the paragraph closing a document, which was already hidden. Its exact hairline line already left no room to type in.
• Word's own 0.1–0.15pt a row in CobaltRota does not come from these paragraphs and is unchanged.

Lane: shared-engine (render-docx). No public API change.