fix(docx): name in the report a paragraph or list item the page reads as markdown

[DemchaAV]

Why

A session reads markdown unless it is told not to: GraphCompose.document() defaults to markdown(true).

• What the page does. It reads a paragraph, or a list item, of plain text holding a mark of emphasis or code (*, _, `) as markdown, through flexmark (MarkDownParser):
  • it sets the text the emphasis marks style bold or italic, and a heading bold, the first three levels larger;
  • it drops the marks its parser reads as syntax, a code span's backticks and a link's address among them.
• What the file did. The DOCX export writes the text as authored, so Word showed **bold** with its asterisks and none of the bold. The report named none of it, and the markdown flag never reaches a semantic backend.
• The corpus has one such paragraph. TimelineMinimal's open-source project line ends in *(Open source)*: the page sets it in italic, and Word showed the asterisks.

What changed

• markdownLost names a paragraph the page read as markdown. It is added to the paragraph's note on every body path (paragraphLost: the body, a text box, a side of a pair, a badge) and to a zone paragraph's, on its page zone note (zoneParagraphLost). The phrase: "its markdown marks are written as letters, where the page sets the text they mark and drops them".
• itemsMarkdownLost names a list's items, on the list's note: "its items' markdown marks are written as letters, …".
• How it is read (marksDropped, shared by both).
  • The page's own trigger comes first (holdsAMarkdownTrigger, readsAsMarkdown): plain text, no runs, holding *, _ or `. Only then are the laid-out lines read.
  • The lines hold fewer of the characters markdown syntax is made of (*_`\#[]()>~) than the text, as the page lays it out, where the page read it. Both sides count the same characters, so one the page keeps counts on both.
  • The text is counted as the page lays it out:
    • a paragraph's prefix before the first line has its marks left out;
    • a list item has a marker typed before it taken off (ListMarker.normalizeItemText), as both the page and the file take it off;
    • a list's lines are its items' (itemLines), without a marker laid out on its own.
  • Lines that hold no text, as a paragraph given no width lays out, are not counted.
  • So text the page sets as authored is not named:
    • markdown off;
    • an underscore inside a word, which markdown keeps;
    • a marker typed before a list item;
    • runs, which the page never reads.
• Where the lines are not read, it says so — if the page's parser would drop a mark (parserDropsAMark runs MarkDownParser on the text). That happens with no layout, or composed in a table cell: a paragraph there is matched to its lines by its text, and a list is not matched at all. The phrase then ends "— whether the page reads them is not measured". A cell holding node_js is not named, since the parser keeps that underscore.
• The file is not changed. It still writes the text as authored, and the docs say so.
• Ledger. The entries that now name the loss:
  • a paragraph's text moves from WRITTEN to REPORTED;
  • a list's nestedItems moves from WRITTEN to REPORTED;
  • a list's items, already REPORTED, names it too.
• Docs. These say what is named, that markdown is the default, and that the file does not yet write the text as the page sets it:
  • the recipe's Paragraphs row and its list-report paragraph;
  • the capability matrix's paragraph and list rows;
  • render-docx/README.md;
  • the CHANGELOG.

Verification

• ./mvnw -B -ntp install -pl :graph-compose-render-docx → BUILD SUCCESS: 1096 tests, 0 failures, 1 skipped (the property-gated fidelity probe).
• DocxMarkdownReportTest is new, with 5 tests.
  • Named:
    • a paragraph read as markdown;
    • a heading;
    • a code span alone;
    • one behind a prefix with a mark of its own, and with as many marks as the text drops;
    • a list whose item is read: flat, with markers in a column of their own, and in a tree of items;
    • a zone paragraph, on the zone's note;
    • not measured:
      • a paragraph with no layout;
      • a paragraph composed in a table cell;
      • a list composed in a table cell.
  • Not named:
    • markdown off, for a paragraph, a list and a zone;
    • no mark, with a layout and without;
    • an underscore inside a word, in a paragraph and in a table cell's paragraph and list;
    • runs;
    • a prefix's mark with markdown off;
    • a paragraph given no width;
    • a marker typed before a list item, with markdown on and off.
• The tests fail without the code they cover. 11 sabotages were run, each breaking one thing, and each made the tests that cover it fail:
  • the body, the zone and the list phrase;
  • the comparison, never and at equal counts;
  • the not-measured phrase, never and always;
  • the prefix's marks;
  • a typed marker kept;
  • a tree of items left unread;
  • lines with no text counted.
• The DOCX bytes do not change. The 62 corpus documents exported deterministically are byte-identical to the export before the change: DocxFidelityCorpusTest -Dgraphcompose.docxFidelity=export, SHA-256 per file, 0 of 62 differ.
• Across the corpus the report names one paragraph more, TimelineMinimal's project line: 951 notes, from 950.
• Documentation guards:
  • core: -pl :graph-compose-core -Dtest='com.demcha.documentation.**' → 166 tests, 0 failures;
  • qa: the documentation guards plus DocxPageZoneTest, DocxTransparentWrapperTest, TimelineRailAcrossBackendsTest and RtlAcrossBackendsTest → 50 tests, 0 failures.
• The full reactor gate was not run; no public signature, POM or workflow file changed.

Known limits

• A list marker holding marks can hide the loss. A legacy-layout list sets its marker in each item's line, so ListMarker.custom("*") or "(a)" adds marks to the lines.
• Composed in a table cell, a paragraph or a list is "not measured" wherever the parser would drop a mark, markdown on or off, since its lines are not read there.
• Some of what markdown changes is not counted. A 1. line read as a list item, a --- rule and an entity like & hold none of the counted characters.
• The fix — writing the runs the page sets — is not made here. The export still writes the text as authored.

Lane: render-docx backend (report only, no change to what is written) plus tests and docs.