feat(markdown): open a markdown file as a text document

[andiwand]

🤖 Generated with Claude Code

Implements stages 1 to 3 of the markdown plan that was already in the repo
(src/odr/internal/markdown/PLAN.md). A .md file decodes to a TextRoot, so
the generic html renderer and every binding get it without format-specific code
— rather than a markdown→html renderer next to html/text_file.cpp, which
would produce html only, with no element api and nothing for
JNI/embind/pybind/ObjC.

CommonMark plus the GitHub extensions, parsed with md4c (new dependency,
md4c/0.5.2, MIT): headings, paragraphs, soft and hard breaks, lists (ordered
counting from their start), block quotes, code blocks, emphasis, strong,
strikethrough, inline code, links, GFM tables and task lists.

FileType::markdown moves from FileCategory::text to
FileCategory::document with DocumentType::text. detect_by_content stays
false — markdown has no signature, and a content probe for it is a probe
for "prose with occasional punctuation" — so the only way in is
DecodedFile(file, FileType::markdown), and a .md still opens as a text file
by default. Callers route on the file name, which is what they already have.

Design rationale and the mapping table are in the new
src/odr/internal/markdown/AGENTS.md;
PLAN.md is marked up with what it got wrong.

Three things the plan had wrong, found by rendering a sample

The element-tree tests would not have caught any of these on their own.

1. A heading needs a bold span, not just a bold paragraph style.
   html::translate_paragraph takes only font family and size from a
   paragraph's text style (translate_block_font_style) — weight, slant and
   decoration are expected on the spans inside. Headings rendered unbolded.
   Only the weight goes on the span: the paragraph's size is already in em,
   and repeating it compounded 2em into 4em. A TH cell has the same
   shape.
2. A tight list item needs a paragraph opened for it. md4c omits
   MD_BLOCK_P inside a tight list item, and html::translate_list_item
   writes the marker into the item's first paragraph — so without one, every
   bullet was silently dropped.
3. Table columns cannot be children of the table. The rows are; one sibling
   chain cannot carry both, and the renderer walks table_first_column's
   siblings, so a 3-column table emitted six <col>s — three columns and three
   rows. Now on their own chain via append_column, the shape odf uses.

The plan's fourth claim — that html::translate_list hardcodes <ul> — is
stale; the renderer takes the label from list_item_marker now, so ordered
lists number themselves.

Known gaps, documented rather than dropped quietly

• Named entities beyond the five predefined XML ones (plus  ) stay
  literal. Numeric references resolve. The plan asked whether the conan
  package exports the md4c-html component whose entity_lookup resolves the
  ~2000 named entities: it does, and the symbol is in the archive, but
  entity.h is not shipped — so using it means declaring md4c's private
  MD_ENTITY layout ourselves, where a mismatch corrupts silently instead of
  failing to link. Left out; worth revisiting before claiming CommonMark
  conformance.
• Raw html is dropped, block and inline. There is no passthrough element,
  and inventing one means deciding what Text::content() returns for it in
  four bindings.
• Images are transparent — the alt text flows through as text. Stage 4.
  Note for whoever picks it up: html::translate_image writes the <img> at
  width:100%;height:100% absolutely positioned inside the frame's div, so a
  frame with no dimensions renders nothing at all, and markdown carries none.
• Horizontal rules have nothing in the model.
• A hyperlink's href is passed through and only escape_attributed by the
  renderer, exactly as an odt's is — this is a second consumer of the existing
  document-link policy, not a third policy.

Testing

20 gtests in test/src/internal/markdown/markdown_file_test.cpp, all inline
string literals in and element tree out, no fixture files. The rendered html
was eyeballed end to end for a document exercising every construct above.

FileTypeCapabilities.declaration_matches_the_engines and
color_scheme_matches_the_html now exercise the .md already in odr-public,
so nothing needs adding to the test-data repo.

Before merging

• The reference output needs regenerating. odr-public/README.md was
  skipped by html_output_test while the row declared open = false, and now
  renders.
• The commit subject is feat, i.e. a minor bump. The category flip is
  api-visible, but the row was classification-only and declared no capabilities,
  so nothing could hold a decoded markdown file to be broken — the plan's own
  reasoning. If you read it as breaking, the subject needs a ! and this
  becomes v7.0.0.