/**
* Markdown to the HTML a word processor will take.
*
* ## Why this exists at all
*
* The transcript on screen is *rendered* markdown: a table is box-drawing
* characters, a code block is a bordered panel, a heading is a coloured chip, a
* long line is hard-wrapped to the terminal's width. Selecting that with the
* mouse and pasting it into Word or Confluence pastes exactly what it looks
* like — dotted rules, broken table edges, wrapping frozen at eighty columns —
* because the terminal's clipboard carries glyphs and nothing else.
*
* What the user wanted to copy is the *structure*, and the structure is still
* in the markdown the model wrote. So the copy path goes back to that source
* and offers it twice: the markdown itself as plain text, and this, as HTML,
* for anything that knows how to take a rich paste. Word turns `` into a
* heading and `
` into a table; Confluence does the same; a plain text
* field gets the markdown and is none the wiser.
*
* ## What it is not
*
* Not a markdown implementation. It is the subset an agent transcript actually
* contains — headings, emphasis, code, links, lists, quotes, rules and GFM
* tables — rendered as the plainest HTML that survives a paste. Anything it
* does not recognise comes through as text, which is the failure a reader can
* still work with. There is no HTML passthrough: source markdown is untrusted
* input as far as this is concerned, every `<` is escaped, and a model that
* wrote `