`renderMarkdown` inserts blank lines inside raw-HTML elements, splitting them on reparse

[sandros94]

Environment

• Operating System: Linux (WSL2)
• Node Version: v24.17.0
• Package Manager: pnpm@11.11.0
• Comark Version: 0.5.1 (reproduces on current main)
• Framework: None

Comark Version

v0.5.1

Reproduction

import { parse } from 'comark'
import { renderMarkdown } from 'comark/render'

const md = '<picture>\n<source srcset="a.avif">\n<img src="a.jpg">\n</picture>'
const rendered = await renderMarkdown(await parse(md))
// '<picture>\n<source srcset="a.avif" />\n\n<img src="a.jpg" />\n</picture>'
//                                        ^^ blank line

console.log(JSON.stringify((await parse(rendered)).nodes))
// [["picture",{…}, ["source",…]], ["img",…]]   ← img split out of the element

Description

The child-join loop in the html handler joins a raw-HTML element's children with blockSeparator whenever a child classifies as block (the catch-all arm classifies any unlisted tag as block, so source, figcaption, option… all hit it). A blank line terminates the HTML block per CommonMark, so the rendered markdown can't reparse to the original tree — parse → renderMarkdown → parse splits the element.

Anything with ≥2 children where one is not a known inline tag fails: figure>img+figcaption, details>summary+p, dl, select, audio with two sources, nested divs. Single-child elements escape it.

The inline variant is worse: text <picture><source…><img…></picture> tail gets the separator too (the loop runs regardless of oneLiner), which splits the containing paragraph in two on reparse.

Since the parser never nests markdown blocks inside a $.html element (blank-line-separated markdown is parsed outside the element), a blank line inside one is never legitimate in markdown output — children should be joined with \n, or nothing at all in inline position.

Fix + tests ready locally, PR incoming

Additional context

Parse-side relative: #235 / #236 are about the parser tolerating hand-written blank lines inside <details>; this one is about the stringifier introducing them. No overlap in touched files.

Filing two more round-trip issues alongside this one, all independent, will cross-link once up

Logs