Parse markdown adjacent to a raw HTML block line

[gordonwoodhull]

Since 0.29.0, a line of prose immediately following a raw HTML block line is emitted verbatim. Backticks, asterisks and every other inline construct survive into the rendered page as literal text:

<details>
<summary>Example custom instructions</summary>
This example demonstrates how a `quarto.instructions.md` file shapes …

Quarto 1 parses that, as does pandoc 3.8.1, and so did q2 0.28.0. A blank line is where the parser recovers, which makes it read like a content typo rather than an engine bug.

Cause

The lift added in bd-block-html-wrapped-in-p-w8qebxig turns a paragraph opening with a block-level HTML tag into a RawBlock by re-reading its source text verbatim. The extent is right — a markdown paragraph and a CommonMark type-6 HTML block both run to the next blank line — but re-reading the source also switches off inline parsing inside it.

That conflates two pandoc extensions. Dropping native_divs costs the <p> wrapper and the Div node, which was intended. Dropping markdown_in_html_blocks costs inline parsing, which was not. On a tight <div>, q2's output was byte-identical to pandoc -f markdown-native_divs-markdown_in_html_blocks, and its AST was a single verbatim RawBlock where -native_divs alone gives RawBlock/Plain/RawBlock.

Reader

process_paragraph already holds the fully-parsed inlines and was discarding them. It now partitions that vector: runs of block-level raw HTML coalesce into a RawBlock, text between them becomes a Plain, and the parts return as IntermediateSection, which every block container already splices. Classification is per-inline — no tag matching, no lookahead — so it stays inside the no-backtracking constraint in dev-docs/syntax-notes.md.

Two details are load-bearing. The tag test runs on trimmed text, because tree-sitter folds preceding whitespace into a non-first RawInline. And each part gets its own span; cloning the paragraph's onto all of them gives siblings overlapping ranges and misleads the incremental writer.

<pre>, <script>, <style> and <textarea> are exempt and stay verbatim — CommonMark's raw-text elements, which pandoc also leaves unparsed. All four are in BLOCK_TAGS, so without the exemption a split would parse markdown inside a <script>.

Writer

Block-level raw HTML is written bare rather than inside a ```{=html} fence, reversing the normalisation the lift originally shipped with. The fence is kept for everything else, gated on whether the text would re-read as the same single RawBlock — a pure run of block tags, or a raw-text element. Without that gate, an author's own ```{=html} block loses its fence and its contents are parsed on the next read: a backtick becomes Code, a # line becomes a Header.

Blocks are always separated by a blank line. Pandoc's writer is tight for RawBlock→Plain and Plain→RawBlock, but that is safe only because pandoc's reader guarantees Plain here means "inside an HTML block that is still open". q2 has no such invariant — Plain is emitted for every split interior, and filters emit them freely — so the writer cannot tell whether a Plain and an adjacent RawBlock came from one paragraph or two, and the AST does not record the answer.

Known divergences

• A split interior returns as a Paragraph after an AST → qmd → read cycle, gaining a <p>. For <div> this matches Quarto 1; for a phrasing-only element such as <summary> it is invalid, though only after a write/read cycle and never on first render. bd-8md6k9dv.
• Interiors are always Plain, never Paragraph. Pandoc chooses by tracking which element is open, which needs balanced matching. Every resulting difference is a <p> wrapper — the native_divs gap qmd already accepts. Against Quarto 1 this matches on every markdown-parsing question.
• A block-level tag mid-paragraph still does not interrupt the paragraph, so a trailing </details> lands inside a <p>. Pre-existing and byte-identical on 0.28.0, 0.29.0 and main. bd-495qnexy.

Notes for review

incremental_writer_tests::roundtrip_comment_in_blockquote no longer uses assert_roundtrip, which demands full structural equality after a rewrite. That guarantee no longer holds; the test now asserts tags in block position, prose parsed, and a fixed point from the first rewrite.

Two snapshots change, both the same class — html-comment-17 and html-comment-44, where a comment followed by same-line text becomes RawBlock + Plain instead of one verbatim RawBlock. Neither the old nor the new output carries the <p> Quarto 1 emits there; the comment simply moves onto its own line.

Workspace 13740 passed / 0 failed / 199 skipped (baseline 13719 at origin/main; +21 are the tests added here). Clippy, cargo fmt and cargo xtask lint clean; cargo xtask verify passes all 14 steps including the WASM leg. Both repro fixtures in q2-positron-docs/llms-info/repros/ are green simultaneously, and q2 matches Quarto 1 on all four rows of the regression repro.

[posit-snyk-bot]

✅ Snyk checks have passed. No issues have been found so far.

Status	Scan Engine	Critical	High	Medium	Low	Total (0)
✅	Open Source Security	0	0	0	0	0 issues
✅	Licenses	0	0	0	0	0 issues

💻 Catch issues earlier using the plugins for VS Code, JetBrains IDEs, Visual Studio, and Eclipse.