Nest markdown inside raw-HTML blocks separated by blank lines (#433)

[srid]

Markdown content between two blank-line-separated raw HTML tags now nests inside the surrounding element instead of escaping to be its sibling. Pandoc emits CommonMark "type 6" HTML blocks like <details> … markdown … </details> as three separate AST blocks (opener RawBlock, Para, closer RawBlock); without grouping, each raw blob ends up wrapped in its own element and the markdown paragraph drifts outside.

The actual fix is upstream in srid/heist-extra#16, which adds a groupRawHtmlBlocks AST-preprocess pass that rewrites those orphan triplets into a B.Div carrying the tag in a directive attribute — the existing Div renderer turns it into the named element with the markdown content as a real DOM child.

This PR pins emanote to that branch and notes the fix under the Bug fixes section of the unreleased changelog.

Closes #433.

Try it locally

nix run github:srid/emanote/empty-group

Generated by /do on Claude Code (model claude-opus-4-7).

[srid]

Evidence

The screenshots below were captured by serving the tests/fixtures/notebook/rawhtml-details.md fixture from a temporary notebook directory using two emanote builds: master (cade1054) and this PR (67ae6800).

Master (broken)	This PR (fixed)

Source HTML structure (the load-bearing diff — fetched via fetch() to bypass the browser's lenient parser):

• Master emits the <details> group as three sibling <rawhtml> wrappers around the opener, the markdown paragraph, and the closer:

  <rawhtml xmlhtmlRaw style='display: contents'><details>
  </rawhtml>
      <p ...>This paragraph ... marker</p>
  <rawhtml xmlhtmlRaw style='display: contents'></details>
  </rawhtml>

• This PR emits the <details> group as a single bare HTML block, with the markdown paragraph as a real direct child:

  <details>
      <p ...>This paragraph ... marker</p>
  </details>

Structural assertion (document.querySelector('[data-marker="RAWHTML_DETAILS_INNER"]').closest("details") !== null):

• Master: true
• This PR: true

Both report true because the browser's HTML parser is lenient enough to recover from the broken master output — the unknown <rawhtml> wrapper around <details> is reparented during DOM construction, so the marker still ends up with a <details> ancestor in the live DOM. The bug is in the emitted HTML, not in the recovered DOM. The before/after source-HTML diff above is therefore the authoritative structural evidence; the closest() assertion is preserved here for completeness but is not load-bearing for this issue.

Visual difference: The "before" screenshot shows the collapsed <details> widget on master — the markdown paragraph is hidden because the browser-recovered DOM does end up with it as a child, and <details> collapses its content when there's no <summary>. The "after" screenshot expands the details (details.setAttribute("open", "")) so the nested paragraph is visible: "This paragraph must render as a child of the <details> element. marker". The visible body text confirms the markdown paragraph renders as a child of <details> in the fix, with cleanly emitted HTML — no more interleaved <rawhtml> siblings around the opener/closer.

[srid]

/do results

Step	Status	Duration	Verification
sync	✓	0s	git fetch ok; forge=github
research	✓	11m 43s	Issue #433: orphan RawBlock html pairs in Pandoc AST get wrapped individually by heist-extra's rawNode. Fix belongs upstream in heist-extra.
branch	✓	15s	Working on existing empty-group branch; created group-orphan-rawhtml-blocks in heist-extra for the dependency fix.
implement	✓	7m 34s	Wrote groupRawHtmlBlocks pass in heist-extra (RawHtmlGroup module) + dropTagAttr fix in Render Div arm; 12 unit + 1 integration test. Updated emanote flake to point at it.
check	✓	9s	cabal build all green.
docs	✓	18s	Bug-fix entry under Unreleased in CHANGELOG.
fmt	✓	1m 12s	just fmt clean on emanote; hlint+fourmolu also clean on heist-extra.
commit	✓	39s	Heist-extra 970c9a3 pushed; emanote 8bca7a37 pushed.
hickey+lowy	✓	9m 41s	1 Hickey fix + 1 hoist (defer flipped); 1 Lowy fix + 3 hoist/doc (defers flipped). 3 follow-up commits on heist-extra.
police	✓	10m 46s	All 3 passes clean after applying: extracted parseTagAfterPrefix, rehomed tag-directive helpers, trimmed comments.
test	✓	53s	Emanote unit tests: 47/47. Heist-extra: 54 tests pinning AST + rendering.
create-pr	✓	2m 25s	heist-extra#16 + emanote#686 (drafts); Hickey/Lowy analysis on the upstream PR.
ci	✓	2m 32s	vira ci PASSED on aarch64-darwin + x86_64-linux. E2e-static/live/morph all green.
evidence	✓	10m 3s	Before/after screenshots + emitted-HTML diff posted. Tightened e2e regression to fetch raw HTTP and verify no <rawhtml> adjacent to <details> (confirmed it fails on master, passes on fix).
Total		58m 27s

Slowest step: research (11m 43s)

Optimization suggestions

• Research dominated (20% of wall time). Most of that was reading the heist-extra Render.hs / Render/Internal.hs to understand the getTag-as-directive convention before designing the grouping pass. For future bug fixes that touch a downstream library, pre-reading the relevant module(s) before invoking /do would shave the initial map step.
• Evidence step ran two sequential nix builds (master + this PR). The sub-agent could parallelise them (master build runs while feature build cache-hits), reclaiming roughly 2-3 minutes on bug-fix evidence captures.
• Police's elegance pass landed two real refactors (extract parseTagAfterPrefix, rehome the tag-directive helpers) that the initial draft missed. Worth running /simplify (or its equivalent) earlier in implement for non-trivial new modules — would catch the same findings before the first commit, avoiding the police-phase rework cycle.
• Re-running vira ci after the evidence-driven test tightening added ~3 minutes. For follow-up runs on the same branch where only test code changed, /do --from ci-only skips research → police and goes straight to CI.

Workflow completed at 2026-04-28 21:50 UTC.