Group orphan raw-HTML opener/closer blocks around their markdown content

[srid]

Pandoc splits CommonMark "type 6" raw-HTML blocks at the next blank line, so markdown like

<details>

**bold** content

</details>

reaches the renderer as three separate blocks: an opener RawBlock carrying "<details>\n", a Para, and a closer RawBlock carrying "</details>\n". Today's rawNode wraps each raw blob in its own <rawhtml> element to keep xmlhtml's parser from mangling the bytes; the side effect is that the <details> open and close tags get trapped inside their wrappers, so the markdown paragraph ends up a sibling of the (now empty) details element rather than its child. See srid/emanote#433.

This PR adds a new Heist.Extra.Splices.Pandoc.RawHtmlGroup module exporting one function:

groupRawHtmlBlocks :: [B.Block] -> [B.Block]

It walks a block list and, when it sees an unbalanced opening tag followed downstream by a matching closing tag (depth counted only against opens of the same tag name), replaces that span with a B.Div carrying the tag in a "tag" directive attribute. The renderer's Div arm already turns Div with a "tag" attr into the named element via divTag, and now also calls stripTagDirective so the directive doesn't survive into serialised HTML as a literal tag="…". renderPandocWith applies the pass via Text.Pandoc.Walk.walk, so nested block lists (BlockQuote, list items, etc.) are covered too.

How the AST changes

Before: [Para "aaaa", RawBlock "<details>\n", Para "bbb",
         RawBlock "</details>\n", Para "eee"]

After:  [Para "aaaa",
         Div ("",[],[("tag","details")]) [Para "bbb"],
         Para "eee"]

Decisions worth surfacing

Question	Decision
Tags without a matching closer	Left as raw blocks — emitting a synthetic close would change the input's semantics
Tag names	Compared case-insensitively to follow HTML's own rules
Opener attributes (e.g. <details open>)	Accepted by the parser, dropped on the produced Div — deliberate scope until a real case demands otherwise
Self-closing forms (<br />)	Rejected by the opener parser so void elements never start a group
Single-block balanced fragments (<span>x</span>)	Left alone — the renderer already handles those correctly

Where the volatility lives

RawHtmlGroup owns the directive scheme end-to-end: tagDirectiveKey, divTag, and stripTagDirective all live there, and Render imports them as a consumer. Future evolution (alternative wrapping strategies, attribute preservation, void-element awareness, smarter nesting heuristics) starts in RawHtmlGroup so the renderer's interface stays stable.

Test coverage

13 unit tests pin the AST transformation: the issue's exact example, empty-group case, orphan opener/closer, consecutive pairs, same-tag nesting via depth counting, self-closing rejection, balanced-in-one-block rejection, case-insensitive matching, attribute tolerance, hyphenated custom-element names, mismatched-tag orphan, and the malformed-closer (missing >) regression. One end-to-end integration test through renderPandocWith asserts the markdown paragraph lands inside <details> with no <rawhtml> wrapper. Plus a regression test pinning that the tag directive doesn't leak as a literal HTML attribute.

Closes srid/emanote#433. Emanote PR bumping the pin: srid/emanote#686.

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

[srid]

Hickey/Lowy Analysis

#	Lens	Finding	Disposition
1	Hickey	closerTag silently matched when > is absent	Fixed in this PR
2	Hickey	getTag/dropTagAttr buried in a where clause as ad-hoc protocol	Fixed in this PR
3	Hickey	Opener/closer parsing asymmetry (inline vs where-bound)	No-op
4	Hickey	reverse acc accumulator pattern in splitAtMatchingCloser	No-op
5	Hickey	B.Figure arm doesn't strip the tag directive	No-op
6	Lowy	Module header should state its volatility axis	Fixed in this PR
7	Lowy	tag attribute protocol scattered as string literals across three sites	Fixed in this PR
8	Lowy	RawHtmlGroup exposed without an explicit reuse contract	Fixed in this PR
9	Lowy	dropTagAttr strip was anonymous; no named helper	Fixed in this PR
10	Lowy	Module name is functional, not volatility-named	No-op

Hickey rationale

Five findings on the diff. The one real correctness item — closerTag accepting a malformed </details (no >) because T.drop 1 (T.dropWhile (/= '>') after) would silently advance past a missing terminator — got fixed by mirroring the opener's T.stripPrefix ">" guard. Two non-issues sat on the structural side: the parseOpener/closerTag asymmetry is justified (opener is tag-agnostic, closer is tag-parameterised) and the reverse acc accumulator is the standard tail-recursive idiom. The getTag/dropTagAttr "buried in a where clause" item became Lowy's #7+#9 and got addressed there.

Lowy rationale

Five findings on the boundary. The boundary itself passes the "almost expendable" test — if Pandoc stops splitting orphan-HTML blocks, the entire RawHtmlGroup module is trivial to remove. What needed cleanup was the directive scheme leaking through three sites (RawHtmlGroup produces, Render consumes via two helpers buried in a where clause, and the test code hardcodes the key). Resolution: the "tag" key + resolver + stripper now all live in RawHtmlGroup (the producer that owns the wire format's volatility), and Render imports divTag/stripTagDirective as a consumer. The module header gained a Volatility & boundary section that names the axis of change so future maintainers know the boundary is load-bearing, plus a Public surface paragraph that mirrors the Render.Internal arrangement.

The one "No-op" item — RawHtmlGroup is a functional name, not a volatility-named one (OrphanHtmlRebalancer would have been more Lowy-aligned) — is noise: the boundary is correct, the doc comment explains the context, and renaming a public module name carries deprecation cost that doesn't earn its keep.