fix: flip attention marker to avoid nested-emphasis fusion

[ChristianMurphy]

Initial checklist

• I read the support docs
• I read the contributing guide
• I agree to follow the code of conduct
• I searched issues and discussions and couldn’t find anything or linked relevant results below
• I made sure the docs are up to date
• I included tests (or that’s not needed)

Description of changes

Problem: emphasis uses the configured marker unconditionally, so
emphasis > emphasis > text serializes as **a**, which re-parses
as strong > text. The same path erases the outer type from
strong > emphasis > text (***a*** re-parses as emphasis > strong)
and collapses any strict three-or-more nested emphasis chain
(***a*** re-parses as emphasis > strong rather than three
nested emphases).

#12 catalogued a broader family of related shapes:
***emphasis*in emphasis*, *a*_b__, a ***b*c d*, and variants
with different leading run lengths. As wooorm noted in the thread,
"pulling a thread somewhere will have something happen somewhere
entirely different": marker choice in one place interacts with
flanking rules elsewhere, and covering every shape needs
coordination beyond picking a single marker per node.

Scope: land the minimal change that fixes the shapes guaranteed to
drift under the current serializer, without regressing shapes where
CommonMark's attention algorithm already recovers the original tree
through fusion. Escape-based work on the remaining #12 shapes is
left for follow-up.

Approach: introduce lib/util/emphasis-marker.js. Both the
emphasis handler and its peek route marker selection through it,
keeping lookahead in container-phrasing consistent with what the
handler emits. The helper flips in two narrow situations:

1. The emphasis is the only child of an attention parent (emphasis
   or strong), and both its opening and closing markers would sit
   immediately next to the parent's primary marker. Using the
   opposite marker (*_a_*, **_a_**) breaks the fusion into
   strong or em+strong.

2. The emphasis sits at the top of a strict same-type chain of
   depth two or more (every link has exactly one emphasis child)
   when the primary marker is *. Three-deep emphasis only
   round-trips with _ on the outside, because _'s flanking
   rules are stricter than *'s. The rule is asymmetric on
   purpose: with primary _ the first rule's adjacency flip
   alternates correctly on its own.

Strong is never flipped. A run of four asterisks already pairs as
two strong delimiters, six as three, and so on, so strong
round-trips without help.

Journey (what was tried and why the scope narrowed):

• An earlier iteration flipped strong too and regressed ~18 corpus
  fixtures whose nested-strong shapes relied on long fused runs of
  asterisks. Strong was dropped from the flip to recover them.

• Flipping whenever info.before or info.after matches the
  primary caused cascading flips on paragraph-level attention
  siblings: [emphasis, strong, emphasis] serialized as
  _a___a__*a*, where _ + __ at the em/strong boundary
  re-tokenised as a single ___ run. The flip was narrowed to
  attention parents only.

• Widening rule 1 to first-or-last-child of any attention parent
  fixed several roundtripping emphasis in emphasis edge case changes document structure #12 shapes (***emphasis*in emphasis*,
  ***x*y z*, ****x*y z*) but regressed ***a*a*-*
  (emphasis > [emphasis > [emphasis, text], text]): CommonMark's
  rule 17 uses the leading *** fusion to recover the three-deep
  structure, and the flip broke that recovery. The rule was
  tightened to only-child plus strict-chain.

• The only-child formulation plus the strict-chain rule is the
  widest version verified to cause zero transitions from ok to
  finding across 600 corpus files (commonmark, gfm, all
  configurations).

Edge cases covered by new tests:

• Plain emphasis and strong, with primaries * and _, showing
  the helper is inert on non-nested attention.
• emphasis > emphasis with each of primary * and _, yielding
  *_a_* and _*a*_.
• strong > emphasis yielding **_a_**.
• emphasis > strong and strong > strong preserved at ***a***
  and ****a****, proving strong is untouched.
• Strict three-deep emphasis chains with both primaries, both
  yielding _*_a_*_ (chain flip vs adjacency flip arrive at the
  same output by different routes).
• Emphasis parents with more than one child, demonstrating the
  only-child guard preserves shapes vanilla handles.
• Middle-sibling emphasis, confirming no flip at non-boundary
  positions.
• Top-level [emphasis, strong, emphasis] round-trip preserved.
• ***a*a*-* fusion shape explicitly preserved as a regression
  guard against future widening.
• Round-trips for parsed *_a_*, parsed _*_a_*_, synthesised
  emphasis > emphasis, synthesised strong > emphasis, and a
  three-deep chain preceded by a text sibling.

Refs: #12

Disclosure: written with AI assistance

[Copilot]

Pull request overview

Fixes markdown round-tripping for nested attention (emphasis/strong) by choosing markers that avoid delimiter fusion and CommonMark re-tokenization, addressing the edge case described in #12.

Changes:

• Added a new utility to pick * vs _ per emphasis/strong node based on surrounding adjacency and a three-deep boundary-chain rule.
• Updated emphasis and strong handlers to use the new marker-selection utility.
• Added tests for nested attention marker flipping and for some round-trip scenarios.

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 3 comments.

File	Description
lib/util/emphasis-strong-marker.js	New marker-selection logic to avoid attention delimiter fusion and handle three-deep same-type boundary chains.
lib/handle/emphasis.js	Uses the new utility to choose the emphasis marker at serialization time.
lib/handle/strong.js	Uses the new utility to choose the strong marker at serialization time.
test/index.js	Adds tests covering nested attention marker flipping and round-trip behavior for selected cases.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

Pull request overview

This PR adjusts how emphasis delimiters are selected during serialization to prevent CommonMark attention delimiter fusion from changing the parsed AST shape (notably for nested emphasis), and adds regression tests for the affected roundtrip cases.

Changes:

• Add lib/util/emphasis-marker.js to centralize emphasis marker selection and selectively flip markers in specific nesting scenarios.
• Update the emphasis handler (and its peek) to use the new marker-selection helper.
• Expand the emphasis test suite with targeted cases and roundtrip guards for nested attention edge cases.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 1 comment.

File	Description
test/index.js	Adds new serialization and roundtrip tests covering nested emphasis/strong and known fusion edge cases.
lib/util/emphasis-marker.js	Introduces helper to choose * vs _ for emphasis to avoid delimiter fusion changing parse results.
lib/handle/emphasis.js	Routes emphasis delimiter selection through the new helper for both handler and peek behavior.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[ChristianMurphy]

observation is correct: strictChainDepth(node) >= 2 does match multiple nodes in a strict chain of depth 4 or deeper, which I'll grant wasn't by explicit design. But applying the proposed gating is a regression, not an improvement. Comparison across depths:

depth	current output	current reparses as	copilot output	copilot reparses as
2	*_a_*	em > em > text("a") ✅	*_a_*	em > em > text("a") ✅
3	_*_a_*_	em > em > em > text("a") ✅	_*_a_*_	same ✅
4	__*_a_*__	strong > em > em > text("a")	_*_*a*_*_	[em(text("*")), em("a"), em(text("*"))]
5	___*_a_*___	em > strong > em > em > text("a")	_*_*_a_*_*_	4 flat em with literal * text
6	____*_a_*____	strong > strong > em > em > text("a")	_*_*_*a*_*_*_	4 flat em with literal * text

Both approaches drift at depth 4+, CommonMark has no round-tripping representation for a strict 4-deep emphasis chain. But the current output is strictly more faithful to the input tree:

• The text content ("a") is preserved.
• Attention nesting is preserved (the inner em > em survives).
• Only the outermost element type changes (em → strong).

The proposed fix:

• Flattens the nesting entirely into sibling em nodes.
• Injects literal * or _ characters into text nodes that weren't in the original tree (content corruption).

So the __...__ at the outer boundary looks alarming, but it's producing the least-bad degradation at depth 4+. I'd rather keep the accidental over-match and document it than trade it for a drift that corrupts content.

[Murderlon]

Looks high-quality 👍 . Double checked with Devin review as well and no issues found: https://app.devin.ai/review/syntax-tree/mdast-util-to-markdown/pull/74