fix(security): stop the MDX code-span scan at a block boundary

[hivecommons-hive]

Security Fix

findActiveContent() (scripts/lib/mdx-active-content.mjs) is the gate that keeps
imported third-party Markdown from shipping live MDX into the published site. Before
scanning, it blanks inline code spans; the closer search crossed newlines and stopped
only at a blank line.

CommonMark parses inlines one leaf block at a time, and a blank line is not the only
thing that ends a block. A list item, an ATX heading, a block quote, a thematic break
or an HTML block interrupts a paragraph with no blank line present, so an unclosed
backtick at the end of one block cannot pair with a backtick in the next one.
blankInlineSpans() paired them anyway and blanked everything between, and the
payload was never scanned:

- a `
- <script>alert(1)</script> `

returned [] on main, while both the reference commonmark parser and micromark
(the engine Docusaurus/MDX uses) render the element live.

What changed

Both changes are in scripts/lib/mdx-active-content.mjs:

• New BLOCK_INTERRUPT_PATTERNS / startsNewBlock() — the line beginnings that
  start a new block: blank line, ATX heading, block quote, bullet and ordered list
  item, thematic break, and line-initial HTML block.
• blankInlineSpans() stops the closer search at the first line that begins a new
  block instead of at a blank line.

Two deliberate choices, both documented in the source:

• Code fences are absent from the list. blankFences() has already replaced every
  fence line with spaces by the time this runs, so a real fence is matched as a blank
  line. Adding a fence pattern would false-positive on ```<iframe>`, a line the
  real parser resolves into an ordinary code span (verified against micromark).
• Ordered lists match any start number, and every line-initial < matches, though
  CommonMark only lets 1. and HTML block types 1-6 interrupt a paragraph. Stopping
  early can only leave more text scanned — the fail-closed direction this module
  states as its rule.

Verification

• Differential against micromark: 66 documents (6 block prefixes x 11 following
  lines, payload <iframe src="https://evil.test">), comparing "rendered live by
  micromark" against "reported by findActiveContent" — 66/66 missed before this
  change, 0/66 after.
• Two new tests: one covering the block-boundary bypass across a list item, heading,
  ordered item, block quote, line-initial HTML block and thematic break; one pinning
  that a genuine lazy paragraph continuation still forms a span and reports nothing.
• node --test tests/mdx-active-content.test.mjs — 29/29 pass.
• npm run test:unit:coverage:check — exits 0; scripts/lib/mdx-active-content.mjs
  and tests/mdx-active-content.test.mjs both at 100.00% lines / 100.00% regions.
• npx prettier --check clean on both changed files.

Scope

Touches only scripts/lib/mdx-active-content.mjs and
tests/mdx-active-content.test.mjs; no workflow file is involved. Disjoint from every
open hold-gated PR: #1090 is scripts/lib/svg-active-content.mjs, #1112 is
package.json/package-lock.json, and #1104/#1110/#1113/#1114/#1117/#1120 are all
under tests/e2e*.

Closes #1121

---

Filed by sec-check agent (ACMM L4/L5 — hold-gated mode). Hold-gated: human review required.

— hive: agent=sec-check backend=copilot model=claude-opus-5 copilot=1.0.88

[hivecommons-hive]

Important

Held for human review by the hive's ACMM level gate.

This PR was opened by the "sec-check" agent while Hive policy required a human checkpoint for that agent. Non-outreach agents are held at ACMM L3–L5; the outreach agent is always held because it publishes project-facing communication.

Hive will keep the hold label until a human removes it. Operators can make a deliberate one-off release during an ACMM level change with release_level_holds=true, but level changes never release this hold automatically.