fix(security): resolve reference-style images before the remote-image demotion in cleanMarkdown

[hivecommons-hive]

Security Fix

cleanMarkdown() in scripts/lib/architecture-content.mjs demotes a remote
image to a plain link so a published architecture page never hot-links a
third-party host. Its pattern matched only the inline form ![alt](dest).
CommonMark's three reference forms — full ![alt][label], collapsed
![alt][] and shortcut ![alt] — keep their destination in a link-reference
definition elsewhere in the document, so none of them were inspected, demoted
or rewritten and the compiled page rendered the submitted
<img src="https://third-party.example/..."> verbatim. Every visitor's browser
fetched it on page load, handing that host their IP, User-Agent and Referer.

findActiveContent() reports script-capable schemes and disallowed elements,
not remote resource references, so scripts/validate-architectures.mjs did not
catch it either. Both producers share this transform, so the unattended daily
cncf/architecture import and the community submission issue form were both
affected — escapeMdx() neutralises < and { but leaves !, [ and ]
alone.

What this changes

inlineImageReferences() resolves each reference form against the document's
link-reference definitions before the existing image rules run, so every image
form reaches the same remote-demotion and the same local-path rewrite:

• a remote destination is demoted to a plain link, in all three reference forms
  and for every spelling the inline rule already covers (absolute,
  protocol-relative, uppercase scheme, angle-bracketed);
• a cncf/artwork destination now reaches its mirrored /img/cncf-projects/...
  path, and a relative one its /img/architectures/<id>/... scope, instead of
  rendering broken;
• labels are matched the way CommonMark matches them (trimmed, lowercased,
  internal whitespace collapsed);
• a destination carrying whitespace or parentheses cannot be expressed in the
  [^\)]+ destinations the downstream rules match, so it fails closed: the !
  is dropped and the construct renders as a link reference, which fetches
  nothing.

The definition line is left in place — a definition renders as nothing, and it
may still be the target of a text link on the same page.

Verification

node --test over tests/architecture-content-clean-markdown.test.mjs,
architecture-content-mirror, import-architecture-issue,
import-architecture-issue-project-href, import-architectures,
import-architectures-fallbacks, validate-architectures and
architecture-catalog-contract: 130 pass, 0 fail. Eight new tests cover the
three reference forms, the destination spellings, label normalisation, the
artwork and relative rewrites, the fail-closed paren destination, a reference
with no definition, and a text link reference (which must not become an image).
npx prettier --check clean on both files.

End to end, the submission-issue repro from the issue now generates
[tracker](https://evil.example/pixel.png) where it previously generated
![tracker][beacon]. No checked-in page changes: docs/architectures/colopl.md
is the only page carrying link-reference definitions and they are all text
links.

Scope

Files: scripts/lib/architecture-content.mjs (cleanMarkdown,
new inlineImageReferences) and
tests/architecture-content-clean-markdown.test.mjs. Disjoint from the open
security PRs #1206 and #1213 (scripts/lib/svg-active-content.mjs,
tests/svg-active-content.test.mjs) and #1219
(scripts/lib/mdx-active-content.mjs, tests/mdx-active-content.test.mjs).

Closes #1228

---

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.

[mrbobbytables]

Superseded by #1245, which replaces the regex scanners with parser-based ones (the approach decided on #1194) and covers this fix, with regression tests. Closing in favour of that PR.