markdown: link/image/autolink conformance improvements

[jonathanKingston]

Focused, low-risk CommonMark conformance improvements for inline links, link reference definitions, images, and autolinks. Refs #10 (partial batch — does not close the issue).

Per-section conformance (before → after)

Section	Before	After
Links	72/90	77/90
Images	17/22	18/22
Autolinks	14/19	16/19
Link reference definitions	18/27	18/27 (unchanged)

Total pinned passing examples: 510 → 518, additive only (no section decreased; the passing array only gains numbers).

Newly passing spec examples

• Autolinks 597, 603
• Links 547, 548, 551, 552, 568
• Images 590

What changed

Autolinks (src/inline-autolinks.ts) — rewrote angle-autolink parsing to the CommonMark URI/email grammar:

• Try the URI (scheme) production before the email production, so <MAILTO:foo@bar.baz> is a scheme autolink with the href kept verbatim rather than being mangled into mailto:MAILTO:... (597).
• Use the spec's email-address grammar; the local part no longer allows a backslash, so <foo\+@bar.example.com> is not an email autolink (606 — see gaps below).
• Backslashes are no longer treated as escapes inside an autolink, so <https://example.com/\[\> is a valid autolink (603).

Destination encoding (src/link-references.ts) — percent-encode [ and ] in href output (%5B/%5D), matching the reference renderer (603).

Inline links (src/inline-links.ts) — when a [label](...) inline destination fails to parse, fall through to shortcut-reference resolution instead of bailing out, so [foo](not a link) links [foo] and keeps (not a link) as literal text (568).

Reference label validity (src/link-references.ts) — reject empty, whitespace-only, and bracket-nesting reference labels for both definitions and lookups, so [], [[[foo]]], and [ref[bar]] no longer form links or definitions (547, 548, 551, 552, 590). Escaped brackets ([Foo*bar\]]) remain valid.

Tests

• Added unit tests for each behavior in inline-autolinks.test.ts, inline-links.test.ts, and link-references.test.ts (not relying solely on the conformance baseline).
• Regenerated tests/fixtures/commonmark/conformance-baseline.json (additive only).
• npm run typecheck and npm test (385 tests) fully green.

Remaining gaps (not addressed here)

• Autolinks 606, 608, 611: 606 still renders the backslash literally (<foo\+@...> → <foo\+@...) because encodeBackslashEscapes treats the malformed email as a verbatim region; 608/611 involve the app's bare-URL autolinking, which is a deliberate extension over CommonMark.
• Reference labels with emphasis (Links 554/558/559, Images 573/576/585/589): these need reference resolution on the raw label before emphasis runs; deferred to avoid a risky pipeline reorder.
• Link reference definitions (multiline/interrupted defs, defs inside code fences, defs after paragraph text, case-folding ẞ→SS): these are block-tokenizer concerns and left for a separate change.
• Raw-HTML precedence cases (Links 491/493/522/524/526/535/536/538) remain divergent by design (the renderer escapes raw HTML).

🤖 Generated with Claude Code

https://claude.ai/code/session_01Cv1iyJb9PHMFG3YGp6ffkT

---

Generated by Claude Code

[jonathanKingston]

Code review (high effort) — no correctness bugs found. ✅

Verified:

• Autolink safety parity: the new autolinkHref keeps the same ^(?:javascript|data|vbscript): blocklist as safeLinkHref, and skips only the escape-decoding/trim (which is correct — autolinks don't process backslash escapes and can't contain whitespace). <javascript:alert(1)>, <data:…>, <vbscript:…> still return null. No security regression.
• URI-before-email ordering gives <MAILTO:foo@bar> a verbatim scheme href (spec 597), and the anchored [^\s<>]* can't overrun the closing >.
• isValidReferenceLabel is escape-aware and applied symmetrically to both definition storage and lookup, so []/[[foo]]/[ref[bar]] are consistently rejected; escaped brackets ([foo\[bar]) stay valid.
• Link fall-through (#568) correctly resolves [foo](not a link) as a shortcut ref with end at the label close, leaving the parens as literal text — the spec-correct behavior, and it doesn't mis-consume for images either.

Low-severity note (non-blocking): dropping [/] from encodeHrefForOutput's allowed set matches cmark (verified), but it now percent-encodes brackets in every href — including an IPv6-literal host like http://[::1]/ → http://%5B::1%5D/, which browsers reject as a host. It's spec-conformant and IPv6 URLs in chat are vanishingly rare, so this is just a conscious-tradeoff heads-up, not a change request.

---

Generated by Claude Code