fix: map spans to their source and fix CommonMark parse and round-trip defects

[plimeor]

Summary

This implements the plan source-span-mapping-and-commonmark-fixes, now archived at docs/archive/2026-10-06-source-span-mapping-and-commonmark-fixes/. The plan's spec changes are merged into docs/specs/.

Closes #6.

Source spans

Every parsed node's span maps to the source bytes it was read from:

• on every line of every container;
• across CRLF;
• across split tabs;
• inside table cells, which now carry spans.

Spans nest within their parent. The issue's 31-case regression test passes.

Parsing

The parser now matches commonmark.js and micromark where they agree on:

• Emphasis: the rule of three and opener floor use original run lengths.
• Image references: an invalid image resource falls back to a reference.
• Containers: lazy lines and blank lines in block quotes and list items, including lazy list markers and an HTML tag on a lazy line.
• Paragraph interruption: ATX-like lines and backtick runs.
• Open blocks in quotes: unclosed fences, math, and HTML blocks.
• Tab columns at any container depth.
• Whitespace: only spaces and tabs count as block whitespace.
• Labels: label matching, footnote labels, and angle-autolink URIs.
• Breaks: a hard break comes only from source spaces.
• Processing instructions: <?> is text.
• Definitions: a list after a definition.
• GFM tables: a table starts below a delimiter row that is neither lazy nor a setext underline, before any block its header row would open.

Serialization

Many more parsed documents serialize to Markdown that reads back to the same tree:

• Text is escaped against every construct beside it.
• A paragraph or heading that does not read back tries other delimiter choices.

Canonical output changes

• Every backtick in text is written as \`.
• A few constructs move to spellings that read back. The regenerated goldens show each one.

Performance

• Parsing the 400 KB fixture document takes 137M instructions, against 261M before.
• Serialization takes about 51M, against 46M.
• Growth tests now check linear time for:
  • nested emphasis;
  • long ~ runs;
  • inline diagnostics that run to a paragraph's end.

Known issues

The final review left findings open. They were accepted for archive and are tracked in #11:

• A panic that predates this branch: parse("\u{a0}e+@"). It should be fixed before the next release.
• 13 nested quote-and-list inputs that the group-19 column tracking newly breaks. The same tracking fixes 378 others.
• Five round-trip classes listed under the plan's Risks.

#11 replaces the container re-prediction and the serializer's copies of parser rules at their root, and folds autolinks into Link.

Verification

• cargo fmt --check, cargo test with and without html, RUSTDOCFLAGS='-D warnings' cargo doc --no-deps, cargo build --target wasm32-unknown-unknown, cargo +1.82 build, and the release pathological tests all pass.
• Conformance stays at 2233 of 2236.
• The parse comparisons against commonmark.js and micromark, on the inputs where they agree, are recorded per task in the archived plan.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Hyap2mHhP25kLtUxRS8XZ3

---

Generated by Claude Code