Converter P0+P1: tables, images, escape, h4-h6, hr, mark/sub/sup, code lang hints + Vitest harness + CI

[dewierwan]

Summary

Aimed at making paste-to-markdown handle real Notion / Airtable / Google Docs pastes without dropping content or silently corrupting it. Plan reference: ~/.claude/plans/great-now-what-else-sprightly-pony.md.

P0 — correctness and core coverage

• Tables → GFM tables — biggest gap. Walks thead/tbody/tr/th/td, falls back to "first row is header" if no thead, escapes | in cells, recurses for inline formatting.
• Escape Markdown special chars in text nodes — was silent corruption: "5 * 3 = 15" became italic, "John_Smith" became "JohnSmith", "[draft]" became a broken link. Now escapes \ ` * _ [ ] ~ in text nodes only (not inside <code> / <pre>).
• Images → ![alt](src) — were dropped entirely. Title attribute included if present.
• Headings h4–h6 — were falling through to default and emitting plain text.
• <hr> → --- — was dropped.

P1 quick wins

• <mark>, <sub>, <sup> emit HTML pass-through (renders anywhere Markdown allows raw HTML).
• <pre> reads data-language (Notion) and class="language-xyz" (Prism convention) for fenced-code language hints.
• <del> joins <s> / <strike> for strikethrough.
• Final whitespace pass collapses runs of 3+ blank lines to two (Notion's nested <div>s used to produce big gaps).

P2 — testability + CI

• Vitest + jsdom harness loads index.html itself as the artifact under test, so tests exercise the actual published page rather than a duplicate. End users still just open index.html — no install needed.
• 34 tests across headings, inline formatting, escaping, images, tables, code blocks, lists, links, and whitespace cleanup.
• Fixture auto-discovery — drop tests/fixtures/<source>/<scenario>.html paired with <scenario>.expected.md and the harness picks them up. Highest-leverage investment for regression safety as new sources are supported.
• GitHub Actions CI — runs cspell, html-validate, and vitest on every PR.

Out of scope for this PR (P3, design-coordinated)

• "Paste as plain text" toggle
• Source-detection pill
• Don't-auto-clear input behaviour
• Google Docs class-based formatting (currently relies on inline-style fallback)

Test plan

• npm test — all 34 tests pass locally
• npm run lint:spell — clean
• npm run lint:html — clean
• CI green on this PR
• Live site smoke test: paste a Notion page with a table, image, h4, hr, and */_/[ in prose; confirm output is correct

🤖 Generated with Claude Code

[dewierwan]

@claude review