[harshmathurx]

Two extraction bugs found while preparing the launch video.

• Paul Graham's "Cities and Ambition" skipped its opening paragraph.
  • Cause: Readability invents <p>s with no live twin, so the live walk couldn't find its root. The classic fallback dropped the loose text "May 2008 Great cities…".
  • Fix: find the live root by matching Readability's longest paragraphs against live text nodes. The classic path now wraps loose text runs and strips a leading date-only token.
  • Result: the essay starts at "Great cities attract ambitious people.", "May 2008" isn't read, and the endnotes are now read at the end.
• Substack image captions were read as headings. Substack marks them up as an h6 right after the image. A short h5/h6 directly after a figure, image or video is now read as text.

Verification:

• 33-page eval identical to main (no recall, heading or junk change); paulgraham greatwork stays at 100%.
• New fixtures pg-comment-first, caption-h6; all fixtures pass.
• bun test 209 pass; highlight-check pass.
• Real-Kokoro e2e on WebGPU passes.

🤖 Generated with Claude Code