fix: break content truncation at semantic boundaries

[ntegrals]

Summary

When extractMarkdown() exceeds maxLength, the previous logic only broke at a paragraph boundary if it was in the last 20% of content. Otherwise it sliced mid-word/sentence, producing broken markdown.

Before

# Some Article

This is a long paragraph about...  ← sliced here mid-sentence

[... content truncated, ~500 chars remaining]

After

The truncation now tries boundaries in priority order:

1. Paragraph break (\n\n) — cleanest cut
2. Sentence ending (. , .\n, ? , ! ) — preserves complete thoughts
3. Word boundary (space) — avoids mid-word cuts
4. Hard limit — only if no boundary found in the first 50%

All boundaries must be at least 50% into the content to avoid over-truncation.

Code change

packages/core/src/page/content-extractor.ts lines 247-268

Test plan

• bun run build — compiles clean
• bun run test — all 364 tests pass