fix(mcp): reduce Exa API context size sent to LLM (#2026)
Use highlights instead of full page text to minimize context size.
The previous implementation requested `text: true` which returned
entire page contents (5-20K+ chars per result), overwhelming LLM context.
Now uses:
- highlights: 3 sentences, 2 per URL (concise, query-relevant)
- text fallback: max 500 chars (only if highlights empty)
This matches what the MCP server at mcp.exa.ai likely returns.
