fix(mcp): reduce Exa API context size sent to LLM

[gary149]

Summary

• Use highlights instead of full page text to minimize context size
• Add text fallback with 500 char limit when highlights are empty

Problem

The direct Exa API integration (added in #2021) was requesting text: true which returned entire page contents (5-20K+ characters per result). This overwhelmed the LLM context compared to the original MCP server behavior.

Solution

Changed content request from full text to:

• highlights: 3 sentences, 2 per URL (concise, query-relevant snippets)
• text fallback: max 500 chars (only used if highlights are empty)

Before	After
~5,000-20,000+ chars per result	~300-600 chars per result
10 results = 50-200K chars	10 results = 3-6K chars

Test plan

• Test web search with Exa to verify results still contain useful content
• Verify LLM responses are more focused/better quality with reduced context