feat: add Exa AI-powered search tool

[tgonzalezc5]

Summary

The Exa search provider in forgegod/tools/web.py was sending a /search request without a contents block, so the response had no text/highlights/summary fields and r.get(\"text\", \"\") was always empty. In practice the Exa fallback was returning [{url, title, snippet: \"\"}] — useless to the researcher.

This PR aligns the implementation with the current Exa API and adds attribution.

• Request contents.highlights (3 sentences) and contents.text (1000 chars) so snippets are actually populated.
• Drop the legacy useAutoprompt field; type: \"auto\" already covers it.
• Add x-exa-integration: forgegod header for usage attribution.
• Cascade snippet extraction across highlights → summary → text via a small _exa_snippet() helper.

Usage

No config or API surface change. Existing ReconConfig.exa_api_key and the web_search(provider=\"exa\", ...) entrypoint behave the same — they just return non-empty snippets now.

from forgegod.tools.web import web_search

# Returns JSON list of {url, title, snippet}, with snippets populated
# from Exa highlights (falling back to summary/text).
results = await web_search(\"transformer architecture\", provider=\"exa\", exa_api_key=\"...\")

Files changed

• forgegod/tools/web.py — fix _search_exa request body, add integration header, add _exa_snippet() helper.
• tests/test_tools.py — add 7 unit tests covering the request shape, header, snippet fallback, and error paths.

Test plan

• pytest tests/test_tools.py — 22 passed (7 new + 15 existing)
• ruff check forgegod/ tests/ — clean
• No-API-key short-circuit returns []
• Request body includes contents.highlights + contents.text and headers include x-exa-integration: forgegod
• Snippet falls back across highlights → summary → text
• HTTP 5xx is caught and returns [] (no exception bubbled to caller)