feat: migrate search handler to support Tavily alongside Serper

[tavily-integrations]

Summary

• Added Tavily as a configurable alternative search provider alongside the existing Serper (Google Search) integration
• Both providers remain active; callers choose via Hydra agent YAML config
• Tavily results are normalized to match the existing Serper organic result structure (title, link, snippet)

Files changed

New files

• libs/miroflow-tools/src/miroflow_tools/mcp_servers/tavily_mcp_server.py — FastMCP server exposing tavily_search tool via the tavily-python SDK, mirrors serper_mcp_server.py interface
• libs/miroflow-tools/src/miroflow_tools/mcp_servers/searching_tavily_mcp_server.py — Orchestrator wrapper with retry logic, mirrors searching_google_mcp_server.py
• apps/miroflow-agent/conf/agent/tavily_search.yaml — Agent config variant using tool-tavily-search instead of tool-google-search in the browsing sub-agent

Modified files

• libs/miroflow-tools/pyproject.toml — Added tavily-python dependency
• apps/miroflow-agent/src/config/settings.py — Added TAVILY_API_KEY env var, tool-tavily-search MCP server config block, and has_tavily_api_key to env info
• apps/miroflow-agent/.env.example — Added TAVILY_API_KEY entry
• apps/gradio-demo/.env.example — Added TAVILY_API_KEY entry

Dependency changes

• Added tavily-python to libs/miroflow-tools/pyproject.toml

Environment variable changes

• Added TAVILY_API_KEY to apps/miroflow-agent/.env.example and apps/gradio-demo/.env.example

Notes for reviewers

• Existing Serper files are completely untouched
• The Tavily MCP server normalizes results to match Serper's organic result format (title/link/snippet) for downstream compatibility
• HuggingFace dataset/space URL filtering is preserved in the Tavily server
• To use Tavily, set TAVILY_API_KEY and select the tavily_search agent config via Hydra (e.g., agent=tavily_search)

🤖 Generated with Claude Code

Automated Review

• Passed after 2 attempt(s)
• Final review: The Tavily parallel search migration is well-implemented and correct. All four previously-flagged issues have been addressed: scrape_website is properly ported to searching_tavily_mcp_server.py using JINA (feature parity with Google variant), JINA env vars are now actually consumed, the double-retry logic has been eliminated from tavily_mcp_server.py, and tenacity/tavily-python are now explicit dependencies in pyproject.toml. The additive strategy is intact — multi_agent.yaml still uses tool-google-search, while the new tavily_search.yaml provides a Tavily-only alternative. SDK patterns, imports, env var propagation, and MCP server architecture all mirror the existing Serper implementation consistently. One unused import (json) in searching_tavily_mcp_server.py is the only finding.