feat: add hallucination grader and fix web search dead branch

[bhavyakeerthi3]

Summary

This PR introduces two significant improvements to the cross_database agent graph: a new Hallucination Grader for RAG quality control and a fix for the Web Search dead branch bug.

---

Changes

1. Hallucination Grader (New Feature)

While the existing pipeline had a completeness_grader to assess if answers were detailed, it lacked a check for factual grounding. This PR adds a "Hallucination Grader" to verify that generated answers are strictly supported by the retrieved context.

Implementation Details:

• New Task: Added src/agent/tasks/hallucination_grader.py which uses structured LLM output to return a binary Yes/No score and a reason for any detected hallucinations.
• State Updates: CrossDatabaseState now tracks retrieved search results (reactome_context and uniprot_context) to provide the grader with the necessary evidence.
• Graph Logic: Two new parallel nodes (check_reactome_hallucination and check_uniprot_hallucination) have been added to the graph. They run immediately after each database generates an answer and before the completeness check.

---

2. Fix: Web Search Dead Branch (Bug Fix)

Problem: When both Reactome and UniProt returned incomplete answers, the graph was short-circuiting directly to the final response, completely skipping the intended Tavily web search.

The Fix:

• Registered a functional perform_web_search graph node backed by TavilyWrapper.
• Fixed the conditional edge logic in decide_next_steps to route to the actual search node.
• Updated summarize_reactome_uniprot.py to include a {web_results} placeholder so the final response can incorporate and cite external sources when relevant.

---

Files Changed

File	Type	Description
src/agent/tasks/hallucination_grader.py	New	Hallucination grader task logic
src/agent/profiles/cross_database.py	Modified	Added state fields, hallucination nodes, and fixed search edges
src/agent/tasks/cross_database/summarize_reactome_uniprot.py	Modified	Updated prompt template to support citation of web results

Testing

Verified via static code review and graph wiring validation. Full end-to-end verification requires OpenAI and Tavily API keys.