Sanitize user-controlled inputs in XML-structured LLM prompts to prevent injection attacks

[chigwell]

User Story
As a security-conscious developer,
I want to sanitize user-controlled inputs in prompt formatting
so that malicious XML tags can't disrupt LLM response parsing.

Background
The current user_prompt.format() in eknowledge/main.py directly inserts raw text into XML-structured LLM prompts. This allows injection of fake <node> entries through inputs containing XML syntax (e.g., "<node><from_node>HACK</from_node>"). The vulnerability exists in:

# main.py line 92:
HumanMessage(content=user_prompt.format(text=chunk, relationships=relations))

Attackers could manipulate knowledge graph outputs by poisoning text inputs with XML tags, potentially creating虚假 relationships or disrupting parsing logic.

Acceptance Criteria

• Modify execute_graph_generation in eknowledge/main.py to sanitize text inputs
• Replace special XML characters (<, >, &) with entities (<, >, &) before string formatting
• Add test case in tests/test_eknowledge.py that verifies:
  • Inputs containing <node>TEST</node> get converted to <node>TEST</node> in prompts
  • LLM receives sanitized text that doesn't create unintended XML nodes
• Ensure verbose mode logs show original vs sanitized text when enabled
• Maintain existing chunk processing performance (add benchmark assertion if missing)