Summary

• Prompt injection mitigation: Added _sanitize_query() (500-char limit, control char stripping) and _build_decomposition_prompt() with structured delimiters separating system instructions from user input
• LLM error handling: Wrapped await self.llm(prompt) in try/except with single-step fallback plan on failure
• Per-step retrieval error handling: Wrapped mesh_retriever.retrieve() in try/except per step — failures log and continue with empty evidence

Closes #2169

Test plan

• Verify query decomposition works with normal queries
• Test with oversized query (>500 chars) — should be truncated
• Test with control characters in query — should be stripped
• Verify LLM failure returns single-step fallback
• Verify individual step failure doesn't kill the chain