Refine local agent system prompt for better tool usage and response quality #1040

The local agent system prompt in prompt_templates.rs is a placeholder. It needs refinement based on real-world usage patterns with the Ministral 3B/8B models.

Current State

• Basic system prompt with node reference URI format and display formatting instructions
• Model sometimes makes bad tool calls (nesting args under properties, echoing schema)
• Model sometimes retries tool calls unnecessarily instead of responding with results
• Tool definitions are duplicated between MCP handlers and local agent GraphToolExecutor

Areas to Refine

Prompt Quality

• Tool usage guidance: when to call tools vs answer directly
• Response formatting: how to present search results, node references
• Error handling: what to do when tool calls return empty results
• Conversation flow: when to ask clarifying questions vs make assumptions
• Token budget awareness: keep prompt concise for 3B model context limits

Tool Unification

• MCP handlers (nodespace-core) and local agent tools (GraphToolExecutor) have separate implementations of the same operations
• get_node markdown export now calls MCP handler directly — same pattern should apply to all tools
• Refactor shared tool logic into a common service layer that both MCP handlers and GraphToolExecutor use
• Ensures consistent behavior, descriptions, and parameter schemas across both surfaces

Acceptance Criteria

• System prompt refined based on observed model behavior
• Tool call accuracy improved (fewer bad/redundant calls)
• Response quality improved for common queries (search, create, update)
• Prompt fits within 4k token budget for system prompt + tool definitions
• Shared tool logic extracted — MCP handlers and GraphToolExecutor use the same underlying implementation
• Tool descriptions and parameter schemas are consistent between MCP and local agent