MCP chat proxy: expose full chat-with-memory as an MCP tool #39

Summary

Add a chat tool to the MCP server so any MCP-capable client can send a
message and receive a memory-aware response — without needing the CLI or
an HTTP connection.

Behaviour

1. Client calls chat tool with { "message": "...", "session_id": "..." }
2. Ragger searches memory for relevant context
3. Builds prompt: system prompt + injected memories + session history
4. Calls configured LLM (inference provider in settings.ini)
5. Captures both sides of the turn (same path as ragger chat)
6. Returns response text as tool result

Streaming caveat

MCP tool results are returned whole. No token-by-token streaming unless
MCP adds a streaming tool-result mechanism in a future spec revision —
worth revisiting when the spec evolves. For now, suitable for automation
and headless agents; not ideal for interactive use.

Scope

• New tool registered in the MCP server (src/mcp.cpp or equivalent)
• Reuses existing chat/inference logic from ragger chat
• Session history keyed by session_id (optional; defaults to a
  per-connection ephemeral session)
• Turn capture unconditional (same as Tier 2 / Tier 3 — see issue Design note: memory capture tiers and MCP chat proxy #38)

Related

• Issue Design note: memory capture tiers and MCP chat proxy #38 — design note covering capture tiers and MCP proxy feasibility
• Issue Labeled API tokens for non-local users #36 — labeled API tokens (auth for remote MCP callers)