feat(mcp): expose AutoBot memory/KB/graph as an MCP server for external clients #5072

Problem

AutoBot is a consumer of MCP servers (see services/mcp_dispatch.py, services/mcp_isolated_runtime.py) but does not expose its own memory as an MCP server. External Claude Code / Gemini / Cline / other MCP-capable clients cannot query AutoBot's knowledge base or memory graph.

This is a missed integration opportunity: AutoBot has a mature hybrid search, entity-relationship graph, and chat history — all valuable context for an external LLM IDE.

Proposed change

New module: `autobot-backend/mcp/autobot_server.py` — an MCP server that exposes:

Knowledge base tools:

• `kb.search(query, filters?, limit?)` — hybrid search over KB
• `kb.get_document(doc_id)` — fetch full doc
• `kb.list_categories()`, `kb.list_tags()`

Memory graph tools:

• `memory.entity_lookup(name)` — entities + relations
• `memory.timeline(entity, range?)` — temporal walk
• `memory.related(entity, depth=2)` — neighbor traversal

Chat history / verbatim tools (blocked on verbatim-lane issue):

• `chat.session_list(user)`
• `chat.verbatim_search(query)`

Agent introspection:

• `agents.list()` with capabilities + descriptions
• `agents.invoke(id, prompt)` (auth-gated)

Acceptance criteria

• MCP server binary (`autobot-mcp-server`) runnable via stdio and over HTTP
• Auth: token-scoped access per tool category
• All tools documented with JSON schema
• Installation guide for Claude Code / Gemini CLI / Cline
• E2E test: Claude Code connects and queries the KB
• Rate limit + observability

Design notes

• Reuse existing FastAPI endpoints where possible — the MCP server should be a thin transport, not a reimplementation
• Respect existing RBAC (auth_rbac.py)