Expose terraphim_rlm tools via terraphim_mcp_server (agent training / nanocode parity)

[AlexMikhalev]

Goal

Make Terraphim easy to "teach" to coding agents (Claude Code / OpenCode) by providing a stable, self-documenting tool surface and deterministic guardrails.

This issue focuses on the RLM branch work needed to reach nanocode-like parity: expose terraphim_rlm MCP tools through terraphim_mcp_server, with clear session handling and docs.

Why

• crates/terraphim_rlm already implements an RLM query loop + executor abstraction + MCP tool service (rlm_code, rlm_bash, rlm_query, rlm_context, rlm_snapshot, rlm_status).
• But those tools are not currently visible/dispatchable from the main MCP server, which is what external agents integrate with.

Proposed changes

1) Wire RLM MCP tools into terraphim_mcp_server

• Add terraphim_rlm dependency to crates/terraphim_mcp_server.
• Instantiate TerraphimRlm + RlmMcpService during MCP server init.
• Extend MCP list_tools to include RlmMcpService::get_tools().
• Extend MCP call_tool to delegate rlm_* tool calls to RlmMcpService.

2) Decide/implement session UX

Pick one (recommend explicit session tools for deterministic external orchestration):

• Option A: Add rlm_session_create, rlm_session_set_current, rlm_session_destroy tools.
• Option B: Require session_id on all rlm_* calls and provide a default session if omitted.

3) Backend readiness / fallbacks

• Document requirements for Firecracker backend (Linux + /dev/kvm).
• Ensure a non-Firecracker fallback works (e.g., SSH/local) for dev environments.

4) Docs for agent integration (Claude Code / OpenCode)

Add a short doc describing:

• the tool list + expected JSON schemas
• recommended "roles" workflow (Retriever/Planner/Executor/Verifier)
• where to plug deterministic replacements (hooks) vs relying on model memory

Acceptance criteria

• tools/list from MCP includes all rlm_* tools.
• call_tool works for:
  • rlm_bash (simple command)
  • rlm_code (simple Python snippet)
  • rlm_query (iterative loop until FINAL)
• Session handling is documented and testable.
• Minimal E2E demo documented (MCP client → rlm_query to solve a small task).

Notes

If needed, scope to the feat/terraphim-rlm-experimental branch first, then merge back once stable.