feat(mcp): wire per-agent MCP server loading in managed mode

Summary

mcp.Manager.LoadForAgent() exists but is never called. In managed mode, MCP servers are still loaded globally from config.json (gateway.go:200-209), not per-agent from DB.

Current state (infrastructure is ready)

• DB schema ready: mcp_servers, mcp_agent_grants (with config_overrides JSONB), mcp_user_grants
• Store methods ready: ListAccessible(), GrantToAgent(), etc.
• Manager method ready: LoadForAgent(ctx, agentID, userID)
• HTTP API ready: /v1/mcp/servers/{id}/grants/agent

What's missing

1. ListAccessible() doesn't SELECT config_overrides from mcp_agent_grants — the column exists but isn't included in the query result
2. LoadForAgent() doesn't merge config_overrides.env into server env before spawning
3. wireManagedExtras() / resolver doesn't call LoadForAgent() per agent — MCP is only initialized once globally
4. MCP tools register into global registry — need per-agent clone (similar to DynamicToolLoader.LoadForAgent() pattern)

Use case

Running 2+ agents in managed mode with different MCP server configs. Example: separate NeuralMemory brains per agent via env var overrides in config_overrides:

// mcp_agent_grants.config_overrides for agent "tho-ngoc"
{"env": {"NEURALMEMORY_DIR": "/home/user/.neuralmemory-tho-ngoc"}}

// mcp_agent_grants.config_overrides for agent "rua"
{"env": {"NEURALMEMORY_DIR": "/home/user/.neuralmemory-rua"}}

Both agents share the same mcp_servers record (command: nmem mcp) but get different env vars per grant → brain isolation without duplicate server definitions.

Proposed approach

1. Add ConfigOverrides []byte to MCPAccessInfo and include mag.config_overrides in ListAccessible() SQL
2. In LoadForAgent(), parse config_overrides.env and merge into server env map before connectServer()
3. In resolver's resolveAgent(), create per-agent MCP Manager with cloned registry (after DynamicToolLoader clone), call LoadForAgent()
4. Handle MCP process lifecycle cleanup when agent loop stops

Happy to submit a PR if this aligns with your design intent.

Start with gateway.go:200-209, then trace wireManagedExtras() and resolver.resolveAgent(), comparing their flow with DynamicToolLoader.LoadForAgent(). Review MCPAccessInfo, ListAccessible(), and LoadForAgent() for config_overrides merging and per-agent registry cloning. Done means managed agents load their granted MCP servers with isolated environment overrides and MCP processes are cleaned up when agent loops stop.