The slash_worker creates a HermesCLI which imports model_tools, triggering
discover_mcp_tools() at module scope. Meanwhile, the TUI server also calls
MCP discovery independently. Both paths spawn ``hermes mcp serve`` child
processes per session.
Fix: defer the cli import in slash_worker to main() and set
HERMES_MCP_DISCOVERY=0 beforehand. model_tools now checks this env var
and skips MCP discovery when suppressed.