Problem

The TUI's summoning hermes… phase blocks on gateway.ready, which ran MCP tool discovery inline in tui_gateway/entry.py:main(). Any configured-but-unreachable MCP server burns its full connect-retry backoff (1 + 2 + 4s ≈ 7s) before the composer appears.

Repro: put a down stdio/http server in mcp_servers (here: twozero_td on a dead port), launch hermes --tui. Measured spawn → gateway.ready:

~7,500 ms   (before)
~115 ms     (after)   ← ~65× faster

Fix

1. Background MCP discovery (the 7s). Move discover_mcp_tools() into a daemon thread so gateway.ready fires immediately. Discovery is already idempotent and _lock-guarded and registers tools into the shared registry as servers connect; a down server keeps retrying in the background instead of stalling the shell. The MCP-SDK cold-start import guard (skip the ~200ms SDK import when no mcp_servers configured) is preserved, so the no-MCP common case pays nothing.

Backgrounding discovery means it no longer completes before the first agent build. Since AIAgent snapshots its tool list once at build (agent/agent_init.py) and never re-reads the registry, two follow-ons keep that snapshot correct:

1a. Bounded wait before the first agent build. entry.py publishes the discovery thread handle; _make_agent calls wait_for_mcp_discovery(timeout=0.75) before constructing the agent. This lets already-spawning fast servers land in the snapshot while a slow/dead server is never waited on past the bound — so a dead server cannot re-introduce the startup hang. The wait is only ever paid on the first prompt during a still-in-flight discovery; it's a no-op once discovery has finished, and a no-op (≈1µs) for users with no mcp_servers (no thread is ever created). gateway.ready itself is untouched — the full ~65× speedup stands.

1b. /reload-mcp now actually refreshes the cached agent. The TUI /reload-mcp handler previously guarded on hasattr(agent, "refresh_tools"), but no such method exists — so it was dead code and the only recovery for a late-connecting server was /new (which discards history). Replaced it with a real agent.tools / valid_tool_names rebuild via get_tool_definitions(...), mirroring gateway/run.py::_execute_mcp_reload. A server that connects later in the session is now picked up by /reload-mcp without losing conversation history. (The user has already consented to the prompt-cache invalidation via the existing confirm gate.)

The background-discovery failure path also now logs (logger.warning(..., exc_info=True)) instead of swallowing silently — a detached-thread exception was previously invisible.

2. Lazy banner imports (~45ms bonus). tui_gateway.server imports hermes_cli.banner purely to reach the lightweight prefetch_update_check helper, but banner eagerly imported rich.console + prompt_toolkit at module level — ~45ms of wasted imports on the critical path. Made them lazy (imported inside cprint / build_welcome_banner); Console annotation moved under TYPE_CHECKING. import tui_gateway.server: ~115ms → ~69ms.

Verification

• spawn → gateway.ready: ~7,500ms → ~115ms (with a dead MCP server configured) — unchanged by the bounded-wait/reload follow-ons, which live on the first-prompt build path, not the gateway.ready path
• First-prompt agent build: no-op (~1µs) for no-MCP users; waits only until fast servers land for reachable MCP; bounded at 0.75s for a slow/dead server
• import tui_gateway.server: ~115ms → ~69ms
• Importing hermes_cli.banner no longer pulls rich/prompt_toolkit; cprint + build_welcome_banner still render correctly (lazy imports resolve at call time)
• wait_for_mcp_discovery verified: no-op with no thread, joins a fast/finished thread immediately, bounded (≈0.3s with a 0.3s timeout, not forever) on a hung thread — new tests in tests/tui_gateway/test_wait_for_mcp_discovery.py
• /reload-mcp rebuilds the cached agent's tool snapshot (no /new required for late-connecting servers)
• Tests: 330 pass — tests/tui_gateway/ (incl. the 4 new wait_for_mcp_discovery tests + existing test_make_agent_provider, test_entry_sys_path), tests/tools/test_mcp_tool.py + test_mcp_dynamic_discovery.py, tests/hermes_cli/test_banner.py, test_update_check.py, test_cmd_update.py. Ruff clean; no new ty diagnostics.

Net diff: 4 files, +192/-28.