Fix #29726 — MCP server startup hang fix

[zombi3butt]

Fix #29726 — One unhealthy optional MCP integration should not make the whole Hermes session unavailable

Problem

When an MCP server is configured under mcp_servers, Hermes attempts to connect and discover tools during startup. If that MCP server is down, slow, misconfigured, or fails during initialize / list_tools, Hermes startup could hang indefinitely before the normal chat session becomes usable.

Root Cause

The MCPServerTask.start() method called await self._ready.wait() with no timeout wrapper. If the server's internal run() loop never fires _ready.set() (e.g., due to a silently swallowed exception in a nested transport handler), start() blocks forever — and since MCP discovery runs serially per-server, one dead server freezes the entire startup.

Fix

1. MCPServerTask.start() — wrapped await self._ready.wait() in asyncio.wait_for(timeout=connect_timeout). On timeout: cancels the orphaned run-task to prevent resource leaks, then raises TimeoutError.
2. hermes_cli/main.py — changed MCP startup error log from logger.debug → logger.warning so users see a clear "MCP tool discovery failed" message instead of silently losing tools.

Behavior Change

• Before: A hung/stuck MCP server could block Hermes startup indefinitely.
• After: Per-server connection timeout (default 60s, configurable via connect_timeout) limits wait time. Failed servers log a visible warning. The rest of Hermes starts normally with available servers only.

Notes

• No breaking API changes. All existing MCP configs work identically.
• discover_mcp_tools() already used asyncio.gather(return_exceptions=True) and logged warnings for per-server failures — this just adds the missing timeout guard at the lowest level (start()).

[teknium1]

Automated hermes-sweeper review: this startup-hang fix is now covered on main by the later MCP startup refactor.

Evidence:

• tools/mcp_tool.py:4167 now applies each server's connect_timeout with asyncio.wait_for(_connect_server(...), timeout=connect_timeout).
• tools/mcp_tool.py:4244 uses asyncio.gather(..., return_exceptions=True) and tools/mcp_tool.py:4256 records/logs failed servers without failing the whole discovery pass.
• hermes_cli/main.py:12065 backgrounds MCP discovery for normal chat/rl startup instead of running it inline.
• hermes_cli/mcp_startup.py:87 bounds the first-agent wait with thread.join(timeout=_resolve_discovery_timeout(...)), so a dead MCP server cannot freeze startup indefinitely.
• Regression coverage exists in tests/hermes_cli/test_mcp_startup.py:48 for the non-blocking startup path.

The implementation landed via d682f320b35a13084371a541a835e1d988c982b8; the PR's exact inner cleanup shape is not copied verbatim, but the user-facing guarantee from #29726 is implemented on current main.