MCP server headroom connection timeout blocks first model call on every startup (30s delay)

Describe the bug

On every Jcode startup, the first model call is delayed by ~30 seconds because the MCP server headroom connection times out. The login message "MCP: advertised 14 cached tool(s) from 1 server(s) at spawn" appears, followed 30 seconds later by "Failed to connect to MCP server headroom: Request timeout: deadline has elapsed".

Steps to reproduce

1. Start Jcode
2. Observe that the first LLM call hangs for ~30 seconds
3. Check logs: the MCP server headroom connection consistently times out

Expected behavior

Jcode should gracefully handle MCP server connection failures without blocking the first model call.

Actual behavior

Every startup shows this pattern:

[INFO] MCP: advertised 14 cached tool(s) from 1 server(s) at spawn
... 30 seconds later ...
[ERROR] Failed to connect to MCP server headroom: Request timeout

Occurrences today (7/7 restarts)

All 7 Jcode restarts today exhibited this 30s delay (timestamps from logs).

Environment

• Jcode version: v0.60.0 (4f147a0c)
• OS: Linux (Red Hat KVM, x86_64)
• Provider: OpenRouter
• MCP servers: headroom only

Suggested fix

1. Non-blocking MCP init: first model call should not be blocked by MCP init
2. Shorter default timeout: consider 5-10s with background retry
3. Graceful degradation: log warning, don\t block user interaction

No files or tests are named. Start by tracing MCP initialization and the first model-call path, then reproduce startup with the headroom server while checking the logged timeout. Done means an MCP connection failure no longer delays the first model call, while the failure is logged and can be retried in the background.