fix(mcp): MCP status shows connected green after server reconnect even though server is down #24855

Bug Description

After a network disruption or server restart causes OpenCode to reconnect, the MCP status indicator continues to show all servers as green (connected), but the underlying MCP server processes may have been lost. The status is stale because Instance.state() in packages/opencode/src/mcp/index.ts caches MCP client state indefinitely per project instance keyed by directory. When Instance.reload() is triggered by reconnect logic, the old cached status and clients maps are preserved even though the server-side processes are gone, resulting in a false-positive 'all online' display.

Steps to Reproduce

1. Start OpenCode with one or more MCP servers configured and connected.
2. Observe the MCP status indicator in the bottom status bar shows green.
3. Cause a reconnect (e.g. network disruption, or restart the OpenCode backend server).
4. After reconnect completes, the status bar still shows all MCP servers green.
5. Try to invoke an MCP tool — it may fail because the actual stdio/HTTP transport is no longer alive.

Expected Behavior

When the server reconnects, the MCP cached state should be invalidated so the next status() call re-evaluates actual server connectivity.

Actual Behavior

The cached Instance.state() for MCP survives across server reconnects, displaying stale connected statuses.

Additional Context

Instance.state() uses a ScopedCache keyed by directory without an explicit expiration or invalidation hook on reconnect events. The MCP init callback only runs once per instance, and Instance.reload() disposes the old instance but the new instance may re-use the same directory key before the old cache entry is fully cleared, depending on timing.