Feature request: `claude mcp restart` command for reconnecting hanging MCP servers

Problem

MCP servers frequently hang, disconnect, or become unresponsive during long Claude Code sessions. Common causes include:

• SSE connections dropping (e.g., Baserow MCP returning "Invalid request parameters" after being connected)
• npx-spawned processes stalling silently
• HTTP/SSE servers timing out without proper error propagation
• OAuth tokens expiring mid-session (e.g., Zoho MCP)

Currently, the only recovery options are:

1. VS Code: Manually find and restart the server in the MCP panel UI
2. CLI: claude mcp remove <name> + claude mcp add <name> (loses config, error-prone)
3. Nuclear: Reload entire VS Code window (Ctrl+Shift+P → Reload Window)

None of these are programmatic. When running automated workflows (e.g., a /start-session command that pings all MCPs), there's no way to recover from failures without human intervention in the UI.

Proposed Solution

Add a restart subcommand to claude mcp:

# Restart a specific server
claude mcp restart <name>

# Restart only servers that fail health check
claude mcp restart --failed

# Restart all servers
claude mcp restart --all

Behavior: Kill the existing MCP server process (or close the SSE/HTTP connection), then re-spawn/reconnect using the existing configuration. No config modification needed.

Health check integration: claude mcp list already performs health checks and reports ✓ Connected / ✗ Failed to connect. The restart --failed variant would use this same health check to identify which servers need restarting.

Real-World Example

This morning, starting a session with 21 configured MCP servers:

• claude mcp list showed 20/21 connected, 1 known-disabled
• However, 3 servers (Baserow, Supabase, Apify) returned MCP errors when actually queried via tools
• The health check in claude mcp list showed them as "Connected" but they were functionally broken
• Required manual reconnection in VS Code MCP panel for each one
• After reconnecting, all 3 recovered immediately

A claude mcp restart --failed (or even claude mcp restart baserow supabase apify) would have saved several minutes of manual troubleshooting.

Additional Context

• Platform: Windows 11, VS Code extension + CLI
• MCP server count: 21 configured (mix of stdio, SSE, HTTP transports)
• Frequency: Happens ~2-3 times per week across different servers
• Claude Code version: v2.1.74

This would also enable building self-healing automation — e.g., a startup command that detects failed MCPs and auto-restarts them before proceeding.