Auto-reconnect MCP streamable-HTTP servers on `Session not found`

Summary

When an MCP server using streamable HTTP transport evicts a session
(idle timeout, restart, deploy), the client keeps using its stale
session ID and every tool call fails with:

Streamable HTTP error: Error POSTing to endpoint:
{"error":{"message":"Session not found","code":-32600},...}

/mcp reports the server as "Connected" because that status reflects
the last successful handshake, not live session validity. Users must
manually reconnect via /mcp to recover.

Proposed behavior

On receiving JSON-RPC error code -32600 with message "Session not found"
(or equivalent session-eviction signal), Claude Code should transparently
re-establish the session and retry the failed call once, rather than
surfacing the error to the user.

Repro

1. Connect to any streamable-HTTP MCP server (e.g. Fastmail).
2. Leave the session idle long enough for the server to evict it

(or wait for a server-side restart).

1. Invoke any tool from that server.
2. Observe Session not found error; /mcp still shows "Connected".
3. Manually /mcp → reconnect → tool works again.

Environment

Claude Code (CLI), darwin.