Clear dead kernel client after execution failure

[PiotrCzapla]

Claude was still trying to use the old connection even when it saw the message. So it hung for 2 minutes and was really confused. It is likely due to the fact that we were investigating rustygate issues before, but still, the idea to close the connection and start a new kernel automatically seems like a good default.

Here is the PR made by codex, intended for your llm, I haven't read it just the code changes and they look fine.

Summary

Close and clear the kernel client after DeadKernelError. The next MCP execute can then create a fresh kernel instead of waiting on a dead WebSocket.

The user-facing death message now describes the recovery behavior, and the existing real lifecycle test checks that the stale client pointer is cleared.

Root cause

execute_outs returned a death note but left Client.kc set. The MCP frontend checks that pointer to decide whether it should auto-connect. After a kernel or RustyGate restart, the stale object prevented auto-connect and later executions could hang indefinitely.

Checks

• nbdev-test --path nbs/00_core.ipynb --n_workers 1 --cell_timeout 90
• nbdev-test --path nbs/01_mcp.ipynb --n_workers 1 --cell_timeout 90
• nbdev-export
• git diff --check