fix(runtimed-py): keep connect runtime alive — fixes kernel launch from Python bindings

[rgbkrk]

Fixes #992.

The bug

Every start_kernel() call from the Python bindings failed with RuntimedError: Disconnected from sync task. The daemon log showed Broken pipe (os error 32) during streaming load. The Tauri frontend also couldn't start kernels for saved notebooks.

Root cause

The sync Session methods (open_notebook_with_socket, create_notebook_with_socket, join_notebook_with_socket) all followed this pattern:

let runtime = Runtime::new()?;
let (notebook_id, state, _) = runtime.block_on(connect_open(...))?;
drop(runtime);  // ← KILLS THE SYNC TASK
Self::from_state(notebook_id, state, peer_label)  // creates a NEW runtime

connect_open spawns a sync task via tokio::spawn inside build_and_spawn. That task runs on runtime. When runtime is dropped, the sync task is cancelled. from_state creates a fresh runtime, but the sync task is gone — the DocHandle's command channels are dead.

Why reads worked: get_cells() reads directly from the local Automerge doc (no sync task needed).

Why writes failed: start_kernel() sends LaunchKernel through DocHandle::send_request() → cmd_tx.send() → the sync task's cmd_rx is dead → SyncError::Disconnected.

Why the daemon saw broken pipe: The sync task owned the socket reader. When the runtime dropped, the socket closed. The daemon's next send_typed_frame() got EPIPE.

Fix

Pass the connect runtime through to the Session instead of dropping it and creating a new one. New method from_state_with_runtime(runtime, ...) keeps the sync task alive.

Verification

client = runtimed.Client()
session = client.create_notebook(runtime='python')
session.start_kernel(kernel_type='python', env_source='uv:prewarmed')
result = session.run('print("Hello!")')
# success=True → "Hello!"

Also verified with open_notebook on a saved .ipynb file with env_source='auto'.