JPY_PARENT_PID is a PID, but ipykernel treats it as a HANDLE on Windows: parent-death kernel shutdown (#1559) is inactive on win32

[chimin-roh]

Summary

v0.8.0 ships "Kernels exit when their owner dies" (#1559). On Windows the mechanism does not arm, so kernels can outlive their owner exactly as before.

packages/coding-agent/src/core/kernel/index.ts:745 passes:

JPY_PARENT_PID: String(process.pid)

On POSIX, ipykernel's ParentPollerUnix treats this as a PID and the chain works. On Windows, kernelapp.init_poller() selects ParentPollerWindows, which passes the value to WaitForMultipleObjects as a process HANDLE, not a PID. jupyter_client does this correctly: it creates an inheritable handle with DuplicateHandle and passes that handle value (jupyter_client/launcher.py).

Evidence

• Measured on Windows 11: calling WaitForSingleObject with a live node.exe PID returns WAIT_FAILED (0xFFFFFFFF) with GetLastError = 6 (ERROR_INVALID_HANDLE).
• ipykernel logs Parent poll failed. If the frontend dies, the kernel may be left running. and the poller thread exits, so no parent watch is active.
• The feature's own integration test is skipped on Windows (test/kernel-parent-watchdog.test.ts:492), so CI does not cover this path.
• The PR description states the chain is closed "on every platform"; on win32 it is not.

Impact

The remaining safety net is the supervisor orphan journal (cli/owned-session-worker.ts:299-313), which only helps while the supervisor is alive. If the process tree is killed (for example taskkill /f /im node.exe, which Windows users do reach for), the kernel survives.

Suggested fix

On win32, duplicate the current process handle as inheritable and pass the handle value instead of the PID, mirroring jupyter_client.launcher. Keep the PID for POSIX.

Happy to test a patch on Windows 11 (ACP 949) if that is useful.

[snimu]

Addressed by #1685