fix(windows): _process_alive misreports process state

[tradewithmeai]

Fixes #149.

What's wrong

_process_alive() uses os.kill(pid, 0), the POSIX existence idiom. On Windows os.kill is implemented over OpenProcess rather than signals, and the function returns the wrong answer in two different ways — measured on Windows 11 / CPython 3.13:

PID state	old behaviour	correct
never existed	raises OSError(WinError 87) — propagates out of a status check	False
recently exited	returns True	False
0 / negative	returns True	False

The middle row is the reason this is worth fixing. A Windows process handle stays openable while any handle to it remains, so os.kill(pid, 0) on a just-exited PID doesn't raise — cce services status reports a crashed service as healthy, with nothing to notice. The WinError 87 crash is the loud symptom; this is the quiet one.

Non-positive PIDs slipped through too: on POSIX os.kill(0, 0) signals the caller's whole process group and returns cleanly, so pid=0 reported alive.

The change

psutil>=5.9 is already a core dependency and already imported by config.py, so this uses it rather than hand-rolling a ctypes OpenProcess + WaitForSingleObject path:

if pid <= 0:
    return False
try:
    return psutil.Process(pid).status() != psutil.STATUS_ZOMBIE
except psutil.NoSuchProcess:
    return False
except psutil.AccessDenied:
    return True   # exists, owned by another user — preserves the old PermissionError branch

Zombies are excluded deliberately: on POSIX a reaped-but-not-collected child still "exists" while not running, which is the same wrong answer in a different shape.

Tests

Three added alongside the existing max-PID case:

• an actually-exited child reports dead — the case the current suite cannot reach. test_process_alive_dead_pid uses PID 2**22, beyond the max PID on Linux/macOS, so it only exercises "never existed"; the silent wrong answer lives in the exited-process path.
• a known-running child reports alive — positive control, so the fix can't pass by returning False for everything.
• 0 and negative PIDs report dead.

Verified by mutation rather than by watching them pass: against the previous implementation 5 tests in this file fail; against the fix all 23 pass.

CI

This also turns tests/test_cli_smoke.py::test_services green — it was failing for the same root cause.

One thing worth flagging separately: ci.yml runs the Windows job with || true, so both of these failures were being swallowed rather than surfaced. That mask was hiding a real defect, not runner flakiness. Removing it is out of scope here, but you may want to once Windows is green.