fix(test): stop child-process cleanup from masking real test failures

[jamesaphoenix]

What broke

CI went red on main again after #79, on ralph-watchdog-script.test.ts > exits cleanly on SIGINT.

The reported error was Timed out waiting for process 16248 to exit — but that message comes from the cleanup helper, not the test body:

• SIGNAL_EXIT_TIMEOUT_MS (what the test itself waits) is 15s
• The test failed after 4065ms — which is exactly terminateChild's two 2s waits
• SIGTERM passed in 86ms in the same run, so process teardown itself was fine

Root cause

terminateChild runs from finally blocks, but its SIGKILL path awaited waitForExit without a try/catch:

proc.kill("SIGTERM")
try {
  await waitForExit(proc, 2000)
  return
} catch {
  proc.kill("SIGKILL")
  await waitForExit(proc, 2000)   // <- unguarded, throws out of finally
}

When the exit event was slow to arrive under CI load, cleanup threw and replaced whatever the test had actually asserted. We never got to see the real outcome.

Fix

Cleanup is best-effort and cannot throw. The SIGKILL wait also gets more room: SIGKILL cannot be caught, so a missing exit event inside the window is a reporting artifact rather than a process still running. The harness still reaps by pid afterwards via cleanupHarness.

Applied to both files carrying this helper (ralph-watchdog-script.test.ts, ralph-script.test.ts).

Note on the previous fix

#79 was still worth doing and is what made this diagnosable. Before it, vitest killed the test at its global 10s timeout with a generic "Test timed out in 10000ms" and no information. After it, the failure surfaced a specific message with a pid — which is what pointed at the cleanup helper rather than the watchdog script.