Make invoke timeouts terminate process trees

[oritwoen]

What happened?

Harness.invoke() and listModels() document timeoutMs as killing the harness after the deadline, but the shared runner sends one SIGTERM to the direct child and waits indefinitely for close. A child that ignores the signal keeps the promise pending. If the child exits but has spawned descendants, those descendants remain alive.

This means the timeout is neither bounded nor recursive.

How to reproduce

Using a custom harness backed by node:

1. Run process.on("SIGTERM", () => {}); setTimeout(() => process.exit(0), 500) with timeoutMs: 30. invoke() returned after 531 ms.
2. Run a child that spawns a 60 second grandchild, then waits. invoke() returned with timedOut: true, but the grandchild was still alive and needed an explicit SIGKILL.

Anything else?

Terminate the owned process tree, allow a documented grace period after SIGTERM, then escalate to a hard kill. Add regression tests for ignored termination signals and surviving descendants on each supported platform.

[meridianmindx]

Submitted complete fix with regression tests in PR #5:

• Spawns in isolated process group on POSIX (detached: true) and recursive process tree termination (taskkill /T /F) on Windows.
• Dispatches SIGTERM first with configurable grace period (killGracePeriodMs, default 500ms), escalating to SIGKILL to clean up ignored signals and surviving child/grandchild processes.
• All 54 unit tests passing cleanly.

Base EVM Payout Address: 0xcFDa9f32d292661740a6d0B4c00867E34c05c56D