Solid PR — the approach (process groups on Unix, taskkill /T on Windows) is the canonical solution for orphaned child processes.

What I liked:

• Setpgid: true + syscall.Kill(-pid, SIGKILL) is the textbook Go approach for process group cleanup ✓
• Build tags (!windows / windows) for platform separation are idiomatic ✓
• The regression test spawns sleep 60 and verifies it gets killed — well-structured ✓
• Already-exited process handling is correct (nil checks, discarded ESRCH errors) ✓
• Race between normal completion and timeout is handled correctly by the select ✓
• errors.Is improvement over == for context.DeadlineExceeded ✓

Suggestions (non-blocking):

1. Double-kill interaction: The command is created with exec.CommandContext, so Go's internal goroutine will also try to kill the direct child when the context expires. This is benign (killing an already-killed process is harmless), but slightly confusing. On Go 1.20+, you could use cmd.Cancel to hook terminateProcessTree directly into the CommandContext machinery, eliminating the hybrid approach:

   cmd.Cancel = func() error {
       return terminateProcessTree(cmd)
   }
   cmd.WaitDelay = 2 * time.Second

2. terminateProcessTree always returns nil — either return meaningful errors or change to no return value, since callers always discard it with _ =.

3. Consider adding a normal-completion test alongside the timeout test, to verify the Start + Wait refactor doesn't break the happy path.

4. SIGTERM before SIGKILL: Currently goes straight to SIGKILL. For an exec tool where the command has already timed out this is defensible, but a brief SIGTERM grace period (e.g., 500ms) before SIGKILL would allow children to flush buffers and clean up temp files.

Approving — these are all improvements, not blockers.