Node test stub flake: re-adding a SIGUSR2 listener on each call resets it to terminate

In Node, process.removeListener('SIGUSR2', h) followed by process.on('SIGUSR2', h) is not atomic. When the last listener is removed, Node restores the signal's default disposition, which for SIGUSR2 is terminate. A signal that lands between the remove and the add kills the process. A test stub whose begin() function did remove+add on every call produced a rare flake (about 1 in 36 runs, worse under --debug and memory pressure) that looked like a bug in the code under test: a worker vanished along with its queued work. The real native addon installed its handler once, so production never saw it. Fix: install the handler once at module load and keep the per-call state in a variable the handler reads. If stubs and the real implementation disagree on signal setup, suspect the stub first when a flake appears only in the stub suites.