tests: 🐛 arm the fake CLI's TERM trap before it publishes its pid

[yasyf]

cancelling_run_reaps_the_child_with_sigterm failed intermittently in CI, timing out after 10s waiting for the child to record that it had been signalled.

The race is in the fixture, not in the reaping under test. The fake claude CLI ran these in this order:

if [ -n "$SPAWNLLM_FAKE_PID_OUT" ]; then printf '%s' "$$" > "$SPAWNLLM_FAKE_PID_OUT"; fi
if [ -n "$SPAWNLLM_FAKE_TERM_OUT" ]; then trap 'printf term > "$SPAWNLLM_FAKE_TERM_OUT"; exit 0' TERM; fi

recorded_pid returns the instant the pid file parses, and the test calls task.abort() immediately after — so those two lines bracket exactly the window in which the TERM is delivered. A signal landing there takes the shell's default TERM action: the child dies without writing the term file, process_exists goes false, and the wait loop spins until its 10s timeout.

Arming the trap first closes it. Nothing observes the pid before the trap needs to exist, so the reorder is the whole fix.

Reproduced by widening the gap with a sleep between the two lines:

order	gap	term file
pid, then trap	200ms	(empty — the failure)
pid, then trap	0	term
trap, then pid	200ms	term
trap, then pid	0	term

cargo test -p spawnllm --test integration: 29 passed, 0 failed.

https://claude.ai/code/session_015yr73szXYEmbjtNVLxnetK