#1648 test(acp): keep the delayed fake agent alive until its exit delay is up

The acp test "retries the session directory removal after the agent exits when the first try fails" has failed 12 Linux CI jobs since #1626 (Node 22, 22.22.3, and 24): kill ESRCH at tests/executor.test.ts:522, because the fake agent, which should take 1.5 s to stop after SIGTERM, was gone within 500 ms. It blocks unrelated PRs such as #1647.

• Teardown (packages/acp/src/session.ts) closes the agent's stdin and stdout pipe right before the SIGTERM. A process whose only other hold is a signal listener can exit on stdin's EOF before the signal handler runs, and a late write can hit EPIPE.
• The fixture's delayed agent (ACP_EXIT_DELAY_MS) now holds the event loop until its delay is up. Once it is stopping, it logs a stdout error, an uncaught error, or a rejection instead of dying on it; before that, an error still ends it at once with code 1, so a real failure fails the test fast.
• The delay starts at SIGTERM or stdin's end, whichever comes first, so an agent whose test worker died without a teardown still exits.
• Both tests that use the delay print the agent's last log lines when it is gone early, so a future failure says why.
• Test code only; no changeset.

Verified

Ran it locally: yes

• The failure never reproduced: 6 runs on macOS and 37 in Linux containers (node:22.22.3, node:24, non-root, 1 and 2 CPUs), so the CI logs above are the evidence for the race.
• With the change: pnpm --filter @e2e-dev/acp executor tests pass on macOS, and the full acp suite passed 15 times in a node:22.22.3 container as a non-root user (the test skips as root).
• CI on this PR: the test jobs on all five Node versions passed, 22.22.3 included, where this test failed most. One check run hit an unrelated web-engine timeout (locator-frame.e2e.ts) and passed on rerun.
• A fresh-context review found the orphan case (a held agent whose worker died would never exit); fixed by also starting the delay at stdin's end.

Tracked in ENG-1185.