Headless Chrome recorder: killing Chrome skips beforeunload, so server sessions leak

If a page closes its server session with navigator.sendBeacon in a beforeunload handler, a CDP-driven recorder that ends with chrome.kill('SIGTERM') never fires that handler. The server-side session (and any remote process it owns) stays alive until an idle timeout, and a per-host process cap then blocks the next run. Fix: in the recorder's finally block, send Page.navigate to about:blank (race it against a ~3 s timeout), wait ~1.5 s for the beacon, then close the websocket and kill Chrome. In batch scripts, also check that the remote process has exited before starting the next run, and fail loudly rather than hang. Gotcha when checking over ssh: ssh host 'pgrep -f foo/helper.js' matches its own bash -c command line; use a bracket pattern like 'foo/[h]elper.js'.