hack/e2e-web-local.sh ended with `exec pnpm exec playwright test`, which replaces the shell and discards the EXIT trap — so every *green* run left the harness, its dev-env node and the forked server node running, holding ports and /tmp data dirs that look like the run you are debugging. Start the harness in its own process group (setsid) and kill the group on exit, then sweep any binary from this build dir that appeared during the run: the ingest worker re-setsid's itself and escapes the group kill. PIDs that already existed before the run are left alone, so a scratch node from `make dev` is not collateral damage. Verified: two consecutive green runs (4 passed, 36.5s), no live libstreamplace or dev-env process afterwards (only unreapable zombies under the container's `tail -f /dev/null` PID 1).