Two bugs found by trying to run the lab where cron actually runs it, in the sparse automation checkout. `process.exit(0)` in the CLI's `finally` ran before a rejection could be reported, so a lab that never got the page open printed nothing and exited 0 -- a broken run and a clean run produced identical output. It now catches, prints the stack, and exits non-zero, because a sweep's silence has to mean "no summits", not "no simulation". The main-module guard compared `resolve(argv[1])` against `fileURLToPath(import.meta.url)`. `import.meta.url` reports the REAL path while argv[1] keeps whatever the caller typed, so any symlinked component -- /tmp against /private/tmp on a Mac -- silently decided "imported" and the command line did nothing at all. Uses `import.meta.main` where node offers it, real paths otherwise. What the lab was actually failing on: the automation checkout's sparse cone carried the marketing tools but none of the modules `shell.mjs` serves out of system/public, so the page 404'd and never booted. Those four paths are now in the cone. `oskiewar-clockwork tune` runs the proposer, and only the proposer: it has no ledger to publish and no figures to post, so it stops before both. Claude-Session: https://claude.ai/code/session_01X62vT9M1iQXjNTF1jMcpN5