live_testlisted

Run a live end-to-end test of the **project-init scaffolder**. Deterministic pytest coverage (`tests/smoke/test_packaging.py`, the `wheel-smoke` CI job, ~70 contract tests) already proves files land with the right exec bits. This command tests what those can't: whether a **scaffolded project actually works in real use** — do its hooks fire, do lifecycle scripts run, does `just ci` pass, would an agent landing in the project find coherent instructions. It captures the manual 0.4.0 "external live test" (which found 13 real bugs, epics #440–444) as a repeatable procedure. It builds the wheel once, installs it into a throwaway venv, scaffolds a representative matrix of mini-projects, fans out **one isolated subagent per project** to exercise four depths, and aggregates a timestamped severity-ranked `REPORT.md`. **No GitHub side effects, no auto-issue-filing, no merge path.** ## Execution model (with fallback) The default is a **parallel one-subagent-per-project fan-out** (general-purpose or Explore subagent). If the subagent/Task tool is unavailable or refused, **fall back to a serial single-agent loop** that runs the same per-project steps and writes the same `result.json` files. The fan-out is an optimization, not a hard dependency. Claude (the orchestrator) runs Steps 1–3 centrally, fans out Step 4, then runs Steps 5–6 after the barrier. ## Steps Define the run root up front and print it before any fan-out: ```bash RUNID="${1:-$(date +%s 2>/dev/null || echo run)}" # a