Agent-agnostic worktree runtime migration

New dispatches use only these planes:

workers    ~/.vibecrafted/worktrees/<org>/<repo>/YYYY_MMDD/<cut-id>
artifacts  ~/.vibecrafted/artifacts/<org>/<repo>/YYYY_MMDD/{plans,reports,...}
runtime    ~/.vibecrafted/control_plane/...

Do not move a live legacy checkout behind its worker. Let it finish or recover it from its existing receipt, then start new cuts in canonical worktrees. Legacy .claude/worktrees, .codex, .gemini, and repo-local .vibecrafted locations are read-only recovery inputs and are never created by the new dispatcher.

Before enabling concurrency > 1:

1. Unset any fleet-wide CARGO_TARGET_DIR.
2. Run vibecrafted dispatch <plan> --doctor.
3. Add explicit depends_on edges and a named integrator = true join for non-linear branch tips.
4. Start the plan and inspect its runtime ledger with dispatch-doctor <plan> --run-id <run-id>.
5. After integration settles, run vibecrafted dispatch <plan> --cleanup-settled <run-id>.

The exact hard-failure remediation for an old shared target is:

unset CARGO_TARGET_DIR — Vibecrafted assigns $PWD/target per worker

Cleanup removes settled linked checkouts and their local targets. It retains branches and durable evidence. Runtime receipts remain on the control plane for reconciliation and retention policy; active runs are never cleaned.

The repository smoke command exercises two concurrent workers plus an exclusive join using real linked checkouts and local subprocesses:

python3 scripts/smoke-dispatch-worktrees.py

It prints the durable JSON receipt path.