fix: recreate a missing managed runtime from the recorded bundle in setup

[rogerdigital]

The bug

Removing the managed container while state.json and the content-addressed bundle survive leaves setup unrecoverable:

• setup resolves the deployment as matching (state fields agree, ownership is 'owned' through the surviving volume), readiness times out, and the recovery branch calls docker restart — which refuses when the container is absent: E_STATE_INVALID: Managed Docker resources are incomplete.
• repair refuses for lack of a profile attachment (remove detached it).
• remove --service cannot resolve a compose path (it reads com.docker.compose.project.config_files off container labels that no longer exist).

This is exactly the state remove --service without --purge-data leaves (compose down, volume kept), observed live and recorded as finding 6 of the 2026-09-19 baseline window-2 report.

The fix

Setup's matching-unhealthy recovery branch now mirrors the semantics repair's recreate-runtime action already has:

• container present but wedged → docker restart (unchanged);
• container absent → recreate the runtime with docker up against the bundle named by the recorded configurationSha256 (bundleComposePath);
• no recorded digest → refuse with E_STATE_INVALID instead of guessing a compose path.

The subsequent flow (readiness → real search → attach or validate → commit) is unchanged; with no attachment the same recovery runs first, so the original detach-then-cleanup scenario is covered by the same code path.

Deliberately out of scope, recorded as a follow-up: remove --service --purge-data still needs a present container to resolve its compose path, so a full purge after a container-less leftover requires one setup first. Changing what ownership verification means for a destructive command deserves its own discussion.

Verification

• Two new unit tests in test/cli/setup.test.ts (recreate-from-bundle with the recorded digest, refuse without one); harness deploymentStatus became configurable. All 75 setup tests green; pnpm verify green.
• Live reproduction on a real managed instance: docker compose down (volume kept) into the exact state that previously failed with Managed Docker resources are incomplete, then the fixed setup recreates the runtime and status --json passes 9/9 checks.