# R-649 — a failed install removes what it started (controller v0.266.0), 2026-09-23

**Operator ruling 2026-09-23:** "On failed install, the controller should clean up the containers."
**Method: endpoint-level** on scratch guest 9202 (deploy through `POST /api/stacks/<n>/deploy`, state from
`GET /api/stacks/<n>`), drill catalog: `outline` with its redis healthcheck set to `false` in the DRILL
template only, so `compose up -d` fails after postgres and redis have started.

| proof | result | evidence |
|---|---|---|
| failed install | `deploy failed after 102.7s: exit code 1` → **`the failed deploy's containers were removed (volumes kept) — R-649`**; 0 containers of project `outline`; volumes `outline_outline_data`, `_postgres_data`, `_redis_data` kept; `deployed: false` | `20-*`, `21-*` |
| the household's Remove afterwards | 200, volumes removed with the data | `30-*` |
| floor | 0.266.0 / MinAgent 0.131.0 read back; demo-hp and demo-felhom on 0.266.0 in 20 s | `40-*` |

Unit: `TestR649_AFailedInstallRemovesWhatItStarted` (stub compose; PATH = the stub only, R-650) and its
control `TestR649_ASuccessfulInstallIsNotTakenDown`. Red-proof: the `down` removed → `calls=["up -d"]`.

**Teardown:** outline removed through the product; `outlinewiki/outline:1.9.1` and controller 0.265.0 removed
by name (postgres/redis images kept — in use by standing apps); `controller.yaml` identical to
`.pre-r649`; live catalog `cfcfe52`; drill repo reset to live `main`. Host: nothing. Hub: floor only.
9201 not touched (reached 0.266.0 by the floor).