fix: remove the bundled registry from every host, and stop starting it

netOrk's docker-compose.yml shipped a registry:2 for the satellite image,
and this script started it on every deploy as infrastructure. Satellites
pull from registry.netork.io; on 172.22.8.50 the registry held one old
image, nothing had pulled from it in 30 days, and it accepted anonymous
pushes on port 5000 of every instance (NetOrk/netork#763).

- registry leaves INFRA_SERVICES.
- A new step removes services netOrk no longer ships: the container
  netork-registry-1, then the volume netork_registry_data. `docker
  compose up` never removes a container whose service left the compose
  file, so without this each host would keep it until someone removed it
  by hand. The step is idempotent and works with the old compose file as
  well as the new one, so it can go out before netOrk drops the service.

README.md

@@ -41,6 +41,7 @@ elsewhere.
 | `NETORK_VERSION` | Tag to deploy (default `latest`) |
 | `REGISTRY_USER`, `REGISTRY_PASSWORD` | Registry login used on every server |
 | `REGISTRY_USER_<server>`, `REGISTRY_PASSWORD_<server>`, `NETORK_VERSION_<server>` | Per-server overrides. `<server>` has its dots replaced by underscores, e.g. `_10_0_0_2` |
+| `DEPLOY_LOCK_WAIT` | Seconds to wait for another deploy to the same host (default 900) |
 
 ## Usage
 
@@ -65,14 +66,25 @@ started earlier pulls whatever image the registry held before, which is stale.
 
 ## What a deploy does, per server
 
+0. Takes the host's deploy lock, `flock` on `~/netork/.deploy.lock`, and holds it until
+   the deploy ends. A second deploy to the same host waits for it and says who holds
+   it, since when, and which tag they are deploying. It gives up after 15 minutes
+   (`DEPLOY_LOCK_WAIT`, in seconds) without touching the host. The lock belongs to an
+   ssh session, so a deploy that fails, is interrupted or loses its connection frees it
+   on its own. A host without `flock` is deployed without the lock, with a warning.
 1. Logs in to the registry. The password travels over ssh's stdin, never on a command
    line.
 2. Pulls the engine image and copies `docker-compose.yml` and
    `docker-compose.registry.yml` out of it into `~/netork/`.
 3. Pulls the netOrk images, then force-recreates the API, the workers, `netork-beat`,
-   `flower` and, on UI servers, `netork-ui`.
-4. Reconciles `registry`, `apt-cacher-ng` and `signal-api`: each is recreated only if its
+   `flower` and, on UI servers, `netork-ui`. If the recreate fails, it is tried once
+   more after 5 s (`DEPLOY_RECREATE_RETRY_DELAY`). Compose's parallel recreate can lose
+   a container it just renamed and leave the rest stopped. If the second attempt fails
+   too, the container states are printed and the deploy fails.
+4. Reconciles `apt-cacher-ng` and `signal-api`: each is recreated only if its
    definition changed. It never touches `postgres` or `redis`.
+   It then removes services netOrk no longer ships, container and volume, where a host
+   still has them: today the bundled `registry` (NetOrk/netork#763).
 5. Verifies that the containers run exactly the image that was pulled. If they don't, the
    deploy fails.
 6. Records `REGISTRY_HOST` and `NETORK_VERSION` in `~/netork/.env`, so that a

deploy.sh

@@ -34,6 +34,9 @@ NETORK_VERSION="${NETORK_VERSION:-latest}"
 # The compose files the engine image carries under /app/deploy/.
 COMPOSE_FILES="docker-compose.yml docker-compose.registry.yml"
 
+# How long a deploy waits for another one on the same host before giving up.
+DEPLOY_LOCK_WAIT="${DEPLOY_LOCK_WAIT:-900}"
+
 read -ra SERVERS <<< "$DEPLOY_SERVERS"
 
 # Run one step on a server, and let its failure stop the deploy.
@@ -70,6 +73,63 @@ run_remote() {
   fi
 }
 
+# ── One deploy per host at a time ────────────────────────────────────────────
+# Several sessions deploy to the same test server, and two runs used to overlap:
+# on 2026-10-05 two `up -d --force-recreate` recreated each other's containers,
+# on 2026-10-06 two runs renamed each other's *.new compose files, and with two
+# tags one tag's compose files could start the other's images (NetOrk/deploy#1,
+# #3). So each deploy first takes flock on ~/netork/.deploy.lock on the host.
+#
+# An ssh session holds it: the remote side opens the lock file on fd 9, takes
+# the lock, says LOCKED, and then waits on its stdin. Closing that stdin ends
+# the session and frees the lock, and so does anything else that ends it: the
+# deploy failing, being interrupted, or losing the network. No stale lock file
+# can outlive the deploy that took it.
+#
+# A host without flock (util-linux) is deployed without the lock, with a
+# warning, rather than becoming undeployable.
+#
+#   hold_host_lock <server> <who and what, for the next deploy to read>
+hold_host_lock() {
+  local server=$1 info=$2 line
+  # shellcheck disable=SC2029  # the wait and the info are meant to expand here
+  coproc HOST_LOCK {
+    ssh "$server" "mkdir -p ~/netork && cd ~/netork && exec 9>.deploy.lock && \
+      if ! command -v flock >/dev/null 2>&1; then echo NOLOCK; exec cat >/dev/null; fi; \
+      if ! flock -n 9; then \
+        echo \"BUSY \$(cat .deploy.lock.info 2>/dev/null)\"; \
+        flock -w ${DEPLOY_LOCK_WAIT} 9 || { echo TIMEOUT; exit 1; }; \
+      fi; \
+      echo '${info}' > .deploy.lock.info; echo LOCKED; exec cat >/dev/null" 2>&1
+  }
+  while IFS= read -r -t "$(( DEPLOY_LOCK_WAIT + 60 ))" line <&"${HOST_LOCK[0]}"; do
+    case "$line" in
+      LOCKED) return 0 ;;
+      NOLOCK)
+        echo "[${server}] WARNING: no flock on ${server}; deploying without the host lock." >&2
+        return 0
+        ;;
+      BUSY*)
+        echo "[${server}] another deploy holds ${server}: ${line#BUSY } — waiting up to ${DEPLOY_LOCK_WAIT}s..." >&2
+        ;;
+      TIMEOUT)
+        echo "[${server}] the deploy lock on ${server} is still held after ${DEPLOY_LOCK_WAIT}s; giving up." >&2
+        return 1
+        ;;
+      *) echo "[${server}] ${line}" >&2 ;;
+    esac
+  done
+  echo "[${server}] could not take the deploy lock on ${server} (the ssh session ended)." >&2
+  return 1
+}
+
+# Free the lock: closing the session's stdin ends the remote side.
+release_host_lock() {
+  local fd="${HOST_LOCK[1]}" pid="${HOST_LOCK_PID}"
+  exec {fd}>&-
+  wait "$pid" 2>/dev/null || true
+}
+
 VERSION_OVERRIDE=""
 EXPLICIT_SERVERS=()
 i=1
@@ -128,7 +188,20 @@ if [[ ${#SERVERS[@]} -eq 0 ]]; then
   exit 1
 fi
 
+# One server: under its host lock, the steps below.
+#
+# A step that fails stops the subshell (set -e) before release_host_lock runs.
+# That is fine: the subshell's end closes the lock session's stdin all the same.
 deploy_server() {
+  local SERVER="$1" KEY="${1//./_}" VER_VAR
+  VER_VAR="NETORK_VERSION_${KEY}"
+  hold_host_lock "$SERVER" \
+    "since $(date -u +%Y-%m-%dT%H:%M:%SZ), by ${USER:-?}@$(hostname -s), deploying ${VERSION_OVERRIDE:-${!VER_VAR:-${NETORK_VERSION}}}"
+  deploy_steps "$SERVER"
+  release_host_lock
+}
+
+deploy_steps() {
   local SERVER="$1"
 
   # Per-server lookup — falls back to global values
@@ -179,7 +252,7 @@ deploy_server() {
   # Containers on upstream images. Reconciled, not force-recreated (below).
   # postgres and redis are deliberately absent: restarting a database on every
   # deploy would be worse than any compose change one could miss.
-  local INFRA_SERVICES="registry apt-cacher-ng signal-api"
+  local INFRA_SERVICES="apt-cacher-ng signal-api"
 
   echo "[${SERVER}] Pulling images..."
   run_remote "$SERVER" "pulling images for ${SERVER_VERSION}" 'Pulled|Already|Error' \
@@ -194,12 +267,31 @@ deploy_server() {
   # observed on netork-ui, which kept serving a stale image after its tag had
   # already advanced. Recreating unconditionally costs a restart per deploy,
   # which a deliberate rollout wants anyway. See NetOrk/netork#90.
+  #
+  # Tried twice. compose recreates the services in parallel, and on 2026-10-05 it
+  # lost a container it had just renamed ("No such container") and stopped with
+  # every engine container Created and none running: the old API was gone, so
+  # netOrk was down until someone ran the deploy again, which then went through
+  # (NetOrk/netork#586). The second attempt is that rerun. Should it fail as well,
+  # the container states go to the log and the deploy fails as before.
   echo "[${SERVER}] Starting containers..."
-  run_remote "$SERVER" "recreating ${SERVICES}" "" \
-    "cd ~/netork && \
+  local RECREATE="cd ~/netork && \
     REGISTRY_HOST='${REGISTRY_HOST}' NETORK_VERSION='${SERVER_VERSION}' \
     docker compose -f docker-compose.yml -f docker-compose.registry.yml \
       up -d --no-deps --force-recreate ${SERVICES}"
+  if ! run_remote "$SERVER" "recreating ${SERVICES}" "" "$RECREATE"; then
+    local delay="${DEPLOY_RECREATE_RETRY_DELAY:-5}"
+    echo "[${SERVER}] Recreating failed; trying once more in ${delay}s..." >&2
+    sleep "$delay"
+    if ! run_remote "$SERVER" "recreating ${SERVICES} (second attempt)" "" "$RECREATE"; then
+      echo "[${SERVER}] Containers after two failed attempts:" >&2
+      ssh -n "$SERVER" "cd ~/netork && \
+        REGISTRY_HOST='${REGISTRY_HOST}' NETORK_VERSION='${SERVER_VERSION}' \
+        docker compose -f docker-compose.yml -f docker-compose.registry.yml ps -a" 2>&1 \
+        | sed "s/^/[${SERVER}] /" >&2 || true
+      return 1
+    fi
+  fi
 
   # Infrastructure containers, deliberately left out of the force-recreate
   # above. Plain `up -d`: compose compares each service definition against the
@@ -219,6 +311,24 @@ deploy_server() {
     docker compose -f docker-compose.yml -f docker-compose.registry.yml \
       up -d --no-deps ${INFRA_SERVICES}" || true
 
+  # Services netOrk no longer ships, with the data only they used. `docker
+  # compose up` never removes a container whose service has left the compose
+  # file, so each would keep running on every host until someone removed it by
+  # hand. Idempotent: a host that never had one, or was cleaned already, passes
+  # untouched. Containers first; a volume still in use cannot be removed.
+  #   registry (netork-registry-1, netork_registry_data): the bundled registry:2
+  #     for satellite images. Satellites pull from registry.netork.io; it held one
+  #     old image, nothing had pulled from it in 30 days, and it accepted
+  #     anonymous pushes on port 5000 (NetOrk/netork#763).
+  local RETIRED_CONTAINERS="netork-registry-1"
+  local RETIRED_VOLUMES="netork_registry_data"
+  echo "[${SERVER}] Removing retired services..."
+  run_remote "$SERVER" "removing ${RETIRED_CONTAINERS} ${RETIRED_VOLUMES}" "" \
+    "for c in ${RETIRED_CONTAINERS}; do \
+       docker rm -f \$c >/dev/null 2>&1 && echo \"removed container \$c\"; done; \
+     for v in ${RETIRED_VOLUMES}; do \
+       docker volume rm \$v >/dev/null 2>&1 && echo \"removed volume \$v\"; done; true" || true
+
   # Verify the containers actually run the image we just pulled. `docker
   # compose up -d` reports "Running" (not "Started") when it decides nothing
   # changed, and `pull` prints "Pulled" even when the tag was already local —
@@ -298,9 +408,9 @@ REMOTE_ENV
 
 echo "Deploying from ${REGISTRY_HOST} to: ${SERVERS[*]}"
 
-export -f deploy_server _validate_version run_remote
+export -f deploy_server deploy_steps hold_host_lock release_host_lock _validate_version run_remote
 export UI_SERVER REGISTRY_HOST REGISTRY_USER REGISTRY_PASSWORD NETORK_VERSION VERSION_OVERRIDE \
-  VALID_VERSION_RE COMPOSE_FILES
+  VALID_VERSION_RE COMPOSE_FILES DEPLOY_LOCK_WAIT
 # Export all per-server credential vars so subshells can resolve them
 while IFS='=' read -r key _; do
   if [[ "$key" =~ ^(REGISTRY_(USER|PASSWORD)|NETORK_VERSION)_ ]]; then

tests/test_deploy_host_lock.py

@@ -0,0 +1,185 @@
+"""One deploy per host at a time (NetOrk/deploy#1, #3).
+
+Several sessions deploy to the same test server, and nothing stopped two runs
+from overlapping:
+
+- 2026-10-05: two `up -d --force-recreate` runs recreated each other's
+  containers. The API was down for a minute, with leftover `<id>_netork-…`
+  containers.
+- 2026-10-06: two runs renamed each other's `*.new` compose files, and one broke
+  off at `mv: cannot stat 'docker-compose.yml.new'`. With two different tags,
+  one tag's compose files could have started the other tag's images.
+
+Now each deploy first takes `flock` on `~/netork/.deploy.lock` on the host,
+through an ssh session that holds it until the deploy ends. A second deploy
+says who holds it, since when and with which tag, and waits. The lock lives as
+long as that session does, so a deploy that dies, or a laptop that loses its
+network, frees it on its own.
+
+The "host" here is this machine with `HOME` in a temporary directory: the fake
+`ssh` runs the lock holder for real, so the lock is a real `flock`.
+"""
+
+from __future__ import annotations
+
+import os
+import subprocess
+import time
+from pathlib import Path
+
+import pytest
+
+DEPLOY = Path(__file__).resolve().parent.parent / "deploy.sh"
+
+_FAKE_SSH = """#!/bin/sh
+for arg in "$@"; do cmd="$arg"; done
+case "$cmd" in
+  *.deploy.lock*)
+    # The lock holder runs for real, on this "host".
+    if [ -n "$FAKE_NO_FLOCK" ]; then exec env PATH="$FAKE_NO_FLOCK" sh -c "$cmd"; fi
+    exec sh -c "$cmd"
+    ;;
+esac
+cat > /dev/null
+case "$cmd" in
+  *--force-recreate*)
+    echo "start $RUN_ID" >> "$FAKE_LOG"
+    sleep "${FAKE_RECREATE_SLEEP:-0}"
+    echo "end $RUN_ID" >> "$FAKE_LOG"
+    if [ -n "$FAKE_RECREATE_FAILS" ]; then echo "Error response from daemon"; exit 1; fi
+    ;;
+esac
+exit 0
+"""
+
+
+@pytest.fixture
+def host(tmp_path: Path) -> dict[str, Path]:
+    bin_dir = tmp_path / "bin"
+    bin_dir.mkdir()
+    fake = bin_dir / "ssh"
+    fake.write_text(_FAKE_SSH)
+    fake.chmod(0o755)
+    home = tmp_path / "home"
+    home.mkdir()
+    return {"bin": bin_dir, "home": home, "log": tmp_path / "log"}
+
+
+def _start(host: dict[str, Path], run_id: str, **env: str) -> subprocess.Popen[str]:
+    return subprocess.Popen(
+        ["bash", str(DEPLOY), "testhost"],
+        env={
+            "PATH": f"{host['bin']}:{os.environ['PATH']}",
+            "HOME": str(host["home"]),
+            "USER": "tester",
+            "DEPLOY_ENV_FILE": "/dev/null",
+            "REGISTRY_HOST": "registry.example",
+            "NETORK_VERSION": "latest-dev",
+            "DEPLOY_RECREATE_RETRY_DELAY": "0",
+            "RUN_ID": run_id,
+            "FAKE_LOG": str(host["log"]),
+            **env,
+        },
+        stdin=subprocess.DEVNULL,
+        stdout=subprocess.PIPE,
+        stderr=subprocess.PIPE,
+        text=True,
+    )
+
+
+def _run(host: dict[str, Path], run_id: str, **env: str) -> subprocess.CompletedProcess[str]:
+    proc = _start(host, run_id, **env)
+    out, err = proc.communicate(timeout=60)
+    return subprocess.CompletedProcess(proc.args, proc.returncode, out, err)
+
+
+def _lock_file(host: dict[str, Path]) -> Path:
+    return host["home"] / "netork" / ".deploy.lock"
+
+
+def _lock_is_free(host: dict[str, Path], seconds: float = 5.0) -> bool:
+    deadline = time.monotonic() + seconds
+    while time.monotonic() < deadline:
+        probe = subprocess.run(["flock", "-n", str(_lock_file(host)), "true"], check=False)
+        if probe.returncode == 0:
+            return True
+        time.sleep(0.05)
+    return False
+
+
+def _events(host: dict[str, Path]) -> list[str]:
+    return host["log"].read_text().split("\n")[:-1] if host["log"].exists() else []
+
+
+def _wait_for(host: dict[str, Path], event: str, seconds: float = 20.0) -> None:
+    deadline = time.monotonic() + seconds
+    while time.monotonic() < deadline:
+        if event in _events(host):
+            return
+        time.sleep(0.05)
+    raise AssertionError(f"{event!r} never happened: {_events(host)}")
+
+
+def test_two_deploys_to_one_host_take_turns(host) -> None:
+    first = _start(host, "A", FAKE_RECREATE_SLEEP="1.5")
+    try:
+        _wait_for(host, "start A")
+        second = _run(host, "B")
+    finally:
+        _out, err = first.communicate(timeout=60)
+
+    assert first.returncode == 0, err
+    assert second.returncode == 0, second.stderr
+    assert _events(host) == ["start A", "end A", "start B", "end B"]
+    assert "another deploy holds testhost" in second.stderr
+    assert "deploying latest-dev" in second.stderr
+    assert "by tester@" in second.stderr
+
+
+def test_the_lock_is_free_once_a_deploy_is_done(host) -> None:
+    result = _run(host, "A")
+
+    assert result.returncode == 0, result.stderr
+    assert _lock_file(host).exists()
+    assert _lock_is_free(host)
+
+
+def test_a_failed_deploy_frees_the_lock_too(host) -> None:
+    result = _run(host, "A", FAKE_RECREATE_FAILS="1")
+
+    assert result.returncode != 0
+    assert _lock_is_free(host)
+
+
+def test_a_deploy_that_waits_too_long_gives_up_without_touching_the_host(host) -> None:
+    lock = _lock_file(host)
+    lock.parent.mkdir(parents=True)
+    holder = subprocess.Popen(["flock", str(lock), "sleep", "30"])
+    try:
+        time.sleep(0.2)
+        result = _run(host, "B", DEPLOY_LOCK_WAIT="1")
+    finally:
+        holder.kill()
+        holder.wait()
+
+    assert result.returncode != 0
+    assert "still held after 1s" in result.stderr
+    assert "FAILED on: testhost" in result.stderr
+    assert _events(host) == []
+
+
+def test_a_host_without_flock_is_deployed_with_a_warning(host, tmp_path) -> None:
+    """flock is util-linux; a host without it should not become undeployable."""
+    no_flock = tmp_path / "no-flock"
+    no_flock.mkdir()
+    for tool in ("sh", "mkdir", "cat", "date"):
+        found = subprocess.run(
+            ["sh", "-c", f"command -v {tool}"], capture_output=True, text=True, check=True
+        )
+        (no_flock / tool).symlink_to(found.stdout.strip())
+
+    result = _run(host, "A", FAKE_NO_FLOCK=str(no_flock))
+
+    assert result.returncode == 0, result.stderr
+    assert "no flock on testhost" in result.stderr
+    assert _events(host) == ["start A", "end A"]

tests/test_deploy_recreate_retry.py

@@ -0,0 +1,108 @@
+"""A failed container recreate is tried once more before the deploy gives up.
+
+On 2026-10-05 a deploy to 172.22.8.50 failed in "Starting containers":
+`docker compose up -d --force-recreate` recreates the services in parallel, lost
+a container it had just renamed ("No such container: 02586df7…"), and stopped
+with every engine container `Created` and none running. The old API was already
+gone, so netOrk was down until somebody ran the same deploy again, which then
+went through cleanly (NetOrk/netork#586).
+
+So the script runs that step a second time on its own. Should the second attempt
+fail too, it shows what the containers look like and fails loudly, as before.
+
+The remote side is a fake `ssh` that answers each step and can fail the recreate
+a given number of times.
+"""
+
+from __future__ import annotations
+
+import os
+import subprocess
+from pathlib import Path
+
+import pytest
+
+DEPLOY = Path(__file__).resolve().parent.parent / "deploy.sh"
+
+_FAKE_SSH = """#!/bin/sh
+for arg in "$@"; do cmd="$arg"; done
+# The host lock's holder runs for real, here (test_deploy_host_lock.py).
+case "$cmd" in *.deploy.lock*) exec sh -c "$cmd" ;; esac
+# Swallow stdin (docker login, the .env heredoc), then answer by command.
+cat > /dev/null
+case "$cmd" in
+  *--force-recreate*)
+    n=$(cat "{state}/recreates" 2>/dev/null || echo 0)
+    n=$((n + 1))
+    echo "$n" > "{state}/recreates"
+    if [ "$n" -le "{failures}" ]; then
+      echo "Error response from daemon: No such container: 02586df7e659"
+      exit 1
+    fi
+    echo " Container netork-netork-api-1 Started"
+    ;;
+  *" ps -a"*)
+    echo "netork-netork-api-1 Created"
+    ;;
+esac
+exit 0
+"""
+
+
+@pytest.fixture
+def deploy(tmp_path: Path):
+    """Run a deploy against the fake ssh; *failures* recreates fail before one works."""
+
+    def _deploy(failures: int) -> tuple[subprocess.CompletedProcess[str], int]:
+        bin_dir = tmp_path / "bin"
+        bin_dir.mkdir(exist_ok=True)
+        fake = bin_dir / "ssh"
+        fake.write_text(
+            _FAKE_SSH.replace("{state}", str(tmp_path)).replace("{failures}", str(failures))
+        )
+        fake.chmod(0o755)
+        result = subprocess.run(
+            ["bash", str(DEPLOY), "testhost"],
+            env={
+                "PATH": f"{bin_dir}:{os.environ['PATH']}",
+                "HOME": str(tmp_path),
+                "DEPLOY_ENV_FILE": "/dev/null",
+                "REGISTRY_HOST": "registry.example",
+                "NETORK_VERSION": "latest-dev",
+                "DEPLOY_RECREATE_RETRY_DELAY": "0",
+            },
+            stdin=subprocess.DEVNULL,
+            capture_output=True,
+            text=True,
+            timeout=60,
+            check=False,
+        )
+        recreates = tmp_path / "recreates"
+        return result, int(recreates.read_text()) if recreates.exists() else 0
+
+    return _deploy
+
+
+def test_a_recreate_that_works_runs_once(deploy) -> None:
+    result, recreates = deploy(failures=0)
+
+    assert result.returncode == 0, result.stderr
+    assert recreates == 1
+
+
+def test_a_recreate_that_fails_once_is_tried_again(deploy) -> None:
+    result, recreates = deploy(failures=1)
+
+    assert result.returncode == 0, result.stderr
+    assert recreates == 2
+    assert "trying once more" in result.stderr
+    assert "All servers deployed successfully." in result.stdout
+
+
+def test_two_failed_recreates_fail_the_deploy_and_show_the_containers(deploy) -> None:
+    result, recreates = deploy(failures=2)
+
+    assert result.returncode != 0
+    assert recreates == 2
+    assert "netork-netork-api-1 Created" in result.stderr
+    assert "FAILED on: testhost" in result.stderr

tests/test_deploy_retired_services.py

@@ -0,0 +1,73 @@
+"""A service netOrk no longer ships is removed from the host (NetOrk/netork#763).
+
+`docker compose up` never removes a container whose service has left the compose
+file, so it would keep running on every host until somebody removed it by hand.
+The first of these is the bundled `registry:2`: it held only an old satellite
+image, nothing had pulled from it in 30 days, and it accepted anonymous pushes on
+port 5000 of every instance.
+
+The remote side is a fake `ssh` that records every command it is given.
+"""
+
+from __future__ import annotations
+
+import os
+import subprocess
+from pathlib import Path
+
+DEPLOY = Path(__file__).resolve().parent.parent / "deploy.sh"
+
+_FAKE_SSH = """#!/bin/sh
+for arg in "$@"; do cmd="$arg"; done
+# The host lock's holder runs for real, here (test_deploy_host_lock.py).
+case "$cmd" in *.deploy.lock*) exec sh -c "$cmd" ;; esac
+cat > /dev/null
+printf '%s\\n----\\n' "$cmd" >> "{state}/commands"
+exit 0
+"""
+
+
+def _deploy(tmp_path: Path) -> tuple[subprocess.CompletedProcess[str], list[str]]:
+    bin_dir = tmp_path / "bin"
+    bin_dir.mkdir()
+    fake = bin_dir / "ssh"
+    fake.write_text(_FAKE_SSH.replace("{state}", str(tmp_path)))
+    fake.chmod(0o755)
+    result = subprocess.run(
+        ["bash", str(DEPLOY), "testhost"],
+        env={
+            "PATH": f"{bin_dir}:{os.environ['PATH']}",
+            "HOME": str(tmp_path),
+            "DEPLOY_ENV_FILE": "/dev/null",
+            "REGISTRY_HOST": "registry.example",
+            "NETORK_VERSION": "latest-dev",
+        },
+        stdin=subprocess.DEVNULL,
+        capture_output=True,
+        text=True,
+        timeout=60,
+        check=False,
+    )
+    commands = (tmp_path / "commands").read_text().split("\n----\n")
+    return result, commands
+
+
+def test_a_deploy_removes_the_retired_registry_and_its_volume(tmp_path: Path) -> None:
+    result, commands = _deploy(tmp_path)
+
+    assert result.returncode == 0, result.stderr
+    removal = [c for c in commands if "netork-registry-1" in c]
+    assert removal, "no command removes the netork-registry-1 container"
+    assert "docker rm -f" in removal[0]
+    assert "docker volume rm" in removal[0] and "netork_registry_data" in removal[0]
+    # The container goes first: a volume still in use cannot be removed.
+    assert removal[0].index("docker rm -f") < removal[0].index("docker volume rm")
+
+
+def test_the_infrastructure_reconcile_no_longer_starts_the_registry(tmp_path: Path) -> None:
+    result, commands = _deploy(tmp_path)
+
+    assert result.returncode == 0, result.stderr
+    reconcile = [c for c in commands if "up -d --no-deps" in c and "--force-recreate" not in c]
+    assert reconcile, "the infrastructure reconcile step is missing"
+    assert " registry" not in reconcile[0].split("up -d --no-deps", 1)[1]