fix(bootstrap): never stop the dashboard from inside its own systemd cgroup

A plugin self-update took the dashboard down and left it down. The chain
ran correctly up to the point where the bootstrap SIGTERMs the server it is
replacing — and that server was systemd-managed, with the bootstrap running
as its descendant, i.e. inside its cgroup. Two failures fired at once:

  * the server exits 0, and `Restart=on-failure` reads a clean exit as
    "meant to stop", so systemd did not bring it back;
  * `KillMode=control-group` reaped the bootstrap itself, mid-run, before it
    could repoint the unit at the new plugin root or write state.json.

Evidence from the incident: bootstrap.log ends at "stopped the server
started from the previous plugin version" with no systemd line after it,
.bootstrap.lock survived with a dead pid, state.json still named the
previous version, and the unit's ExecStart still pointed at the old cache
dir. The same truncation is visible on three earlier updates — those only
recovered because a later SessionStart bootstrapped again.

systemd-service.js now splits install into writeUserService (repoints the
unit, starts nothing) and restartUserService (--no-block). The bootstrap
skips the manual stop whenever a unit of ours owns the server, and defers
the restart past writeState and releaseLock, so being killed by its own
request costs nothing. The unit is Restart=always: a graceful SIGTERM from
anything must no longer read as "leave it stopped".

installUserService keeps its old semantics for `agenthub service install`,
which runs from a user's shell and is not in the cgroup.

scripts/systemd-service.js

@@ -26,11 +26,43 @@ function unitPath(name) {
   return path.join(unitDir(), `${name}.service`);
 }
 
-function systemctlAvailable() {
-  if (process.platform !== "linux") return false;
-  return spawnSync("systemctl", ["--user", "--version"], { stdio: "ignore" }).status === 0;
+// Test seam: swap the systemctl implementation so unit tests never talk to a
+// real service manager. Mirrors tmux.js's __setExecImpl/__reset style.
+let execImpl = (args, opts) => spawnSync("systemctl", ["--user", ...args], opts);
+
+function __setExecImpl(fn) {
+  execImpl = fn;
+}
+function __reset() {
+  execImpl = (args, opts) => spawnSync("systemctl", ["--user", ...args], opts);
 }
 
+function systemctlAvailable() {
+  if (process.platform !== "linux") return false;
+  return execImpl(["--version"], { stdio: "ignore" }).status === 0;
+}
+
+/**
+ * Is there a unit by this name, and did WE write it? The distinction decides
+ * whether the dashboard may be stopped by hand: a unit of ours means systemd
+ * owns the process, and SIGTERM-ing it out of band is the one thing that must
+ * never happen (see `restartUserService`).
+ */
+function unitIsOurs(name) {
+  try {
+    return fs.readFileSync(unitPath(name), "utf8").includes(MARKER);
+  } catch {
+    return false;
+  }
+}
+
+/**
+ * `Restart=always`, not `on-failure`: a graceful SIGTERM exits 0, which
+ * `on-failure` reads as "meant to stop". One stray kill — or an out-of-band
+ * stop during a plugin self-update — then left the dashboard down until some
+ * later Claude Code session happened to run the bootstrap. An explicit
+ * `systemctl stop` (what uninstall does) still stops it for good.
+ */
 function unitContent({ description, workingDir, execStart, env }) {
   const envLines = Object.entries(env)
     .map(([k, v]) => `Environment=${k}=${v}`)
@@ -44,7 +76,7 @@ After=network.target
 Type=simple
 WorkingDirectory=${workingDir}
 ExecStart=${execStart}
-Restart=on-failure
+Restart=always
 RestartSec=3
 ${envLines}
 
@@ -54,17 +86,18 @@ WantedBy=default.target
 }
 
 function systemctl(args) {
-  return spawnSync("systemctl", ["--user", ...args], { stdio: "inherit" });
+  return execImpl(args, { stdio: "inherit" });
 }
 
 /**
- * Writes (or refreshes) the unit file and enables + starts it. Idempotent:
- * an unchanged unit is left running as-is; a changed one (e.g. a plugin
- * cache path after an update) is reloaded and restarted.
+ * Writes (or refreshes) the unit file and enables it for boot — but starts
+ * nothing. Split out of `installUserService` for the plugin self-update, which
+ * must be able to point the unit at the new cache directory *before* anything
+ * stops the running server, and defer the restart to its very last act.
  *
  * @returns {"written"|"unchanged"|"foreign"|"unavailable"|"failed"}
  */
-function installUserService({ name, description, workingDir, execStart, env = {} }) {
+function writeUserService({ name, description, workingDir, execStart, env = {} }) {
   if (!systemctlAvailable()) return "unavailable";
   const body = unitContent({ description, workingDir, execStart, env });
   const target = unitPath(name);
@@ -72,22 +105,56 @@ function installUserService({ name, description, workingDir, execStart, env = {}
     if (fs.existsSync(target)) {
       const current = fs.readFileSync(target, "utf8");
       if (!current.includes(MARKER)) return "foreign";
-      if (current === body) {
-        systemctl(["enable", "--now", `${name}.service`]);
-        return "unchanged";
-      }
+      if (current === body) return "unchanged";
     }
     fs.mkdirSync(unitDir(), { recursive: true });
     fs.writeFileSync(target, body, "utf8");
     systemctl(["daemon-reload"]);
-    systemctl(["enable", "--now", `${name}.service`]);
-    systemctl(["restart", `${name}.service`]);
+    systemctl(["enable", `${name}.service`]);
     return "written";
   } catch {
     return "failed";
   }
 }
 
+/**
+ * Ask systemd to restart the unit and do NOT wait for it.
+ *
+ * `--no-block` is the whole point, not a nicety. The caller is the bootstrap
+ * spawned by the running server, so it lives in that server's cgroup; the
+ * moment systemd stops the old main process, `KillMode=control-group` reaps
+ * the caller too. Blocking here means being killed mid-call, which is why the
+ * restart must be the last thing the bootstrap does — by then there is nothing
+ * left to lose.
+ */
+function restartUserService(name) {
+  if (!systemctlAvailable()) return "unavailable";
+  const res = systemctl(["restart", "--no-block", `${name}.service`]);
+  return res.status === 0 ? "restarted" : "failed";
+}
+
+/**
+ * Writes (or refreshes) the unit file and enables + starts it. Idempotent:
+ * an unchanged unit is left running as-is; a changed one (e.g. a plugin
+ * cache path after an update) is reloaded and restarted.
+ *
+ * Safe for `agenthub service install`, which is invoked from a user's own
+ * shell and is therefore not inside the unit's cgroup. The plugin bootstrap
+ * is, and uses `writeUserService` + `restartUserService` instead.
+ *
+ * @returns {"written"|"unchanged"|"foreign"|"unavailable"|"failed"}
+ */
+function installUserService(opts) {
+  const result = writeUserService(opts);
+  if (result === "unchanged") {
+    systemctl(["enable", "--now", `${opts.name}.service`]);
+  } else if (result === "written") {
+    systemctl(["enable", "--now", `${opts.name}.service`]);
+    systemctl(["restart", `${opts.name}.service`]);
+  }
+  return result;
+}
+
 /** @returns {"removed"|"absent"|"foreign"|"unavailable"} */
 function uninstallUserService(name) {
   if (!systemctlAvailable()) return "unavailable";
@@ -110,8 +177,13 @@ function enableLinger() {
 
 module.exports = {
   systemctlAvailable,
+  unitIsOurs,
+  writeUserService,
+  restartUserService,
   installUserService,
   uninstallUserService,
   enableLinger,
   unitPath,
+  __setExecImpl,
+  __reset,
 };