fix(install): 刚装完的服务会失败，单元要说清楚而不是每 5 秒重启一次

我在真内核上把安装序列完整走了一遍（容器里补了 useradd/systemctl 的替身，其余全是真的），
暴露出一个只有真跑才看得见的缺陷：

装完 → `install` 启动服务 → 但那时**还没有配对的手机**，而守护进程**故意**拒绝空的设备列表
（一个什么都不能授权的 agent 不该坐在那里显得健康）。于是 `Restart=always` + `RestartSec=5`
变成每 5 秒重启一次，永远，journal 里刷同一行。实测：exit 1，"devices is empty, so no phone
could ever be authorised"。

配对发生在安装**之后**（运维在节点上跑 `--pair`，然后在手机上配对），所以这不是异常状态，
而是流程的必经一步。修法两条：

- 单元加 `StartLimitIntervalSec=300` / `StartLimitBurst=10`：十次之后 systemd 放弃，单元进入
  `failed`——一个运维看得见、能处理的状态，而不是无限重启。
- `install` 的结束语改成说清这一步：服务会退出并停住，下一步是 `--pair`，用
  `systemctl status` 看状态。此前它说"installed and running"，而那是假的。

同一轮演练确认了此前所有修复在真内核上成立：服务用户建了、配置与日志目录交给它了、配置对
该用户可读、单元来自二进制且内容一致、卸载按序拆干净（用户/目录/单元全无残留）。

Go 全包通过（新增一项守着"单元必须有启动上限，且 install 必须说出下一步"）。

internal/install/uninstall_test.go

@@ -426,3 +426,51 @@ func TestTheRemovalHintIsTheOneItWasGiven(t *testing.T) {
 		t.Fatalf("the uninstall still claims every node is in agents.json:\n%s", output)
 	}
 }
+
+// The install ends by starting a service that cannot run yet, and the unit has to say what to do about
+// that. A freshly installed node has no paired devices — pairing happens after install — and the daemon
+// refuses to run in that state on purpose, because an agent that could authorise nothing must not look
+// healthy. With `Restart=always` and `RestartSec=5` and no start limit, that is a restart every five
+// seconds until a phone pairs.
+//
+// Found by rehearsing the install on a real kernel: exit 1, "devices is empty, so no phone could ever be
+// authorised", and the journal filling with the same line.
+func TestTheUnitGivesUpInsteadOfRestartingForever(t *testing.T) {
+	unit := string(UnitFile())
+	if !strings.Contains(unit, "Restart=always") {
+		t.Skip("the unit no longer restarts, so there is nothing to bound")
+	}
+	if !strings.Contains(unit, "StartLimitIntervalSec=") || !strings.Contains(unit, "StartLimitBurst=") {
+		t.Fatal("the unit restarts without a start limit: an unpaired node would restart every RestartSec forever")
+	}
+	// And the install tells the operator what to do next, so the failed state is expected rather than
+	// surprising.
+	var lines []string
+	root := t.TempDir()
+	paths := Paths{
+		ServiceName: "msa-agent",
+		UnitPath:    filepath.Join(root, "etc/systemd/system/msa-agent.service"),
+		ConfigPath:  filepath.Join(root, "etc/msa-agent/config.json"),
+		LogDir:      filepath.Join(root, "var/log/msa-agent"),
+		BinaryPath:  filepath.Join(root, "opt/msa-agent/msa-agent"),
+		User:        "msa-agent",
+	}
+	for _, path := range []string{paths.BinaryPath, paths.ConfigPath} {
+		if err := os.MkdirAll(filepath.Dir(path), 0o755); err != nil {
+			t.Fatal(err)
+		}
+		if err := os.WriteFile(path, []byte("x"), 0o600); err != nil {
+			t.Fatal(err)
+		}
+	}
+	if err := Install(InstallOptions{
+		Paths:  paths,
+		Runner: &recorder{failOn: "getent"},
+		Log:    func(format string, args ...any) { lines = append(lines, fmt.Sprintf(format, args...)) },
+	}); err != nil {
+		t.Fatalf("Install: %v", err)
+	}
+	if !strings.Contains(strings.Join(lines, "\n"), "--pair") {
+		t.Fatalf("the install does not say to pair next:\n%s", strings.Join(lines, "\n"))
+	}
+}