fix(k3s): route audit logs through journald, codify etcd member recovery

Two changes prompted by today's etcd raft panic on worker1-k8s0
(tocommit out of range, lost-write on follower) and the cascading
disk pressure that surfaced underneath it.

Audit logs to journald
- kube-apiserver now uses audit-log-path=- so audit events flow to
  k3s.service stdout and into journald instead of growing files in
  /var/log/kubernetes. The previous setup combined apiserver's
  internal rotation with a logrotate *.log glob that double-rotated
  the rotated files into permanent orphans (observed: 7+ GB).
- New journald-limits operation writes a SystemMaxUse=2G drop-in so
  audit volume cannot fill /var/log even under bursty load.
- log-rotation operation repurposed to decommission the obsolete
  logrotate rule and reap leftover audit files. Idempotent: no-op
  on fresh installs.

Etcd member recovery
- New recoverEtcdMember(broken, peer, hostname) codifies the
  documented k3s recovery: stop k3s, etcdctl member remove, wipe
  /var/lib/rancher/k3s/server/{db,tls,cred}, restart, poll for
  rejoin. Refuses to operate when cluster size < 3 to preserve
  quorum.

Tests
- 7 new unit tests covering both decommission paths and the
  recovery procedure (54 total, all green).
- install.test.ts asserts the file-based audit args are gone.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

bastion/src/modules/modules/k3s/src/groups/hardening.ts

@@ -1,19 +1,21 @@
-// Hardening: Pod Security Standards, certificate check, log rotation.
+// Hardening: Pod Security Standards, certificate check, journald cap, storage.
 
 import type { OperationContext, OperationResult, OperationGroup } from "../types.js";
 import { runSequential } from "../utils.js";
 import { applyPodSecurityStandards } from "../operations/pod-security.js";
 import { checkCertExpiry } from "../operations/cert-check.js";
 import { configureLogRotation } from "../operations/log-rotation.js";
+import { configureJournaldLimits } from "../operations/journald-limits.js";
 import { configureLonghornDisk } from "../operations/longhorn-disk.js";
 
 export const hardeningGroup: OperationGroup = {
   name: "hardening",
-  description: "Pod security, certificate check, log rotation, storage",
+  description: "Pod security, certificate check, journald cap, storage",
   operations: [
     { name: "Apply Pod Security Standards", fn: applyPodSecurityStandards },
     { name: "Check certificate expiry", fn: checkCertExpiry },
-    { name: "Configure log rotation", fn: configureLogRotation },
+    { name: "Decommission file-based audit logs", fn: configureLogRotation },
+    { name: "Configure journald disk cap", fn: configureJournaldLimits },
     { name: "Configure Longhorn disk", fn: configureLonghornDisk },
   ],
 };

bastion/src/modules/modules/k3s/src/install.ts

@@ -76,7 +76,6 @@ sed -i 's/^SELINUX=enforcing/SELINUX=permissive/' /etc/selinux/config 2>/dev/nul
 # ── 5b. Create k3s config directory ──
 echo "[5/10] Writing k3s server configuration..."
 mkdir -p /etc/rancher/k3s
-mkdir -p /var/log/kubernetes
 
 cat > /etc/rancher/k3s/config.yaml << 'K3S_CONFIG'
 # k3s server configuration — CIS hardened
@@ -91,13 +90,10 @@ disable:
   - servicelb
   - traefik
 
-# API server hardening
+# API server hardening (audit-log-path=- routes audit to journald via stdout)
 kube-apiserver-arg:
   - "anonymous-auth=false"
-  - "audit-log-path=/var/log/kubernetes/audit.log"
+  - "audit-log-path=-"
-  - "audit-log-maxage=30"
-  - "audit-log-maxbackup=10"
-  - "audit-log-maxsize=100"
   - "audit-policy-file=/etc/rancher/k3s/audit-policy.yaml"
   - "enable-admission-plugins=NodeRestriction,PodSecurity"
   - "request-timeout=300s"

bastion/src/modules/modules/k3s/src/operations/etcd-recover.ts

@@ -0,0 +1,194 @@
+// Recover a broken etcd member by removing it from the cluster, wiping its
+// local state, and restarting k3s so it rejoins as a fresh member.
+//
+// Use case: a node panics on startup with
+//   "tocommit(N+1) is out of range [lastIndex(N)]. Was the raft log corrupted,
+//    truncated, or lost?"
+// This means the local raft WAL is missing the last entry the leader thinks
+// the follower acknowledged (lost write, unclean shutdown, etc). The fix is
+// always the same and well-documented; this codifies it so we don't fumble
+// the procedure under pressure.
+//
+// Preconditions:
+//   - At least one healthy peer is reachable so the cluster has quorum after
+//     we remove the broken member. (For a 3-node cluster: 2 healthy. For a
+//     5-node: 3 healthy.) If quorum would be lost, this function refuses.
+//   - SSH access to both the broken node and a healthy peer.
+//   - etcdctl available on the healthy peer (k3s does not bundle it; the
+//     procedure installs it on demand on Fedora).
+
+import type { SshClient } from "../types.js";
+
+const ETCD_TLS = {
+  ca: "/var/lib/rancher/k3s/server/tls/etcd/server-ca.crt",
+  cert: "/var/lib/rancher/k3s/server/tls/etcd/server-client.crt",
+  key: "/var/lib/rancher/k3s/server/tls/etcd/server-client.key",
+} as const;
+
+const SSH_TIMEOUT = 60_000;
+
+export interface RecoverEtcdMemberOptions {
+  /** SSH client for the broken node (the one panicking). */
+  broken: SshClient;
+  /** SSH client for any healthy server peer in the same cluster. */
+  peer: SshClient;
+  /** Hostname (k8s node name) of the broken node. Used to find its etcd member id. */
+  brokenHostname: string;
+  /** Logger for progress output. */
+  log?: (msg: string) => void;
+}
+
+export interface RecoverEtcdMemberResult {
+  success: boolean;
+  changed: boolean;
+  message: string;
+  /** New etcd member id assigned after rejoin (when known). */
+  newMemberId?: string;
+  /** Old etcd member id that was removed. */
+  removedMemberId?: string;
+  error?: string;
+}
+
+function etcdctl(subcmd: string): string {
+  return [
+    "ETCDCTL_API=3 etcdctl",
+    `--cacert=${ETCD_TLS.ca}`,
+    `--cert=${ETCD_TLS.cert}`,
+    `--key=${ETCD_TLS.key}`,
+    "--endpoints=https://127.0.0.1:2379",
+    "--command-timeout=10s",
+    subcmd,
+  ].join(" ");
+}
+
+async function ensureEtcdctl(peer: SshClient): Promise<void> {
+  const probe = await peer.exec("command -v etcdctl 2>/dev/null", { timeoutMs: 5_000 });
+  if (probe.exitCode === 0 && probe.stdout.trim()) return;
+  // Best-effort install on Fedora. If the host isn't dnf-based, surface the
+  // error to the caller via the next etcdctl invocation.
+  await peer.exec("dnf install -y etcd 2>&1", { timeoutMs: 120_000 });
+}
+
+async function getMemberList(peer: SshClient): Promise<Array<{ id: string; name: string }>> {
+  const result = await peer.exec(etcdctl("member list"), { timeoutMs: SSH_TIMEOUT });
+  if (result.exitCode !== 0) {
+    throw new Error(`etcdctl member list failed: ${result.stderr || result.stdout}`);
+  }
+  // Format: <hex-id>, started, <name>, <peer-urls>, <client-urls>, <isLearner>
+  return result.stdout
+    .split("\n")
+    .map((line) => line.trim())
+    .filter(Boolean)
+    .map((line) => {
+      const [id, , name] = line.split(",").map((p) => p.trim());
+      return { id: id ?? "", name: name ?? "" };
+    })
+    .filter((m) => m.id);
+}
+
+export async function recoverEtcdMember(
+  opts: RecoverEtcdMemberOptions,
+): Promise<RecoverEtcdMemberResult> {
+  const log = opts.log ?? (() => {});
+
+  try {
+    log(`Looking up etcd member id for ${opts.brokenHostname} via peer...`);
+    await ensureEtcdctl(opts.peer);
+
+    const members = await getMemberList(opts.peer);
+    if (members.length < 3) {
+      return {
+        success: false,
+        changed: false,
+        message: "Refusing to remove a member from a cluster with <3 members (quorum would be lost)",
+        error: `member count = ${members.length}`,
+      };
+    }
+
+    // Member names are <hostname>-<random-suffix>; match by hostname prefix.
+    const broken = members.find((m) => m.name.startsWith(opts.brokenHostname));
+    if (!broken) {
+      return {
+        success: false,
+        changed: false,
+        message: `No etcd member found matching hostname ${opts.brokenHostname}`,
+        error: `members: ${members.map((m) => m.name).join(", ")}`,
+      };
+    }
+    log(`Broken member: ${broken.id} (${broken.name})`);
+
+    log("Step 1/4: stopping k3s on broken node");
+    await opts.broken.exec("systemctl stop k3s 2>&1", { timeoutMs: SSH_TIMEOUT });
+
+    log("Step 2/4: removing broken etcd member from cluster");
+    const remove = await opts.peer.exec(
+      etcdctl(`member remove ${broken.id}`),
+      { timeoutMs: SSH_TIMEOUT },
+    );
+    if (remove.exitCode !== 0) {
+      return {
+        success: false,
+        changed: false,
+        message: "etcdctl member remove failed",
+        error: remove.stderr || remove.stdout,
+        removedMemberId: broken.id,
+      };
+    }
+
+    log("Step 3/4: archiving corrupt etcd state and stale TLS/cred dirs on broken node");
+    const ts = Math.floor(Date.now() / 1000);
+    await opts.broken.exec(
+      [
+        `mv /var/lib/rancher/k3s/server/db /var/lib/rancher/k3s/server/db.corrupt-${ts} 2>/dev/null || true`,
+        "rm -rf /var/lib/rancher/k3s/server/tls /var/lib/rancher/k3s/server/cred",
+      ].join(" && "),
+      { timeoutMs: SSH_TIMEOUT },
+    );
+
+    log("Step 4/4: starting k3s on broken node — it will rejoin");
+    await opts.broken.exec("systemctl start k3s 2>&1", { timeoutMs: SSH_TIMEOUT });
+
+    // Poll for rejoin. The new member-id is what the cluster assigns on join.
+    let newMemberId: string | undefined;
+    for (let i = 0; i < 60; i++) {
+      await new Promise((r) => setTimeout(r, 5_000));
+      try {
+        const after = await getMemberList(opts.peer);
+        const rejoined = after.find(
+          (m) => m.name.startsWith(opts.brokenHostname) && m.id !== broken.id,
+        );
+        if (rejoined) {
+          newMemberId = rejoined.id;
+          break;
+        }
+      } catch {
+        // peer may briefly be unreachable mid-rejoin — keep polling
+      }
+    }
+
+    if (!newMemberId) {
+      return {
+        success: false,
+        changed: true,
+        message: "k3s started but new member did not appear in cluster within 5 minutes",
+        removedMemberId: broken.id,
+      };
+    }
+
+    log(`Rejoined as ${newMemberId}`);
+    return {
+      success: true,
+      changed: true,
+      message: `Recovered: removed ${broken.id}, rejoined as ${newMemberId}`,
+      removedMemberId: broken.id,
+      newMemberId,
+    };
+  } catch (err) {
+    return {
+      success: false,
+      changed: false,
+      message: "Recovery failed",
+      error: err instanceof Error ? err.message : String(err),
+    };
+  }
+}

bastion/src/modules/modules/k3s/src/operations/index.ts

@@ -11,7 +11,13 @@ export { installK3sBinary } from "./k3s-install.js";
 export { installCilium } from "./cilium.js";
 export { fixCoreDnsUpstream } from "./dns-fix.js";
 export { configureLogRotation } from "./log-rotation.js";
+export { configureJournaldLimits } from "./journald-limits.js";
 export { applyDefaultNetworkPolicies } from "./network-policy.js";
 export { applyPodSecurityStandards } from "./pod-security.js";
 export { checkCertExpiry } from "./cert-check.js";
 export { configureLonghornDisk } from "./longhorn-disk.js";
+export { recoverEtcdMember } from "./etcd-recover.js";
+export type {
+  RecoverEtcdMemberOptions,
+  RecoverEtcdMemberResult,
+} from "./etcd-recover.js";

bastion/src/modules/modules/k3s/src/operations/journald-limits.ts

@@ -0,0 +1,33 @@
+// Cap journald disk usage so audit logs (which now flow through journald via
+// kube-apiserver's stdout) cannot fill /var/log. Default journald uses up to
+// 10% of the filesystem, capped at 4 GB. In a /var/log of ~10 GB shared with
+// other services, that's still room for audit volume to evict useful logs.
+// 2 GB / 200 MB-per-file is a comfortable middle.
+
+import type { Operation, OperationResult } from "../types.js";
+import { sshOpts, writeRemoteFile } from "../utils.js";
+
+const DROPIN_CONTENT = `[Journal]
+SystemMaxUse=2G
+SystemKeepFree=1G
+SystemMaxFileSize=200M
+`;
+
+const DROPIN_PATH = "/etc/systemd/journald.conf.d/10-k3s-audit-cap.conf";
+
+export const configureJournaldLimits: Operation = async (ctx): Promise<OperationResult> => {
+  const changed = await writeRemoteFile(ctx, DROPIN_PATH, DROPIN_CONTENT);
+  if (changed) {
+    // Reload journald so the new limit applies without a reboot.
+    await ctx.ssh.exec(
+      "systemctl kill --signal=SIGUSR2 systemd-journald 2>/dev/null; " +
+      "systemctl restart systemd-journald 2>&1 || true",
+      sshOpts(ctx),
+    );
+  }
+  return {
+    success: true,
+    changed,
+    message: changed ? "journald limits configured (2 GB cap)" : "journald limits already configured",
+  };
+};

bastion/src/modules/modules/k3s/src/operations/k3s-config.ts

@@ -13,6 +13,12 @@ function generateServerConfig(config: K3sConfig): string {
   const clusterLines = isJoining
     ? `server: "${config.k3sServerUrl}"\ntoken: "${config.k3sToken}"`
     : "cluster-init: true";
+  // audit-log-path=- routes audit events to k3s.service's stdout, which systemd
+  // forwards to journald. journald enforces its own size caps (see
+  // configureJournaldLimits) so audit volume cannot fill the disk. File-based
+  // audit logs led to /var/log/kubernetes growing to 7+ GB because apiserver's
+  // own rotation produced files that any logrotate glob would double-rotate
+  // and never expire.
   return `# k3s server configuration — CIS hardened, etcd HA
 ${clusterLines}
 protect-kernel-defaults: true
@@ -30,10 +36,7 @@ node-label:
 
 kube-apiserver-arg:
   - "anonymous-auth=false"
-  - "audit-log-path=/var/log/kubernetes/audit.log"
+  - "audit-log-path=-"
-  - "audit-log-maxage=30"
-  - "audit-log-maxbackup=10"
-  - "audit-log-maxsize=100"
   - "audit-policy-file=/etc/rancher/k3s/audit-policy.yaml"
   - "enable-admission-plugins=NodeRestriction,PodSecurity"
   - "request-timeout=300s"
@@ -61,7 +64,7 @@ kubelet-arg:
 }
 
 export const writeK3sConfig: Operation = async (ctx): Promise<OperationResult> => {
-  await ctx.ssh.exec("mkdir -p /etc/rancher/k3s /var/log/kubernetes", sshOpts(ctx));
+  await ctx.ssh.exec("mkdir -p /etc/rancher/k3s", sshOpts(ctx));
 
   const content = isServerRole(ctx.config.role)
     ? generateServerConfig(ctx.config)

bastion/src/modules/modules/k3s/src/operations/log-rotation.ts

@@ -1,25 +1,44 @@
-// Configure log rotation for k3s.
+// Decommission file-based k8s audit logging in favor of journald.
+//
+// Earlier versions wrote audit events to /var/log/kubernetes/audit.log and
+// rotated them with a logrotate rule. Two failure modes followed: kube-apiserver
+// rotated internally (audit-{ts}.log), the *.log glob in logrotate
+// double-rotated those (-{date}), and the resulting filename matched no
+// retention policy, so the directory grew unbounded (we observed 7+ GB).
+//
+// k3s now sets audit-log-path=- so audit goes to stdout → journald, which
+// enforces SystemMaxUse caps. This operation removes the obsolete logrotate
+// rule and reaps any audit files left behind by the old setup. Idempotent: on
+// fresh installs everything is already absent and the operation is a no-op.
 
 import type { Operation, OperationResult } from "../types.js";
-import { writeRemoteFile } from "../utils.js";
+import { sshOpts } from "../utils.js";
 
-const LOGROTATE_CONFIG = `/var/log/kubernetes/*.log {
+const REMOVE_LOGROTATE = "rm -f /etc/logrotate.d/k3s";
-  daily
+
-  rotate 14
+// Bounded by a max-depth and explicit name pattern so we never reach outside
-  compress
+// the deprecated audit-log directory.
-  delaycompress
+const REAP_OLD_AUDIT_FILES =
-  missingok
+  "find /var/log/kubernetes -maxdepth 1 -type f " +
-  notifempty
+  "\\( -name 'audit*.log*' -o -name 'audit-*.log' \\) " +
-  copytruncate
+  "-delete 2>/dev/null; " +
-  maxsize 100M
+  "rmdir /var/log/kubernetes 2>/dev/null; true";
-}`;
 
 export const configureLogRotation: Operation = async (ctx): Promise<OperationResult> => {
-  const changed = await writeRemoteFile(ctx, "/etc/logrotate.d/k3s", LOGROTATE_CONFIG);
+  const before = await ctx.ssh.exec(
+    "test -e /etc/logrotate.d/k3s -o -d /var/log/kubernetes && echo present || echo absent",
+    sshOpts(ctx),
+  );
+  const wasPresent = before.stdout.trim() === "present";
+
+  await ctx.ssh.exec(REMOVE_LOGROTATE, sshOpts(ctx));
+  await ctx.ssh.exec(REAP_OLD_AUDIT_FILES, sshOpts(ctx));
 
   return {
     success: true,
-    changed,
+    changed: wasPresent,
-    message: changed ? "Log rotation configured" : "Log rotation already configured",
+    message: wasPresent
+      ? "Removed legacy file-based audit logging (now via journald)"
+      : "No legacy audit log artifacts present",
   };
 };

bastion/src/modules/modules/k3s/tests/install.test.ts

@@ -71,9 +71,14 @@ describe("k3s install script — server role", () => {
     expect(script).toContain("enable-admission-plugins=NodeRestriction,PodSecurity");
   });
 
-  it("configures audit logging", () => {
+  it("configures audit logging via journald (stdout)", () => {
-    expect(script).toContain("audit-log-path=/var/log/kubernetes/audit.log");
+    expect(script).toContain("audit-log-path=-");
-    expect(script).toContain("audit-log-maxage=30");
+    // file-based fields and the now-obsolete log directory must be gone
+    expect(script).not.toContain("/var/log/kubernetes/audit.log");
+    expect(script).not.toContain("audit-log-maxage");
+    expect(script).not.toContain("audit-log-maxbackup");
+    expect(script).not.toContain("audit-log-maxsize");
+    expect(script).not.toContain("mkdir -p /var/log/kubernetes");
   });
 
   it("cleans stale flannel vxlan before Cilium install", () => {

bastion/src/modules/modules/k3s/tests/operations.test.ts

@@ -348,3 +348,143 @@ describe("applyPodSecurityStandards", () => {
     expectCommand(ctx.ssh, "pod-security.kubernetes.io/audit=restricted");
   });
 });
+
+// --- Audit Logging Decommission (file-based → journald) ---
+
+import { configureLogRotation } from "../src/operations/log-rotation.js";
+import { configureJournaldLimits } from "../src/operations/journald-limits.js";
+
+describe("configureLogRotation (decommission file-based audit logs)", () => {
+  it("removes the legacy logrotate rule and reaps obsolete audit files", async () => {
+    const ctx = mockCtx();
+    ctx.ssh.exec.mockResolvedValueOnce(stdout("present")); // probe: legacy artifacts exist
+    ctx.ssh.exec.mockResolvedValue(OK);
+
+    const result = await configureLogRotation(ctx);
+
+    expect(result.success).toBe(true);
+    expect(result.changed).toBe(true);
+    expectCommand(ctx.ssh, "rm -f /etc/logrotate.d/k3s");
+    expectCommand(ctx.ssh, /find \/var\/log\/kubernetes.*audit.*-delete/);
+    expectCommand(ctx.ssh, "rmdir /var/log/kubernetes");
+  });
+
+  it("is a no-op when nothing legacy is present", async () => {
+    const ctx = mockCtx();
+    ctx.ssh.exec.mockResolvedValueOnce(stdout("absent"));
+    ctx.ssh.exec.mockResolvedValue(OK);
+
+    const result = await configureLogRotation(ctx);
+    expect(result.success).toBe(true);
+    expect(result.changed).toBe(false);
+  });
+});
+
+describe("configureJournaldLimits", () => {
+  it("writes a 2 GB SystemMaxUse drop-in and reloads journald when changed", async () => {
+    const ctx = mockCtx();
+    ctx.ssh.exec.mockResolvedValueOnce(stdout("__LABCTL_NOT_FOUND__")); // no existing drop-in
+    ctx.ssh.exec.mockResolvedValue(OK);
+
+    const result = await configureJournaldLimits(ctx);
+
+    expect(result.success).toBe(true);
+    expect(result.changed).toBe(true);
+    const writeCall = ctx.ssh.exec.mock.calls.find((c) => {
+      const cmd = c[0] as string;
+      return cmd.includes("10-k3s-audit-cap.conf") && cmd.includes("LABCTL_EOF");
+    });
+    expect(writeCall).toBeTruthy();
+    const written = writeCall?.[0] as string;
+    expect(written).toContain("SystemMaxUse=2G");
+    expect(written).toContain("SystemKeepFree=1G");
+    expectCommand(ctx.ssh, "systemctl restart systemd-journald");
+  });
+
+  it("does not restart journald when the drop-in is already correct", async () => {
+    const ctx = mockCtx();
+    const existing =
+      "[Journal]\nSystemMaxUse=2G\nSystemKeepFree=1G\nSystemMaxFileSize=200M\n";
+    ctx.ssh.exec.mockResolvedValueOnce(stdout(existing));
+    ctx.ssh.exec.mockResolvedValue(OK);
+
+    const result = await configureJournaldLimits(ctx);
+
+    expect(result.success).toBe(true);
+    expect(result.changed).toBe(false);
+    expectNoCommand(ctx.ssh, "systemctl restart systemd-journald");
+  });
+});
+
+// --- Etcd Recovery ---
+
+import { recoverEtcdMember } from "../src/operations/etcd-recover.js";
+import { mockSsh } from "./helpers.js";
+
+describe("recoverEtcdMember", () => {
+  it("refuses to operate when cluster is below 3 members (quorum risk)", async () => {
+    const broken = mockSsh();
+    const peer = mockSsh();
+    peer.exec.mockResolvedValueOnce(stdout("/usr/bin/etcdctl")); // etcdctl present
+    peer.exec.mockResolvedValueOnce(stdout(
+      "111, started, host-a-aaa, https://10.0.0.1:2380, https://10.0.0.1:2379, false\n" +
+      "222, started, host-b-bbb, https://10.0.0.2:2380, https://10.0.0.2:2379, false",
+    ));
+
+    const result = await recoverEtcdMember({ broken, peer, brokenHostname: "host-b" });
+
+    expect(result.success).toBe(false);
+    expect(result.message).toMatch(/quorum/i);
+    // Critically: must NOT have stopped k3s or removed anything
+    expect(broken.exec).not.toHaveBeenCalledWith(expect.stringContaining("systemctl stop k3s"), expect.anything());
+  });
+
+  it("performs full procedure when quorum is preserved", async () => {
+    const broken = mockSsh();
+    const peer = mockSsh();
+    // ensureEtcdctl: present
+    peer.exec.mockResolvedValueOnce(stdout("/usr/bin/etcdctl"));
+    // member list (3 members, target = host-b)
+    peer.exec.mockResolvedValueOnce(stdout(
+      "111, started, host-a-aaa, https://10.0.0.1:2380, https://10.0.0.1:2379, false\n" +
+      "222, started, host-b-bbb, https://10.0.0.2:2380, https://10.0.0.2:2379, false\n" +
+      "333, started, host-c-ccc, https://10.0.0.3:2380, https://10.0.0.3:2379, false",
+    ));
+    // member remove
+    peer.exec.mockResolvedValueOnce(stdout("Member 222 removed"));
+    // post-rejoin member list — new id 444 for host-b
+    peer.exec.mockResolvedValueOnce(stdout(
+      "111, started, host-a-aaa, https://10.0.0.1:2380, https://10.0.0.1:2379, false\n" +
+      "333, started, host-c-ccc, https://10.0.0.3:2380, https://10.0.0.3:2379, false\n" +
+      "444, started, host-b-zzz, https://10.0.0.2:2380, https://10.0.0.2:2379, false",
+    ));
+
+    const result = await recoverEtcdMember({ broken, peer, brokenHostname: "host-b" });
+
+    expect(result.success).toBe(true);
+    expect(result.removedMemberId).toBe("222");
+    expect(result.newMemberId).toBe("444");
+    expectCommand(broken,"systemctl stop k3s");
+    expectCommand(peer,"member remove 222");
+    expectCommand(broken,/db\.corrupt-/);
+    expectCommand(broken,/rm -rf .*\/server\/tls/);
+    expectCommand(broken,"systemctl start k3s");
+  });
+
+  it("fails clearly when no member matches the broken hostname", async () => {
+    const broken = mockSsh();
+    const peer = mockSsh();
+    peer.exec.mockResolvedValueOnce(stdout("/usr/bin/etcdctl"));
+    peer.exec.mockResolvedValueOnce(stdout(
+      "111, started, host-a-aaa, https://10.0.0.1:2380, https://10.0.0.1:2379, false\n" +
+      "222, started, host-b-bbb, https://10.0.0.2:2380, https://10.0.0.2:2379, false\n" +
+      "333, started, host-c-ccc, https://10.0.0.3:2380, https://10.0.0.3:2379, false",
+    ));
+
+    const result = await recoverEtcdMember({ broken, peer, brokenHostname: "host-d" });
+
+    expect(result.success).toBe(false);
+    expect(result.message).toMatch(/No etcd member found/);
+    expect(broken.exec).not.toHaveBeenCalledWith(expect.stringContaining("systemctl stop k3s"), expect.anything());
+  });
+});