fix: make online updates recoverable

server/cli/update.ts

@@ -163,7 +163,10 @@ function writeJob(sqlite: Database.Database | undefined, jobId: string, values:
       requested_at=COALESCE(excluded.requested_at, update_jobs.requested_at),
       started_at=COALESCE(excluded.started_at, update_jobs.started_at),
       operation=excluded.operation,
-      status=CASE WHEN update_jobs.status='cancelled' THEN update_jobs.status ELSE excluded.status END,
+      -- Terminal rows are immutable from the runner's ordinary progress
+      -- writes. In particular, a stale/replayed request must not resurrect a
+      -- failed job as queued/downloading/etc.
+      status=CASE WHEN update_jobs.status IN ('cancelled', 'failed', 'completed') THEN update_jobs.status ELSE excluded.status END,
       version=excluded.version, platform=excluded.platform,
       release_url=COALESCE(excluded.release_url, update_jobs.release_url),
       asset_name=COALESCE(excluded.asset_name, update_jobs.asset_name),
@@ -176,6 +179,11 @@ function writeJob(sqlite: Database.Database | undefined, jobId: string, values:
       error_message=COALESCE(excluded.error_message, update_jobs.error_message),
       updated_at=excluded.updated_at,
       completed_at=COALESCE(excluded.completed_at, update_jobs.completed_at)
+    -- Do not let a delayed runner replay overwrite any field on a terminal
+    -- row. The predicate is part of the same SQLite upsert, so a finalizer
+    -- racing this write still wins atomically instead of leaving a partially
+    -- mutated completed/failed/cancelled record.
+    WHERE update_jobs.status NOT IN ('cancelled', 'failed', 'completed')
   `).run(
     jobId,
     values.adminId ?? null,
@@ -230,6 +238,7 @@ async function resolveRelease(options: UpdateRunOptions, platform: ReturnType<ty
       allowedHosts: options.allowedHosts ?? [],
       baseUrl: metadataUrl.toString(),
       maxBytes: options.maxBytes ?? 512 * 1024 * 1024,
+      timeoutMs: options.timeoutMs,
       publicKey: options.publicKey,
       requireSignature: options.requireSignature,
     });
@@ -398,19 +407,28 @@ export function finalizeUpdateJob(
   status: "completed" | "failed",
   message?: string,
 ): void {
-  const row = sqlite.prepare(`
-    SELECT id, status, version, platform, admin_id AS adminId,
-      request_id AS requestId, session_hash AS sessionHash
-    FROM update_jobs WHERE id=?
-  `).get(jobId) as { id: string; status: UpdateJobStatus; version: string; platform: string; adminId: string | null; requestId: string | null; sessionHash: string | null } | undefined;
-  if (!row) throw new Error("更新任务不存在");
-  const canComplete = row.status === "applying" || row.status === "completed";
-  const canFail = ACTIVE_UPDATE_STATUSES.includes(row.status) || row.status === "completed" || row.status === "failed";
-  if (status === "completed" ? !canComplete : !canFail) throw new Error("更新任务状态不允许完成");
-  const now = Date.now();
-  const safeFailureMessage = status === "failed" ? "新版本健康检查失败，已恢复上一版本" : null;
   sqlite.transaction(() => {
-    sqlite.prepare("UPDATE update_jobs SET status=?, error_message=?, completed_at=?, updated_at=? WHERE id=?").run(status, safeFailureMessage, now, now, jobId);
+    const row = sqlite.prepare(`
+      SELECT id, status, version, platform, admin_id AS adminId,
+        request_id AS requestId, session_hash AS sessionHash
+      FROM update_jobs WHERE id=?
+    `).get(jobId) as { id: string; status: UpdateJobStatus; version: string; platform: string; adminId: string | null; requestId: string | null; sessionHash: string | null } | undefined;
+    if (!row) throw new Error("更新任务不存在");
+    // A failed finalization can be retried by the runner. Once it has been
+    // committed, make retries a no-op so the error and audit trail stay stable.
+    if (row.status === status) return;
+    // A completed release is terminal. A delayed recovery process must never
+    // be able to downgrade it to failed after the service was healthy.
+    if (row.status === "completed" && status === "failed") throw new Error("更新任务状态不允许完成");
+    const canComplete = row.status === "applying" || row.status === "completed";
+    const canFail = ACTIVE_UPDATE_STATUSES.includes(row.status) || row.status === "completed" || row.status === "failed";
+    if (status === "completed" ? !canComplete : !canFail) throw new Error("更新任务状态不允许完成");
+    const now = Date.now();
+    const safeFailureMessage = status === "failed"
+      ? (message?.trim() ? safeErrorMessage(new Error(message)) : "新版本健康检查失败，已恢复上一版本")
+      : null;
+    const result = sqlite.prepare("UPDATE update_jobs SET status=?, error_message=?, completed_at=?, updated_at=? WHERE id=? AND status=?").run(status, safeFailureMessage, now, now, jobId, row.status);
+    if (result.changes !== 1) return;
     writeAudit(sqlite, {
       requestId: row.requestId || randomUUID(),
       actorAdminId: row.adminId,
@@ -569,6 +587,7 @@ export async function main(config: AppConfig = loadConfig()): Promise<void> {
       ...(dataBackupArchive ? { dataBackupArchivePath: dataBackupArchive, dataBackupSource: config.dataDir } : {}),
       ...((arg("--backup-dir")) ? { backupDir: arg("--backup-dir") } : {}),
       allowedHosts: allowedHosts.length ? allowedHosts : config.updateAllowedHosts,
+      timeoutMs: config.updateTimeoutMs,
       maxBytes: config.updateMaxBytes,
       dataBackupMaxBytes: config.maxTotalBytes,
       currentVersion: config.appVersion,