diff --git a/apps/server/src/modules/search/releases.ts b/apps/server/src/modules/search/releases.ts
index 220019f7..5d4d2769 100644
--- a/apps/server/src/modules/search/releases.ts
+++ b/apps/server/src/modules/search/releases.ts
@@ -79,13 +79,12 @@ export class ReleaseFanout {
 
   /**
    * 在给定站点上搜一组查询词:每站每词一次请求(同站多个词的结果按站点资源 ID 去重)。
-   * 缓存按(站点, 查询词):失败 / 超时也缓存,但一轮里全部站点都失败时不缓存(下次直接重试)。
+   * 缓存按(站点, 查询词):只缓存成功的结果;失败 / 超时 / 验证拦截 / 跳过不写缓存,下次搜索直接重试。
    */
   async search(terms: string[], sites: SiteRow[], opts: { refresh?: boolean } = {}): Promise<FanoutHit> {
     const words = [...new Map(terms.map((t) => [cacheKey(t), t.trim()])).entries()].filter(([k]) => k);
     let cached = true;
     let coalesced = false;
-    const fresh: string[] = [];
     const perSite = await mapLimit(sites, Math.max(1, this.opts.concurrency), async (site) => {
       const parts: SiteResult[] = [];
       for (const [k, term] of words) {
@@ -105,16 +104,11 @@ export class ReleaseFanout {
         const p = this.searchSite(this.rt(), site, term).finally(() => this.inflight.delete(key));
         this.inflight.set(key, p);
         const r = await p;
-        this.cache.set(key, { expires: this.opts.now() + this.opts.ttlMs, result: r });
-        fresh.push(key);
+        if (r.coverage.status === "ok") this.cache.set(key, { expires: this.opts.now() + this.opts.ttlMs, result: r });
         parts.push(r);
       }
       return mergeSite(site, parts);
     });
-    // 本轮新查的站点全部失败:撤掉这些缓存
-    const freshResults = fresh.map((k) => this.cache.get(k)?.result).filter((r): r is SiteResult => !!r);
-    if (freshResults.length && freshResults.every((r) => r.coverage.status !== "ok"))
-      for (const k of fresh) this.cache.delete(k);
     this.prune();
     const fetched = perSite.map((r) => r.fetchedAt);
     return {
diff --git a/apps/server/test/search/search.test.ts b/apps/server/test/search/search.test.ts
index 45c449ef..efadc6af 100644
--- a/apps/server/test/search/search.test.ts
+++ b/apps/server/test/search/search.test.ts
@@ -289,13 +289,20 @@ describe.skipIf(!up)("全局搜索", () => {
     expect(rows.length).toBe(3);
   });
 
-  it("30s 内相同查询命中缓存,不再请求站点", async () => {
+  it("30s 内相同查询:成功的站点命中缓存不再请求;超时的站点不缓存,每次重试", async () => {
     const before = [nexusSearches(), mteamSearches(), slowCalls];
     const r = await req("/search?q=%20dune%20&scope=releases");
-    expect(r.body.data.releases).toMatchObject({ cached: true, total: 3, partial: true });
+    // 超时的站点重新查了一次,所以整体不算缓存结果
+    expect(r.body.data.releases).toMatchObject({ cached: false, total: 3, partial: true });
+    expect([nexusSearches(), mteamSearches()]).toEqual(before.slice(0, 2));
+    expect(slowCalls).toBe(before[2]! + 1);
     const page = await req("/search/releases?q=Dune");
-    expect(page.body.data).toMatchObject({ cached: true, total: 3 });
-    expect([nexusSearches(), mteamSearches(), slowCalls]).toEqual(before);
+    expect(page.body.data).toMatchObject({ cached: false, total: 3, partial: true });
+    expect([nexusSearches(), mteamSearches()]).toEqual(before.slice(0, 2));
+    expect(slowCalls).toBe(before[2]! + 2);
+    // 只查成功过的站点:全部命中缓存
+    const okOnly = await req(`/search/releases?q=Dune&sites=${sites.nexus},${sites.mteam}`);
+    expect(okOnly.body.data).toMatchObject({ cached: true, total: 3, partial: false });
     // refresh=true 跳过缓存
     const fresh = await req("/search/releases?q=Dune&refresh=true");
     expect(fresh.body.data.cached).toBe(false);
@@ -387,7 +394,9 @@ describe.skipIf(!up)("全局搜索", () => {
   });
 
   it("筛选 / 排序 / 游标 / 分面", async () => {
-    const q = (s: string) => req(`/search/releases?q=Dune&${s}`).then((r) => r.body.data);
+    // 只查成功的两站(超时站点不缓存,每次重试会拖慢用例)
+    const ok = `sites=${sites.nexus},${sites.mteam}`;
+    const q = (s: string) => req(`/search/releases?q=Dune&${ok}&${s}`).then((r) => r.body.data);
     let d = await q("resolution=2160p");
     expect(d.items.map((i: Loose) => i.siteReleaseId).sort()).toEqual(["812345", "990001"]);
     expect(d).toMatchObject({ total: 2, unfilteredTotal: 3 });
@@ -399,7 +408,7 @@ describe.skipIf(!up)("全局搜索", () => {
 
     d = await q("source=web-dl,REMUX&free=false");
     expect(d.items.map((i: Loose) => i.siteReleaseId).sort()).toEqual(["900001", "990001"]);
-    d = await q(`sites=${sites.mteam}`);
+    d = (await req(`/search/releases?q=Dune&sites=${sites.mteam}`)).body.data;
     expect(d.items.map((i: Loose) => i.siteReleaseId)).toEqual(["990001"]);
     d = await q(`minSize=${10 * 1024 ** 3}&maxSize=${50 * 1024 ** 3}`);
     expect(d.items.map((i: Loose) => i.siteReleaseId)).toEqual(["990001"]);
@@ -420,8 +429,9 @@ describe.skipIf(!up)("全局搜索", () => {
   });
 
   it("快速筛选 qf(SEARCH-3):分面计数、同维度 OR / 跨维度 AND、分页一致", async () => {
+    const ok = `sites=${sites.nexus},${sites.mteam}`;
     const q = (qf: string, extra = "") =>
-      req(`/search/releases?q=Dune&qf=${encodeURIComponent(qf)}${extra}`).then((r) => r.body.data);
+      req(`/search/releases?q=Dune&${ok}&qf=${encodeURIComponent(qf)}${extra}`).then((r) => r.body.data);
     const counts = (opts: Loose[]) => Object.fromEntries(opts.map((o) => [o.value, o.count]));
     const ids = (d: Loose) => d.items.map((i: Loose) => i.siteReleaseId).sort();