diff --git a/app/pages/index.vue b/app/pages/index.vue
index beaf516a..ce72a3de 100644
--- a/app/pages/index.vue
+++ b/app/pages/index.vue
@@ -87,7 +87,7 @@ const [
   })),
   // Server-rendered rather than lazy: the section is one of the few places on
   // the homepage whose content changes hourly, and the endpoint is cached at
-  // the edge for 5 minutes, so it costs a cache read rather than a query.
+  // the edge for 30 minutes, so it costs a cache read rather than a query.
   withHomeDataTiming('home-trending', useFetch('/api/feed/trending', {
     key: 'home-trending-v3',
     // The hero texture prints Skill names from every repo it gets, so it asks
diff --git a/layers/marketing/app/pages/skills/index.vue b/layers/marketing/app/pages/skills/index.vue
index 965283dd..17a9de05 100644
--- a/layers/marketing/app/pages/skills/index.vue
+++ b/layers/marketing/app/pages/skills/index.vue
@@ -114,7 +114,7 @@ const tagBySlug = computed(() => {
  * it must never hold up the registry render or fail it. If the feed is slow or
  * errors, rows draw without flames and nothing else changes.
  *
- * The endpoint caches for 300s, so this costs a shared read rather than a
+ * The endpoint caches for 1800s, so this costs a shared read rather than a
  * per-visitor query on the site's highest-traffic surface.
  */
 const { data: trendingFeed } = useFetch('/api/feed/trending', {
diff --git a/layers/registry/server/api/v1/trending.get.ts b/layers/registry/server/api/v1/trending.get.ts
index 4ec74cca..77d6ef50 100644
--- a/layers/registry/server/api/v1/trending.get.ts
+++ b/layers/registry/server/api/v1/trending.get.ts
@@ -7,7 +7,7 @@ import { findSkillsByKeys } from '../../utils/skills-registry'
 
 /**
  * Reads the trending feed in process with the exact query the trending page
- * sends, so v1 ranks the same board from the same five-minute cache, then
+ * sends, so v1 ranks the same board from the same 30-minute cache, then
  * slices it to `limit`.
  */
 export default defineApiOperation({
diff --git a/nuxt.config.ts b/nuxt.config.ts
index d1fad7c4..22cdda55 100644
--- a/nuxt.config.ts
+++ b/nuxt.config.ts
@@ -596,7 +596,7 @@ export default defineNuxtConfig({
     // named in Vary. The cache key includes the query string, so each `?range=`
     // board is its own entry. Browsers still get no-store and ask the edge.
     //
-    // 60s fresh: the board's feed already caches for 300s, so this adds at most
+    // 60s fresh: the board's feed already caches for 1800s, so this adds at most
     // a minute. 3600s stale: the board's inputs move every 15 minutes at most
     // (X mentions) and hourly (social mentions, engagement), so a quiet colo
     // may serve one copy up to an hour old while it refreshes in the
diff --git a/server/api/feed/trending.get.ts b/server/api/feed/trending.get.ts
index a9c1739a..6c9f018f 100644
--- a/server/api/feed/trending.get.ts
+++ b/server/api/feed/trending.get.ts
@@ -7,6 +7,7 @@
  * catalog. The admin review surface reads the same loader without the filter.
  */
 
+import type { FeedCacheWindow } from '#server/utils/feed-cache'
 import type { TrendingBoard } from '#shared/server/trending-board'
 import type { FallbackSkill } from '#shared/server/trending-fallback'
 import type { TrendingRepo } from '#shared/server/trending-repos'
@@ -168,6 +169,16 @@ export interface TrendingFeedResponse {
   computedAt: number
 }
 
+/**
+ * The board's fastest input is an X post, and those land every 15 minutes at
+ * most (`sync-x-mentions` runs on a 15-minute cron). A five-minute recompute
+ * therefore reran the full post-window scan three times per cadence that
+ * could change the answer, measured at 3,196 executions and 112.9 million
+ * row reads in one week. Half an hour sits a full cadence above ingest;
+ * `staleTtl` still keeps the last good board through a bounded D1 outage.
+ */
+const TRENDING_FEED_WINDOW: FeedCacheWindow = { ttl: 1800, staleTtl: 3600 }
+
 function toSkillItem(entry: TrendingSkill, starSeries: StarPoint[]): TrendingSkillFeedItem {
   return {
     owner: entry.owner,
@@ -302,6 +313,6 @@ export default defineEventHandler(
       const fallback = fallbackSkills.map(entry => toFallbackItem(entry, seriesOf(entry)))
 
       return { items, namedSkills: skillItems, fallback, computedAt: now }
-    }, [limit, windowHours])
+    }, [limit, windowHours], TRENDING_FEED_WINDOW)
   },
 )
diff --git a/test/unit/feed-cache-resilience.test.ts b/test/unit/feed-cache-resilience.test.ts
index f2a96a73..b50d0756 100644
--- a/test/unit/feed-cache-resilience.test.ts
+++ b/test/unit/feed-cache-resilience.test.ts
@@ -47,12 +47,15 @@ function request(route: string, query = '', deployment = 'release-a'): H3Event {
 }
 
 const routes = [
-  { route: 'recent-updates', load: () => import('../../server/api/feed/recent-updates.get') },
-  { route: 'recent-publishes', load: () => import('../../server/api/feed/recent-publishes.get') },
-  { route: 'trending', load: () => import('../../server/api/feed/trending.get') },
+  // `staleMs` is one second past the feed's fresh window; `deadMs` is one
+  // second past fresh + stale. Trending rides a 30-minute window (X posts
+  // land every 15 minutes at most), the others the 300 s default.
+  { route: 'recent-updates', load: () => import('../../server/api/feed/recent-updates.get'), staleMs: 301_000, deadMs: 3_901_000 },
+  { route: 'recent-publishes', load: () => import('../../server/api/feed/recent-publishes.get'), staleMs: 301_000, deadMs: 3_901_000 },
+  { route: 'trending', load: () => import('../../server/api/feed/trending.get'), staleMs: 1_801_000, deadMs: 5_401_000 },
 ]
 
-describe.each(routes)('$route feed cache', ({ route, load }) => {
+describe.each(routes)('$route feed cache', ({ route, load, staleMs, deadMs }) => {
   it('serves its last good response when an expired entry cannot refresh', async () => {
     const handler = (await load()).default
     const fresh = await handler(request(route))
@@ -61,9 +64,9 @@ describe.each(routes)('$route feed cache', ({ route, load }) => {
     boundary.board.mockRejectedValue(new Error('D1 DB is overloaded'))
 
     await expect(handler(request(route))).resolves.toEqual(fresh)
-    vi.advanceTimersByTime(301_000)
+    vi.advanceTimersByTime(staleMs)
     await expect(handler(request(route))).resolves.toEqual(fresh)
-    vi.advanceTimersByTime(3600_000)
+    vi.advanceTimersByTime(deadMs - staleMs)
     await expect(handler(request(route))).rejects.toThrow('D1 DB is overloaded')
   })
 
@@ -73,17 +76,28 @@ describe.each(routes)('$route feed cache', ({ route, load }) => {
     boundary.all.mockRejectedValue(new Error('D1 DB is overloaded'))
     boundary.board.mockRejectedValue(new Error('D1 DB is overloaded'))
 
-    vi.advanceTimersByTime(301_000)
+    vi.advanceTimersByTime(staleMs)
     if (route === 'trending')
       await expect(handler(request(route, '?window=24'))).rejects.toThrow('D1 DB is overloaded')
     else
       await expect(handler(request(route, '?ignored=24'))).resolves.toEqual({ items: [] })
     await expect(handler(request(route, '', 'release-b'))).rejects.toThrow('D1 DB is overloaded')
-    vi.advanceTimersByTime(3600_000)
+    vi.advanceTimersByTime(deadMs - staleMs)
     await expect(handler(request(route))).rejects.toThrow('D1 DB is overloaded')
   })
 })
 
+it('trending serves one compute across a 15-minute ingest cadence', async () => {
+  // X posts land every 15 minutes (`sync-x-mentions` runs `*/15 * * * *`), so
+  // the board's inputs cannot change inside this gap. A second compute there
+  // only repeats the full post-window scan.
+  const handler = (await import('../../server/api/feed/trending.get')).default
+  const first = await handler(request('trending'))
+  vi.advanceTimersByTime(900_000)
+  await expect(handler(request('trending'))).resolves.toEqual(first)
+  expect(boundary.board).toHaveBeenCalledTimes(1)
+})
+
 it('shares equivalent trending options without sharing a different window', async () => {
   const handler = (await import('../../server/api/feed/trending.get')).default
   const fresh = await handler(request('trending', '?limit=500&window=900'))