console: keep the redesigned cluster page's live data in TanStack DB collections

[leedqin]

Motivation

Fifth PR in the cluster details redesign stack, on top of #39510. The resource usage charts and the memory by object card read their live data from TanStack DB collections instead of per-hook subscribe state, so switching windows reads data that's already loaded, and coming back to a cluster paints its last data at once.

Description

Charts. Each cluster gets a collection fed by the 3h or 24h utilization SUBSCRIBE, and each subscribe covers its view's full retention. Switching windows within a tier reads the collection instead of resubscribing. Leaving the Metrics tab closes the subscribe; coming back paints the held rows at once and shows the refreshing spinner until the new snapshot lands. If the held rows have all left the window, the card shows loading rather than "No resource usage". Windows past 24h still poll, through usePolledReplicaUtilizationHistory, which mounts no subscribe hooks and so opens no idle WebSocket. The page no longer calls useReplicaUtilizationHistory, so its includeOfflineEvents option goes away.

Memory by object. Each replica gets a collection fed by a SUBSCRIBE to mz_object_arrangement_sizes, an index lookup on replica_id. A live query joins it to allObjectsCollection for names and keeps the ten largest. The card waits for that collection, so objects never show as raw ids, and a replica without a heap limit drops the bar and lists each object's size. The replica id comes from the largest-replica query. Sizes update live, and the card no longer polls sizes every 60s or object names every 5 minutes. Environments before 26.35 keep the polled card.

Collections are keyed by environment and id, because cluster and replica ids repeat across environments, through one createScopedSubscribeCollections helper that drops an earlier environment's collections when the scope changes. They aren't persisted. A subscribe collection's status gains isRefreshing, set while held rows are served and no running subscribe has delivered a newer snapshot. Ending a collection's session sets it, so coming back before the next subscribe starts (for example while the cluster's lineage refetches) shows loading rather than an empty window.

Measured in headless Chrome against a dev server (React dev mode, so absolute times run high), median of 3 runs, time until the charts are drawn:

Step	Before	After
Open a cluster	594 ms	630 ms
Switch 1h to 3h	222 ms	25 ms
Switch back to 1h	238 ms	24 ms
Objects tab back to Metrics	338 ms	102 ms
Back to the cluster from the list	347 ms	95 ms

The card's first open doesn't get faster: it waits on the largest replica query, which this PR doesn't change. On mz_catalog_server the card's subscribe costs the same as the polls it replaces. Each held-open subscribe adds about 0.2 mCPU per viewer on environmentd for its progress updates.

Verification

• useLiveClusterUtilization.test.ts: 3h samples bin into the window's buckets, keeping each bucket's peak, and 24h buckets clip to the window in time order.
• useLiveLargestObjects.test.ts: the largest objects are named from the catalog, largest first; an object missing from the catalog is marked as an orphaned dataflow; only the ten largest are kept.
• MemoryByObject.test.tsx: on an environment that streams sizes, objects are named from the catalog; a replica without a heap limit shows no share; the card waits for the catalog before naming objects.
• ResourceUsage.test.tsx: a 1h window draws from the live subscribe.
• useLiveClusterUtilization.test.ts: coming back with held rows that have left the window reports loading and refreshing on every render until the new snapshot lands, including while an evicted lineage query refetches.
• useSubscribe.test.tsx: ending a collection session marks its rows as refreshing.
• subscribeCollection.test.ts: a scope change drops the earlier scope's collections.
• subscribeCollection.test.ts: held rows are marked refreshing until a new subscribe's snapshot lands.

🤖 Generated with Claude Code

[jdonelson]

heapLimit is Number(largestReplica.heapLimit), which is 0 when no replica has reported metrics yet (or on the emulator). Dividing by it makes the largest object fill the whole bar and the legend read "Infinity%". Returning a null percentage like the polled path avoids that, but then the bar is empty and formatMemoryUsage shows "-" for every object, hiding sizes we do have. Could the card drop the bar when the heap limit is unknown and list each object's size on its own?

[leedqin]

Good point, done in 40df6d7: without a heap limit the card drops the bar and lists each object's size, with "Size on " as the subtitle.

[linear-code]

CNS-203