Web viewer: every server-tab switch re-downloads the whole fleet roll-up (/api/fleet, 66–81 KB) to render one server's header card #4190

Problem

The server page's header card (band, CPU, memory, blocking, and so on) comes from loadServerCard(). That calls apiGetFleet() and searches the whole roll-up for the one matching card. The server page re-renders on every sub-tab switch and on every 60 s poll, so each tab click fetches the entire fleet roll-up again. apiGetFleet only shares a request between callers that are in flight at the same moment (#3895). Nothing survives across navigation.

Measured

Production stores, build 3.8.0-nightly.20260924.467, 2026-09-25. Browser timing on one server's 12 SQL Server tabs and one PostgreSQL target's 8 tabs:

• Every one of the 20 tab opens included GET /api/fleet: 65.6 KB on a 43-server store and 78.9 KB on a 50-target store.
• On an otherwise idle tab it took 0.12–0.33 s.
• When it ran beside a heavy tab's own reads it was one of the slowest calls on the page: 0.94 s on I/O, 1.0 s on Config, 1.8–2.1 s on Queries. That's the header card waiting behind the store's widest read.
• It's 66–81 KB per tab click for one card of roughly 1.5 KB. Over a 12-tab walk that's ~0.8 MB of fleet data, most of it other servers'.

Where

• Darling/PerformanceMonitor.Darling.Service/wwwroot/js/pages/server.js, loadServerCard() ~:199–215: const res = await apiGetFleet(); then res.data.cards.find(...).
• wwwroot/js/util.js, apiGetFleet() ~:366+: joins an in-flight request only. By design nothing outlives the response.

Fix shape

Any one of these, cheapest first:

1. Reuse within the page. Cache the server's card in the server page's module state for the life of the route and its sub-tab switches. Refresh it on the 60 s poll, not on tab clicks.
2. Keep the last fleet response in the shell with its generated_at, and let loadServerCard use it when it's newer than the poll interval. The sidebar already fetched it seconds earlier.
3. A per-server card endpoint, GET /api/fleet/card?server=…, that computes one card instead of 43–50. That also helps saved views that show one server.

Pin: a web test that switching between two tabs of one server issues no /api/fleet request (with 1 or 2).