Home answers under its budget, cached as docs/CACHING.md says #3157

Home's KPI row and pinned panels now come from a cache, read only after the viewer is confirmed as an admin. bun test passes on the files I ran (home-cache, snappy, board-pins, home-kpi*, caching-doc, counts), and tsc --noEmit is clean. I did not run the whole suite.

What changed

• A new src/web/studio/home_cache.ts holds both caches. kpiRow reads the figures through cachedKpis. pinnedPanels runs each panel's query through cachedPanel, passed into drawPanel as an optional run argument. Only data is cached; the page is still rendered for each request.
• Invalidation: entries are stamped with writeStamp, the same write clock the feed cache uses, rather than the forge-wide MAX(event.seq) the doc names.

- Reason: the figures read run, pull, usage and page_vital rows, and some of those writes append no event. - Nothing is cached inside a transaction, and an entry lives at most 30 s. - A panel's key is its board slug plus its own definition, so an edited board is a new key. The doc's boardOid isn't available at that point. - I added a paragraph to docs/CACHING.md recording both differences.

• Tests: test/home-cache.test.ts checks three things.

- A repeat request is all hits, and one write causes exactly one KPI miss and one panel miss. - Home warmed by the owner shows no KPIs or pinned panels to a member or an anonymous visitor. - Warm home p95 stays under the 200 ms floor the other tests use. The doc's 50 ms comes from the bench.

Numbers (bench/snappy.ts --n 50, / as owner, before → after)

The fixture pins no panel, so the gain is the KPI queries only. The cached 15.6 statements per request is the part this change removed. With pins, a hit saves one query per panel; I did not measure that.

Left out

• No per-repository or event-sequence invalidation for these two caches. Any write drops them, which costs hit rate, not correctness.
• The feed and the "On you" card are untouched.
• The doc's / p95 ≤ 50 ms is met only on the small fixture (about 20 ms). I did not run the 69-repository bench/map-home.ts.

To try it: pin a panel from /boards and reload / as the admin. The second load costs no KPI or panel queries. Load / as a member and neither appears.

Closes #2717

Home's KPI row and pinned panels now come from a cache, read only after the viewer is confirmed as an admin. bun test passes on the files I ran (home-cache, snappy, board-pins, home-kpi*, caching-doc, counts), and tsc --noEmit is clean. I did not run the whole suite.

What changed

• A new src/web/studio/home_cache.ts holds both caches. kpiRow reads the figures through cachedKpis. pinnedPanels runs each panel's query through cachedPanel, passed into drawPanel as an optional run argument. Only data is cached; the page is still rendered for each request.
• Invalidation: entries are stamped with writeStamp, the same write clock the feed cache uses, rather than the forge-wide MAX(event.seq) the doc names.

- Reason: the figures read run, pull, usage and page_vital rows, and some of those writes append no event. - Nothing is cached inside a transaction, and an entry lives at most 30 s. - A panel's key is its board slug plus its own definition, so an edited board is a new key. The doc's boardOid isn't available at that point. - I added a paragraph to docs/CACHING.md recording both differences.

• Tests: test/home-cache.test.ts checks three things.

- A repeat request is all hits, and one write causes exactly one KPI miss and one panel miss. - Home warmed by the owner shows no KPIs or pinned panels to a member or an anonymous visitor. - Warm home p95 stays under the 200 ms floor the other tests use. The doc's 50 ms comes from the bench.

Numbers (bench/snappy.ts --n 50, / as owner, before → after)

The fixture pins no panel, so the gain is the KPI queries only. The cached 15.6 statements per request is the part this change removed. With pins, a hit saves one query per panel; I did not measure that.

Left out

• No per-repository or event-sequence invalidation for these two caches. Any write drops them, which costs hit rate, not correctness.
• The feed and the "On you" card are untouched.
• The doc's / p95 ≤ 50 ms is met only on the small fixture (about 20 ms). I did not run the 69-repository bench/map-home.ts.

To try it: pin a panel from /boards and reload / as the admin. The second load costs no KPI or panel queries. Load / as a member and neither appears.