perf: cache SSR data fetching to eliminate upstream pool fanout

[erubboli]

Problem

Homepage and /pools SSR took 12-18s TTFB. Each render issued ~177 upstream calls (measured via a request-counting proxy in front of the testnet api-web-server):

• fetchAllPoolsFromApi paged /pool?offset=N 10 items at a time → 34 sequential calls per crawl
• one /pool/{id}/delegations call per pool (155 pools) per crawl
• /pools crawled twice per render (summary + list each fetched independently)
• 10 sequential /block/{id} fetches for the recent-blocks widget
• in testnet mode (no DATABASE_URL) none of this was cached — every render paid the full cost

Fix

• src/lib/server-cache.ts (new): in-process TTL cache with in-flight de-duplication and stale-while-revalidate, persisted on globalThis so page bundles and route-handler bundles in the same server process share one store (same pattern as the existing pg pool singleton in lib/postgres.ts).
• src/lib/explorer-ssr.ts: single shared getPoolsSnapshot() (120s SWR) now feeds the homepage summary, /pools, /api/pool/list and /api/pool/summary; computed pool summary cached 30s, mempool 10s, transaction total 60s.
• src/lib/explorer-source.ts:
  • pool crawl uses items=100 pages fetched in adaptive parallel rounds with retries (upstream breaks for items > 100; verified)
  • recent blocks fetched in parallel via /chain/{height} id lookups + Promise.all block fetches (sequential previous-block walk kept as fallback for older API versions)
  • chain/tip cached 5s, recent transactions 10s, immutable block/height lookups 1h (in-memory + Next force-cache data cache)
• API routes /api/pool/list, /api/pool/summary, /api/block/tip: route through the shared cached loaders; response shapes unchanged. /api/block/tip stays fast under k8s probe hammering (in-flight dedup, no locks).

Compatibility

• Mainnet mode unchanged: DB-first logic preserved — with a cron-populated Postgres, verified 0 upstream pool calls end-to-end (worker sync + server against a scratch Postgres).
• Cache is per-pod in-memory — safe for multi-pod deployments; every pod independently converges within the TTL.
• No API response shapes changed.

Measurements (testnet API via local counting proxy)

	before	after
Homepage cold / warm TTFB	17.1s / 15.8s	4.5s / 0.015s
Upstream pool calls per render	34 + 155 delegations	0 warm; one snapshot build (3 + 155) per 120s TTL per pod
10 concurrent cold renders	~1,770 upstream calls	~161 calls, one shared build, all renders ~4.5s
/api/block/tip under 20-way hammering	1 upstream call per hit	5-46ms, 0 upstream within 5s TTL
Mainnet (DB populated)	DB-first	0 pool calls (verified)

Cold is bounded by one shared crawl (~4.5s through a latency-adding local proxy; lower against production). next build, tsc --noEmit and 14/14 jest tests pass (4 new tests cover the cache: dedup, TTL, SWR, error propagation).