spec: unstable_noStore() on draft-placeholder fallback (kill poisoned aggregate)

[Flotapponnier]

Why

Every SRH stuck-pool incident (twice in 8 days, PR #1190 + #1287) needs a manual cache-key bump + deploy to purge the poisoned aggregate cache. Root cause: when SRH goes stuck, loadAllBenchmarksCached throws AllBenchmarksDraftError, loadAllBenchmarksSafe catches and returns draft placeholders, homepage renders that as HTML, and Vercel Edge caches that HTML for revalidate=60 seconds — under stale-while-revalidate the poisoned HTML then sits for hours after backend recovery, since SWR only refreshes on cache miss.

Fix

When loadAllBenchmarksSafe falls back to placeholders, call unstable_noStore() first. That marks the render as no-store and instructs Vercel to skip every cache layer (ISR + CDN edge) for this response. Every subsequent hit re-renders server-side, and as soon as SRH is back the next hit produces real HTML immediately.

Effect

• Zero-manual-action recovery once SRH is healed again
• Kills the recurring pattern of "SRH stuck 2 min → homepage broken 2 hours + PR + deploy"
• No behavioral change on the happy path (only fires when the aggregate quorum guard trips)

Follow-up: separate PR will add a VPS cron auto-heal (recreate SRH if Redis client count < 10 for 60s) so the recovery path shortens further.

Cheap-tier roadmap

• noStore fallback (this PR)
• VPS cron auto-heal SRH stuck-pool
• (later, real refactor) worker → Vercel Blob broadcast, kill SRH from read path