spec: throttle aggregate Redis fan-out to 12 concurrent

[Flotapponnier]

Root cause found

Homepage "Awaiting samples" persistent state: the aggregate loader fires Promise.allSettled(specs.map(loadOne)) = 68 read chains at once. Each chain does 2-3 Redis GETs (pointer → blob → optional LKG), so ~150-200 concurrent Redis ops per aggregate cycle. SRH's request pool saturates and the alphabetically-later specs (perp-open-interest → zksync-rpc, exactly 31 slugs) systematically time out at 8s and render as draft placeholders, even though their blobs are sitting in Redis the whole time (v=2, correct pointers, valid LKG).

Deterministic pattern: sorted slug list, cutoff at perp-funding-stability / perp-open-interest.

Fix

Wrap the fan-out in a small limitedAllSettled (12 concurrent workers, same result shape as Promise.allSettled). All 68 chains complete in ~4-5 s worst-case wall time instead of 37 completing in 200 ms + 31 timing out.

Behind the unstable_cache layer (revalidate 300 s) so the slightly-longer cache-miss wall time is invisible to end users. On cache hit (99% of requests) zero difference.

Follow-up work still open

• (later, architectural) worker → Vercel Blob broadcast, remove SRH from read path entirely
• (later) batch pointer reads via MGET so the 68 chains become 1 pipelined request
  Both would replace this throttle but aren't cheap.