Rates Engine v0.5.0-rc.67

[v0.5.0-rc.67] — 2026-05-21

Fixed

• /v1/assets/{id} cache TTL was shorter than the prewarm
  cadence — p95/p99 blew up (#52). The assetDetailResponseCache
  TTL was 30s but selfPrewarmAssetEndpoints only refreshes every
  60s, so the cache sat expired for 30 of every 60 seconds. The
  status page polls /v1/assets/native every 30s and kept landing
  in the cold window, paying the full ~700ms handler rebuild every
  time (inflated to ~2.8s under concurrent-backfill CPU
  contention). Every OTHER status-page endpoint measured 1-200ms —
  the entire API p95/p99 tail (and the three latency SLO alerts)
  was this one endpoint's cold misses. Bumped the TTL to 120s —
  one full prewarm interval of headroom, so a prewarm pass always
  refreshes an entry before it expires and native stays
  permanently warm. Matches the sibling F2-path caches (1–2 min
  TTL, same 60s prewarm). 120s staleness still fits the ADR-0015
  closed-bucket-only contract.

Changed

• Status page endpoint probe is now two-shot (#52). The
  per-endpoint latency matrix on the status page polls every
  30 s / 2 min; between polls Cloudflare lets the edge→origin
  connection pool go cold, so a single probe's first request paid
  a full CF↔origin TCP+TLS setup (~2-3 s measured) that has
  nothing to do with API latency — the API serves cached asset
  detail in <10 ms. The probe now fires a throwaway warm-up fetch
  first and measures the second request on the warm connection,
  i.e. the latency a returning user actually experiences. Both
  fetches keep cache: 'no-store' so neither the browser nor the
  CDN serves a stale body — it's still a real round trip, just not
  a cold-pool one. A non-2xx or thrown warm-up short-circuits
  (reports down/error without a second request). web/status
  frontend only — no API change.