Rates Engine v0.5.0-rc.52

[v0.5.0-rc.52] — 2026-05-15

Changed

• Status page: removed the "Backfill by decoder" panel. Each
  backfill restart (without -resume) creates fresh cursor rows
  keyed by chunk boundaries, so the panel accumulated "stalled"
  rows from every prior partial run — making the page look like
  data was unreliable even when later runs (or live ingest) had
  filled the same ledger ranges. The density_pct column on
  backfill_coverage already answers "do we have data here"
  honestly via interval union. Wire field retained; only the UI
  panel dropped.

• deploy.yml hardening. Two fixes after the rc.51 deploy
  failures: (1) the migration-staging step's ls … | head
  tripped pipefail via SIGPIPE even though staging succeeded —
  swapped for a SIGPIPE-safe sort | head + explicit count;
  (2) new migrations_skip workflow input so operators who have
  applied migrations out-of-band (or know the schema is unchanged)
  can deploy past the playbook's hardcoded passwordless DSN, which
  fails auth against the live db. Proper DSN-from-target-host fix
  is a follow-up.

Fixed

• Backfill-coverage snapshot was permanently "pending". The
  cache-refresh query (BackfillCoverageStats) ran a single
  SELECT ... GROUP BY source over the trades hypertable. Once
  trades grew past ~2700 chunks, that scan needed >2700 chunk
  AccessShareLocks in one transaction, overflowing
  max_locks_per_transaction (256 on r1) with
  out of shared memory. The error was masked by the API's 30s
  context timeout, so the symptom looked like a slow query rather
  than a hard limit.

  Rewritten to a per-source loop: each source's earliest/latest
  ledger via ORDER BY ts {ASC,DESC} LIMIT 1 (chunk-exclusion
  stops after the first/last chunk — ~3s vs ~68s for MIN()/MAX()
  which seek every per-chunk index), trade count approximated from
  the 24h source/total ratio scaled by approximate_row_count
  (precise per-source COUNT(*) is 2:34s on sdex). Each statement
  runs in its own implicit transaction so the per-transaction lock
  budget resets — the 2700-chunk overflow can't recur. Full refresh
  is now ~23s across all 13 on-chain sources.

  Cache-refresh timeout raised 30s → 2min (it's a background
  goroutine, never bounds an API request; 2min sits below the 5min
  refresh interval so refreshes don't stack).

• /v1/pools?order_by=pair returned 500 on every request. The
  pair-ordered SQL branch in buildPoolsQuery was missing the
  filter.Asset arg in its args slice — postgres returned
  pq: got 6 parameters but the statement requires 7 because the
  CTE references $7. The volume-desc branch already had the
  correct 7-arg slice. Caught live on r1 2026-05-14.

Added

• CORS Allow-Credentials: true opt-in for cookie-bearing
  cross-origin fetches. New [api].allow_credentials config flag
  (default false). Required for the magic-link session on
  /v1/account/me + /v1/account/keys to actually work from a
  cross-origin browser SPA — pre-fix the preflight emitted
  Access-Control-Allow-Origin but no Access-Control-Allow-Credentials,
  so browsers stripped cookies. The middleware now panics at boot
  when both allowed_origins=["*"] and allow_credentials=true
  are set, since browsers reject that combo at the parser.