The reconciler read `should` and `existing` in separate statements under READ
COMMITTED, and its queue one-consumer check wasn't locked — so two concurrent
reconcilers (e.g. two parallel app-creates) could (a) spuriously DELETE a valid
materialized cron copy when one committed a copy between the other's two reads
(a lost copy self-healing only on the next mutation), or (b) both pass the
queue slot check and double-fill a consumer slot. Take a fixed xact-scoped
advisory lock (REMATERIALIZE_LOCK_KEY) at the top of the reconcile so each pass
is atomic w.r.t. the others. Reconciles are infrequent + fast, so serializing
is cheap; the ON CONFLICT guard stays as defense in depth.
Found by the M5 review. Pinned by the existing stateful_templates test + the
full journey suite (which previously flaked on the race).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
05373814b2