Davy7/21/2026, 1:47:24 AM

Issue 1: The startup executor only replayed Publish operations, so every
Gc row scheduled after a successful publish was skipped forever. Add a
Gc dispatch branch in retry_pending_operations that calls execute_gc
with the provider and a dedicated control DB connection.

Issue 2: The credential refresh single-flight lock serialized
concurrent refreshes but did not collapse them — after acquiring the
lock, each waiter re-checked its own stale in-memory secret copy and
proceeded to fire a redundant network refresh. Reload the stored
credential from disk after acquiring the lock so the waiter picks up
the token that the first thread already refreshed and persisted.

Parent8675930