0.8.3

Added

• PostgreSQL: PostgresEventStore::pool_churn returns the new #[non_exhaustive], serializable
  PoolChurn: the connections the handle's pool has opened, the ones it retired by cause (a lease
  dropped without settling, a client already closed, shutdown) and the connections that replaced a
  retirement. A connection counts as a replacement while earlier non-shutdown retirements
  outnumber earlier replacements; every other connection only fills capacity. PoolStatus is
  unchanged. The type is modelled as eventlog.pool.PoolChurn in ess/pool/, and a test holds
  its encoding to the committed schema in ess/pool-generated/. The new default feature
  pool-churn only tells the capacity harness that the accessor exists; the accessor is not gated.
• Capacity laboratory: each capacity worker reports its pool's churn, and capacity-sweep carries
  it per worker into each configuration receipt. The metrics collector records, for a sample over
  eight client connections, each counted backend's pid, backend_start, state,
  state_change and xact_start, and whether each pid is still counted in the next sample. The
  <= 8 connection check is unchanged.

Fixed

• Capacity laboratory: the comparative lane's <= 8 connection budget counts only the workload's
  own backends. capacity-sweep starts every worker of both adapters with
  application_name=eventlog-capacity-workload, and the metrics collector reports
  workload_connections per sample and workload_connections_max per configuration beside the
  unchanged all-backend connections and connections_max_excluding_observer. The CI service's
  pg_isready health check opens short-lived client backends; one sampled beside the eight pool
  connections failed the lane with nine. A maximum of zero workload connections also fails the
  budget. Over-threshold attribution still covers every client backend, and no application name
  is written to the receipt. See docs/design/comparative-capacity-connection-budget.md.
• Capacity laboratory: each comparative configuration starts from an empty pg_stat_statements.
  The metrics collector resets the extension immediately before the configuration's opening
  statement snapshot, so an eviction caused by statements registered earlier on the same server,
  such as the production gate's, can no longer move dealloc inside a configuration and fail its
  envelope. A refused or timed-out reset leaves that snapshot unavailable, which fails the
  configuration. The deallocation and stats_reset equality checks are unchanged. See
  docs/design/comparative-capacity-connection-budget.md.