ActiveRecord unbounded + recovery metrics stores

[bolshakov]

Background

Stoplight decides when to trip by counting outcomes. When a light has no window configured, its metrics are running totals: consecutive successes/failures, the last error, and last-success/last-failure timestamps. The same store, keyed by a different "kind", also tracks the outcomes of recovery probes while a light is healing (recovery metrics). Because writes can arrive out of order across processes, updates are guarded so a delayed write never overwrites a newer value. The Redis store does this with timestamp-guarded field writes; this ticket reproduces it with a locked read-modify-write.

Both stores implement the domain metrics-store port (snapshot, record-success, record-failure, clear). The Redis metrics stores are the reference to mirror, including how a failure is serialized to and from the domain failure object. The epic describes the schema.

Goal

Implement the ActiveRecord unbounded metrics and recovery metrics stores conforming to the domain metrics-store port, with the same out-of-order-write guards Redis has.

What to do

• Add the stoplight_metrics migration (one row per light and kind, per the epic schema).
• Implement record-success and record-failure as a transaction that ensures the row exists then reads-modifies-writes under a row lock, replicating the newer-timestamp guards. Record-failure returns the post-write snapshot.
• Implement the recovery-metrics store as the same behavior keyed by its own kind.
• Give rows a retention timestamp refreshed on write; expired rows are invisible to snapshots.
• Reuse the exception-serialization approach used by the existing metrics stores.
• Add both stores to the benchmark harness.

Acceptance criteria

• The metrics contract shared examples pass for both the unbounded and recovery metrics ActiveRecord stores, on SQLite and Postgres.
• A spec writing an older-timestamped update after a newer one shows the newer last-success/last-error is not regressed, matching the Redis store.
• A spec shows a row past its retention horizon is not reflected in the metrics snapshot.
• Both stores appear in the benchmark harness; Memory / Redis / ActiveRecord numbers are recorded in the PR description.
• bundle exec steep check and bundle exec standardrb pass.

References

• Port: ports/metrics_store.rbs
• Reference to mirror: redis/storage/unbounded_metrics.rb, redis/storage/recovery_metrics.rb, serialization base redis/storage/metrics.rb; memory: memory/storage/unbounded_metrics.rb
• Contract examples: data_store/metrics.rb, recovery_metrics.rb, metrics_snapshot.rb