fix(store): omit empty statuses from findOldestCreatedAt (false lag-gauge reading)

[endrju19]

The bug

PostgresOutboxStore and MysqlOutboxStore both pre-seeded the result map with clock.instant() for every requested status before running the query:

val result = statuses.associateWith { clock.instant() }.toMutableMap()  // ← every status = "now"
// ... SQL overwrites only statuses that have rows ...

A status with no rows stayed at the seeded near-now timestamp instead of being absent from the map.

Why it matters

The lag gauge in okapi-micrometer relies on absence to report 0 for empty statuses:

snapshot.oldest[status]?.let { Duration.between(it, now).toMillis() / 1000.0 } ?: 0.0

With the pre-seed, an idle status (e.g. no PENDING entries) reported a small false lag — the elapsed time between the store query and the gauge publish — instead of 0. A dashboard would show a non-zero lag for a status that has nothing queued.

Why no test caught it

The existing unit test uses a stub whose findOldestCreatedAt does oldest.filterKeys { it in statuses } — i.e. it correctly omits absent statuses. So the stub and the real JDBC adapters disagreed: the test was green while production diverged. This is a stub-more-correct-than-impl trap.

The fix

Start from an empty map so absent statuses are omitted; the gauge's ?: 0.0 fallback then applies.

- val result = statuses.associateWith { clock.instant() }.toMutableMap()
+ val result = mutableMapOf<OutboxStatus, Instant>()

Applied identically to both adapters (the bug was copy-pasted from one to the other).

Test

Added one test to the shared OutboxStoreContractTests — so it runs against real Postgres and real MySQL (Testcontainers), not a stub. It persists PENDING entries only, then asserts findOldestCreatedAt returns a key only for statuses that have rows (PENDING present with the oldest created_at; DELIVERED/FAILED absent). The assertion is on the key set, which catches the structural bug even under a fixed test clock (where a value-only assertion would deceptively pass).

Test plan

• ./gradlew :okapi-postgres:compileKotlin :okapi-mysql:compileKotlin — clean (no warnings)
• :okapi-integration-tests:compileTestKotlin — new test compiles
• ./gradlew ktlintCheck — green
• ⚠️ The new integration test was not executed locally — Docker/Testcontainers is unavailable in the authoring environment. RED→GREEN was verified by reasoning (buggy code returns all 3 requested keys → key-set assertion fails; fixed code returns only populated keys → passes). Please run ./gradlew :okapi-integration-tests:test with Docker up before merging.

Note — deferred cleanup

clock is now unused by both stores (it existed only for the buggy pre-seed; Kotlin emits no warning for an unused constructor val). Removing it is a breaking constructor change touching ~20 call sites (Spring autoconfig bean factories, benchmarks, ~15 tests) and is intentionally left to a separate follow-up to keep this bugfix focused and easy to review/cherry-pick.