feat(metrics): record eventsourced pg outbox publishes and retries (#182)

## Why
authz-service now publishes through pg/v2's transactional outbox, and nothing measured it: `NewEventsourcedMetrics` ignored the `pg.Outbox*` metrics. Without them a failing or abandoned publish (a lost privilege revocation for running authz_clients) can't be alerted on.

## What
- `pg.OutboxPublish` → histogram `eventsourced.outbox.publish.duration` (`event.type`, `success`)
- `pg.OutboxRetry` → counter `eventsourced.outbox.retries` (`event.type`, `permanent`)
- `pg.OutboxBatch` / `pg.OutboxCleanup` stay ignored.

otelsetup now imports `codeberg.org/eventsourced/pg/v2` v2.1.1 (`go` directive becomes 1.26.0; grpc patch bump via MVS). All current consumers already depend on pg/v2; a release raises their minimum to v2.1.1.

## Tests
`go test -race ./...` green; the contract test records both new metrics.

## Review
Go Backend expert reviewed: no Critical/High. Noted: the test checks instrument names only, not attributes; the pg/v2 dependency is worth a line in the release notes.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

https://claude.ai/code/session_013DJouG8ZZZxKtodF9kDvzj
Reviewed-on: https://gitea.unbound.se/shiny/otelsetup/pulls/182

metrics.go

@@ -5,6 +5,7 @@ import (
 	"errors"
 
 	"codeberg.org/eventsourced/eventsourced"
+	"codeberg.org/eventsourced/pg/v2"
 	"go.opentelemetry.io/otel"
 	"go.opentelemetry.io/otel/attribute"
 	"go.opentelemetry.io/otel/metric"
@@ -40,6 +41,8 @@ type eventsourcedMetrics struct {
 	snapshotStoreDur metric.Float64Histogram
 	snapshotLoadDur  metric.Float64Histogram
 	idempotencyCheck metric.Int64Counter
+	outboxPublishDur metric.Float64Histogram
+	outboxRetries    metric.Int64Counter
 }
 
 // NewEventsourcedMetrics builds an eventsourced.MetricsRecorder that records to
@@ -78,6 +81,8 @@ func NewEventsourcedMetrics() (eventsourced.MetricsRecorder, error) {
 		snapshotStoreDur: hist("eventsourced.snapshot.store.duration", "Time taken to persist a snapshot."),
 		snapshotLoadDur:  hist("eventsourced.snapshot.load.duration", "Time taken to load a snapshot."),
 		idempotencyCheck: counter("eventsourced.idempotency.checks", "Number of command idempotency lookups."),
+		outboxPublishDur: hist("eventsourced.outbox.publish.duration", "Time taken to publish an event from the outbox."),
+		outboxRetries:    counter("eventsourced.outbox.retries", "Failed outbox publishes; permanent ones are never retried."),
 	}
 	if err := errors.Join(errs...); err != nil {
 		return nil, err
@@ -86,8 +91,8 @@ func NewEventsourcedMetrics() (eventsourced.MetricsRecorder, error) {
 }
 
 // Record implements eventsourced.MetricsRecorder. Metric types the adapter does
-// not recognise (for example pg outbox metrics when the outbox is not enabled)
-// are ignored.
+// not recognise are ignored, and so are pg.OutboxBatch and pg.OutboxCleanup: a failed
+// publish already shows in the publish and retry instruments.
 func (e *eventsourcedMetrics) Record(ctx context.Context, raw eventsourced.Metric) {
 	switch m := raw.(type) {
 	case eventsourced.CommandDuration:
@@ -119,5 +124,15 @@ func (e *eventsourcedMetrics) Record(ctx context.Context, raw eventsourced.Metri
 			attribute.String("aggregate.type", m.AggregateType),
 			attribute.Bool("hit", m.Hit),
 		))
+	case pg.OutboxPublish:
+		e.outboxPublishDur.Record(ctx, m.Duration.Seconds(), metric.WithAttributes(
+			attribute.String("event.type", m.EventType),
+			attribute.Bool("success", m.Success),
+		))
+	case pg.OutboxRetry:
+		e.outboxRetries.Add(ctx, 1, metric.WithAttributes(
+			attribute.String("event.type", m.EventType),
+			attribute.Bool("permanent", m.Permanent),
+		))
 	}
 }