feat(otel): unit metadata + decouple OTLP event mirroring from the webhook

[KIvanow]

🔗 Upstream Valkey issue: valkey-io/valkey#4078 — OpenTelemetry (OTLP) metrics export

What

Three OTLP-mirror polish items (valkey#4078) so an OTLP-only deployment — collector configured, no webhook — gets complete, well-typed telemetry rather than a subset gated behind the Pro webhook.

Changes

1. Unit metadata on mirrored metrics. planInstruments now derives a UCUM unit from the prom-client name suffix, and the exporter passes it to createObservable{Gauge,Counter}:

suffix	unit
_bytes	By
_seconds	s
_milliseconds	ms
_ratio	1
_percent	%
_total / unknown	(none)

The mirrored metric name is never rewritten — the unit lives in OTLP metadata, so downstream backends can convert/label. unit is omitted from the instrument options entirely when nothing is derivable (an empty-string unit is meaningful noise).

2. Decouple anomaly OTLP dispatch from the Pro webhook gate. addAnomaly no longer wraps the OTLP emit in webhookEventsProService.isEnabled(). OTLP is its own opt-in channel, gated only by the OTEL_* env vars inside the dispatcher (which no-ops when disabled). A licensed operator can now ship anomalies to their collector without also configuring a webhook.

3. Widen event coverage. cluster.failover (PrometheusService) and the instance.down / instance.up availability edges (HealthService) now mirror to OTLP on the state transition itself, independent of webhook config. The availability edges previously only advanced their state when a webhook dispatcher was wired; the transition + OTLP emit now happen regardless, with the webhook dispatch kept as a separate best-effort step.

Test plan

• deriveUnit suffix table + "unit only when derivable" on the exporter (fake meter captures instrument options).
• Anomaly mirror fires with the Pro webhook disabled (isEnabled → false), via the data-loss Rule A path.
• prom-otel-bridge, otel-metrics-exporter (15 passed) and anomaly.service (128 passed) suites green; apps/api tsc clean.

Addresses valkey#4078.

---

Note

Medium Risk
Changes when operational events are emitted and can increase OTLP traffic for deployments that previously relied on webhook gates; behavior is still bounded by OTEL env config and existing transition logic.

Overview
Improves OTLP-only deployments so telemetry is complete and better typed without depending on Pro webhook configuration.

Mirrored metrics now attach UCUM unit metadata derived from Prometheus name suffixes (_bytes → By, _seconds → s, etc.); metric names are unchanged and empty units are omitted from OTel instrument options.

OTLP event mirroring is decoupled from the Pro webhook gate: anomaly detection always calls otelEvents.dispatch (still no-ops unless OTEL_* is set). Instance up/down in HealthService and cluster.failover in PrometheusService emit on state transitions regardless of webhook wiring; webhooks remain a separate optional step. Cluster failover state is advanced even when only OTLP fires.

^{Reviewed by Cursor Bugbot for commit 60ca87e. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 8cde829. Configure here.}

[jamby77]

Approving — logic traced clean. The health/cluster state-machine restructure preserves edge semantics (one event per transition, no double-fires), OTLP failures can't suppress webhooks (separate try/catch, and OtelEventDispatcherService.dispatch is synchronous so the catch is real), and the @Optional() injection resolves since OtelEventsModule is @Global(). Two follow-ups worth doing:

1. Test gap on the riskiest part. The HealthService instance.down/instance.up and PrometheusService cluster.failover transitions now advance without webhook wiring, and neither path has coverage — health.service.spec.ts never touches handleInstanceDown/handleInstanceUp, and there's no PrometheusService unit spec at all. A future refactor that reintroduces the webhook guard would silently kill OTLP-only availability events again (exactly the regression class this PR fixes) with nothing going red.

2. deriveUnit() misses _seconds_total. The suffix match bails on _total before deriving, so betterdb_cpu_sys_seconds_total / betterdb_cpu_user_seconds_total get no unit even though they measure seconds. The OTel Prometheus convention derives from the token preceding _total — checking endsWith('_seconds_total') before the _total bail-out (or stripping _total first) fixes it. Same family: _us (e.g. betterdb_commandstats_latency_us) is unmapped, though I see the PR scopes to unambiguous suffixes deliberately.

The bridge/exporter half is well-tested (suffix table, unknown/_total cases, exporter omitting the key when underivable, and the anomaly mirror firing with isEnabled() → false).