✨ feat(controller): add stateful replica health classifier

[nitinstp23]

Context

main and the parallel feat/replica-health-detection branch both descended from the replica-health design spec (83964ef) and independently solved the same "silent replication outage" problem — they're siblings, not parent/child. This PR ports the only piece main genuinely lacked.

main was already equal or ahead on:

• per-replica metric cleanup (DeleteClusterMetrics partial-match sweep + RecordReplicaSet sweep-and-rebuild),
• negative behind handling (json.Unmarshal into int64),
• multi-database "worst-status-wins" aggregation,
• legacy (pre-3.x) SHOW REPLICAS format support.

So none of that was ported. The net-new piece is the stateful classifier, re-implemented against main's data model. main's richer per-replica metrics are kept intact; the new gauges are additive.

What's new

• replica_classifier.go — a per-replica state machine that distinguishes a freshly-registered replica (Transient, 30s grace) from a genuinely broken data channel (DataChannelDown), plus Behind / BehindTooLong / Invalid / UnknownStatus. It clears the behind timer when the channel is down so behind_seconds doesn't grow without bound. Status is normalized (lowercase/trim) to match the client's summarizeDataInfo.
• ReplicationSpec.BehindAlertThreshold (default 5m) — configurable duration-based "behind too long" detection. main previously recorded raw lag but never flagged sustained lag.
• Two new gauges — memgraph_replica_data_channel_up and memgraph_replica_behind_seconds, recorded in CheckReplicationHealth and swept on replica unregister / cluster deletion.
• Specific, actionable events — ReplicaDataChannelDown / ReplicaInvalid / ReplicaBehindTooLong (with remediation text) replacing the generic ReplicaUnhealthy.

ReplicationManager now holds per-replica timers (in-memory; reset on operator restart) and an injectable clock for deterministic tests. CheckReplicationHealth keeps its existing signature so RecordReplicaSet continues populating main's snapshot metrics.

Bug fixed vs. the source branch

The source branch declared BehindAlertThreshold as a value-typed metav1.Duration. Because omitempty can't omit a struct, a typed Go client (controller-runtime, envtest) sends "0s" when the field is unset, which the strict >0 CEL rule rejects with a 422 — breaking the existing controller envtest. Fixed here by making the field a pointer: unset → omitted → CRD default 5m applies, while explicit 0s/negatives are still rejected.

Behavior change worth noting

A replica within the 30s grace window now counts toward HealthyReplicas (Transient) instead of being immediately Invalid — the intended anti-flapping behavior.

Testing

• go build ./..., go vet ./... — clean
• go test ./internal/... ./api/... — all green, including the envtest controller reconcile test
• New: 17-case classifier table test (grace period, behind-too-long, invalid-wins, timer-clear, status normalization) + gauge record/delete tests
• CRD + deepcopy regenerated via make generate manifests (diffs contain only the new field)

[irfn]

LGTM