Isolate per-connection failures in graph reconciler

[ArnobKumarSaha]

Problem

The graph adjacency cache (objGraph) is built per-object by the reconciler in pkg/graph/reconciler.go. ListConnectedObjectIDs evaluated every connection in rd.Spec.Connections, and on the first non-NotFound/NoMatch error it did return nil, err — aborting the whole object. The reconciler then returned without calling objGraph.Update, so the object got zero edges.

Observed in the wild: a PostgreSQL 11 instance whose ui.kubedb.com query view runs a pg_stat_statements query using total_exec_time (a column that only exists in PG ≥ 13). The query fails with pq: column "total_exec_time" does not exist, which poisoned the entire graph build — offshoot, auth_secret, catalog, and the working views were all dropped, even though they resolve purely from the K8s API and have nothing to do with the SQL view.

Fix

• ListConnectedObjectIDs: a failing connection is now skipped (error wrapped with src -> target context and accumulated), the remaining connections still resolve, and the function returns (edges, utilerrors.NewAggregate(errs)).
• Reconcile:
  • Discovery errors keep their existing semantics — don't overwrite the graph, requeue fast (500ms). Detected via the new anyDiscoveryError helper (the k8s aggregate type implements Is but not As, so IsDiscoveryError can't see through it).
  • For any other failure, the partial graph is persisted via objGraph.Update first, then the error is returned for exponential-backoff retry.

Net: one bad connection no longer wipes an object's whole graph.

Out of scope

The kubedb-side query bug (total_exec_time vs total_time for PG < 13) lives in a different repo and is unchanged.

Test

• go build ./pkg/graph/..., go vet, gofmt clean
• go test ./pkg/graph/... passes