Add multi-gateway gRPC concurrency and pointlist isolation tests

[takashikasuya]

Summary

Closes the test-coverage gaps identified in #114: BuildingOS previously had no test proving that concurrent gateway gRPC connections (GatewayIngress telemetry ingest, GatewayEgress control plane) and the per-gateway point-list export (#224) stay correctly isolated between gateways.

• GatewayIngress (GatewayIngressServiceTest.cs): two gateway streams processed concurrently (forced onto separate thread-pool threads via Task.Run, since the test fakes complete synchronously and would otherwise never yield) never cross-contaminate telemetry, and an identity-spoofing attempt on one connection doesn't affect a legitimate concurrently-connected gateway.
• GatewayEgress (GatewayEgressReplicaRoutingTest.cs, GatewayConnectionRegistryTest.cs): extended the existing 2-replica routing test to 3 concurrent gateways; added a point-list-push routing-isolation test (the shared bus's point-list subscription was previously a no-op stub); added a concurrent same-gateway-id connect-race test for GatewayConnectionRegistry.
• Pointlist sync isolation (ListGatewayPointListTest.cs, GatewayProvisioningControllerTest.cs, new OxiGraphSeedHostedServicePointListPushTest.cs): added a fake OxiGraph endpoint that actually scopes SPARQL responses by the gatewayId literal in the query (previous tests always returned one canned response regardless of query content, so a dropped filter would have gone undetected); fixed the controller test's mock to be gatewayId-aware for the same reason; added coverage for the seed-time per-gateway push fan-out (OxiGraphSeedHostedService.PublishPointListUpdatesAsync), which had zero test coverage before.

Production fix (found while writing the push-isolation test): PublishPointListUpdatesAsync's try/catch wrapped the entire per-gateway loop, so one gateway's publish failure silently aborted the push for every gateway after it in that seed run. Moved the try/catch inside the loop so each gateway is published independently, and added a BuildingOsMetrics.PointListPushSignals counter (the codebase's established "logged + metered" convention for best-effort per-item loops) so a partial failure is visible in Prometheus/Grafana, not just logs.

nexus-gateway (the external BOWS agent) was intentionally left out of scope — it's architected as one-gateway-id-per-process and multi-gateway federation is explicitly out of scope there.

This PR was reviewed with a multi-angle automated code review (correctness, reuse, simplification, efficiency, altitude, CLAUDE.md conventions) before being opened; findings that surfaced real issues (tests not actually running concurrently, a metrics gap, some test-fake duplication) were fixed.

Test plan

• dotnet test --filter "FullyQualifiedName!~IntegrationTest" (unit tests) — could not run in this sandboxed environment: the .NET SDK download hosts (dotnetcli.azureedge.net, builds.dotnet.microsoft.com) are blocked by the environment's egress policy, so nothing here has been compiled. All new/changed code was verified by careful manual reading (types, signatures, Moq/xUnit usage, SPARQL query text matched against production constants) and cross-checked by 7 independent review passes, but please run the real test suite (or the manual-trigger oss-ci GitHub Actions workflow) before merging.
• Manually verified against the local OSS stack

Checklist

• I read CONTRIBUTING.md and followed the coding conventions in CLAUDE.md
• I added or updated tests for the change
• I updated relevant docs (README.md, docs/) if behavior or setup changed — no doc changes needed (test-only + one internal metric addition, no behavior/setup change)

---

Generated by Claude Code