fix(datalayer): sync cross-replica state from one goroutine, not one per endpoint

[Lucas-Fernandes-Martins]

Why

Review feedback on the upstream PR llm-d/llm-d-router#2310, which introduced the independent cross-replica sync ticker already on lfm/combined-fixes.

ahg-g asked why the sync runs a goroutine per endpoint:

Note that the polling dispatcher was designed to be per endpoint because it needed to establish a connection and poll information per endpoint. This doesn't seem to be the case for cross replica syncing, a single go routine iterating over the endpoints list and calling sync on each should be sufficient.

He's right, and the per-endpoint design was also leaking. The goroutine only returns on ctx.Done(), but NewEndpoint is called with the process-lifetime parentCtx (datastore.go:438), so ReleaseEndpoint stopped the collector while the sync ticker kept publishing state for an endpoint that no longer exists — a goroutine leak plus stale cross-replica writes. Copilot flagged the same thing independently.

A single loop over the live endpoints fixes both at once: no per-endpoint goroutine to leak, and a released endpoint simply stops being visited.

Summary

• One runCrossReplicaSync(ctx) goroutine, started from the existing Runtime.Start(ctx, mgr), which already receives a context and runs immediately after Configure (where crossReplicaPub is built). Each tick it iterates the live endpoints and dispatches.
• Collector keeps the endpoint it was started with. It already documents itself as running "data collection for a single endpoint" and Start is handed that endpoint, so this makes the existing registry enumerable via collectorManager.RangeEndpoints — no wrapper type, and no second registry that could drift out of step with the collectors.

No change to what is published or how often; only how many goroutines do it, and that removed endpoints stop.

Reviewer guide

File	What to look at
pkg/epp/datalayer/runtime.go	Real logic. The goroutine moves from NewEndpoint to Start, and runCrossReplicaSync drops its ep parameter in favour of RangeEndpoints.
pkg/epp/datalayer/collector.go	Small addition. One ep field set by Start, plus an Endpoint() accessor under the existing mutex.
pkg/epp/datalayer/manager.go	Purely additive. RangeEndpoints only; Register, Remove and StopAll are untouched.
pkg/epp/datalayer/cross_replica_publisher_test.go	New tests only, appended at the end. Existing tests untouched.

Worth a second opinion on placing the launch in Start: it is the only Runtime method taking a lifetime context, and it runs after Configure, so crossReplicaPub is always set by then — but the ordering is load-bearing and the nil check is what keeps it safe if that ever changes.

Note that RangeEndpoints only sees collectors that have been started. A collector is registered a few lines before Start is called, so a brand-new endpoint can miss at most one tick — before which it has nothing worth publishing anyway.

Test plan

• go build ./..., gofmt -l, and go vet ./pkg/epp/datalayer/... all clean.
• go test -race ./pkg/epp/datalayer/... green, including the pre-existing cross-replica publisher and collector tests.
• Two new tests:
  • TestRunCrossReplicaSync_PublishesAllEndpoints — one loop publishes every registered endpoint.
  • TestRunCrossReplicaSync_StopsAfterRelease — regression guard for the leak: publishing stops once the endpoint is removed from the registry.

Note on the third review comment

Copilot also flagged crossReplicaPublisherType as an unused const that would fail to compile. That is stale — the symbol no longer exists anywhere in the tree (grep finds no references) and the build is clean. No change needed.