Release note: Router Memory Prometheus label sets

Upgrading to a release that includes #4326 changes the published label sets for two Router Memory metrics. Update dashboards, recording rules, and alerts that reference the old labels before or immediately after rollout.

llm_memory_retrieval_total

	Label set
Before	backend, status, user_id
After	backend, status

status continues to distinguish retrieval outcomes such as hit, miss, and error on Milvus, Valkey, and Qdrant backends. The counter no longer creates one time series per authenticated user.

Example migration:

# Before (per user — no longer valid)
sum by (backend, status, user_id) (llm_memory_retrieval_total)

# After (aggregate by backend and outcome)
sum by (backend, status) (llm_memory_retrieval_total)

llm_memory_store_size

	Label set
Before	backend, user_id
After	backend

This gauge was not wired to live store counts on earlier releases; the user_id dimension was unused in production paths. The label is removed so a future wiring does not reintroduce unbounded identity cardinality.

Qdrant retrieval telemetry

On earlier releases, Qdrant Retrieve did not increment llm_memory_retrieval_total. It recorded a generic successful store operation with zero duration instead. After this change, Qdrant retrieval emits the same hit/miss/error retrieval counters and durations as Milvus and Valkey.

Per-user analysis

Router Memory still scopes retrieval and storage by user in the store and request path. Per-user debugging belongs in logs, traces, and management APIs—not Prometheus labels on these counters.