What would you like to be added:

Improve llm-d-router (EPP) observability documentation and operator-facing guidance so metrics, tracing, and embedded llm-d-kv-cache telemetry are discoverable and consistent with the current implementation.
Concrete deliverables:

1. Update docs/metrics.md
   • Document the canonical subsystem llm_d_router_epp_* (introduced in Rename metrics and introduce a single subsystem for epp #851 / PR Add new metrics with single subsystem #1071).
   • Mark deprecated series (inference_*, llm_d_inference_scheduler_*) and point to replacements.
   • Expand beyond disagg-only metrics: request/latency, pool, scheduler, plugin, datalayer errors (high-level catalog + labels).
2. Document scrape topology for operators
   • EPP Prometheus endpoint: default :9090/metrics (serveMetrics / controller-runtime metrics server).
   • Model-server metrics scrape path (data layer / Helm metricsDataSource).
   • Embedded llm-d-kv-cache: when precise-prefix-cache-scorer is enabled with indexerConfig.kvBlockIndexConfig.enableMetrics: true, kvcache_index_* (and related) metrics are registered on the same EPP registry—no separate kv-cache HTTP endpoint is required in standard deployments.
3. Cross-link related work (implement in separate PRs/issues where noted)
   • Metrics naming / plugin labels: Align on name/label for plugin scoped metrics #1243
   • Deprecate/remove legacy metrics: [Deprecation]: Deprecate old metrics #1070, [deprecation] Remove lagacy data and metrics collection #962
   • EPP operations guide: llm-d router (particularly EPP) operations guide #1291
   • E2E metrics stability: [Testing] Resolve EPP Metrics Verification Flakiness in E2E Test Suites #1192
4. (TBD follow-up, separate issue or sub-task) Add an e2e or integration assertion that kvcache_index_* appears on EPP /metrics when enableMetrics: true and precise-prefix-cache-scorer is configured.

Why is this needed:

• docs/metrics.md still states all metrics use the llm_d_inference_scheduler subsystem, which is outdated after the llm-d Router rename (docs: rename inference scheduler to llm-d Router and fix typos #1233) and metrics migration (Rename metrics and introduce a single subsystem for epp #851).
• Operators and SREs cannot easily answer: what to scrape, which metric names to use in dashboards/alerts, and whether llm-d-kv-cache (separate repo) requires its own /metrics port when used via EPP.
• Today, sample configs often set enableMetrics: false (e.g. deploy/config/sim-epp-kvcache-config.yaml), so kvcache_* series are easy to miss even though the integration path (shared ctrlmetrics.Registry) already exists in code.
• Reduces duplicate work (e.g. opening a standalone kv-cache metrics endpoint) and aligns monitoring with the EPP Helm ServiceMonitor pattern (router.monitoring.prometheus).