[MartinForReal]

What

Implements the GenAI token-usage metrics observability gap from #2148 for the Go ADK runtime. The Go ADK (google.golang.org/adk/v2) ships spans only, no metric instruments, so kagent supplies the stdlib Prometheus instrument itself, in the shape of the open/dirty #2149 but rebuilt from scratch against today's go/adk source and aligned with the Python runtime's attribute set per #2458.

How

• pkg/telemetry/metrics.go — a gen_ai_client_token_usage Prometheus histogram (semconv gen_ai.client.token.usage, semconv-recommended buckets) on the default registry.
• Labels align with what upstream Google ADK Python emits for the same instrument (one dashboard across runtimes): gen_ai_token_type (input/output), gen_ai_request_model, gen_ai_response_model (falls back to request model), gen_ai_provider_name (semconv well-known value via SemconvProviderName), gen_ai_agent_name, gen_ai_operation_name (chat), error_type.
• Gating — recording and the /metrics endpoint are default-OFF behind OTEL_METRICS_ENABLED=true, matching the existing OTEL_TRACING_ENABLED / OTEL_LOGGING_ENABLED gates.
• pkg/a2a/server/server.go — serves /metrics on the A2A mux when the gate is on, excluded from request tracing/span flushing.
• pkg/a2a/executor.go — records token usage from the ADK AfterEventCallback: skips partial (streaming) events so one observation is recorded per LLM call; input = PromptTokenCount, output = CandidatesTokenCount + ThoughtsTokenCount from UsageMetadata.
• cmd/main.go / pkg/config — model + provider labels resolved at startup; agent name from AppName.
• docs/genai-token-metrics.md — documents the metric, labels, gating, and scraping.

Testing

• go build ./adk/..., go vet, and golangci-lint (repo config) are clean.
• go test -race passes for the changed packages (telemetry, a2a, a2a/server, config).
• Two focused test files: metrics_test.go (record / skip-zero / disable-gate / response-model fallback / labels / provider-name mapping / /metrics handler) and executor_metrics_test.go (partial events not double-counted, per-LLM-call aggregation).

Notes / decisions

• Pod scrape annotations: the annotation piece of chore(deps): bump the go-minor-patch group in /go with 3 updates #2099 targeted go/core/internal/controller/translator/agent/manifest_builder.go, which no longer exists on main after the controller/runtime rework. That path was dropped rather than reinvented; operators can add prometheus.io/* annotations and this is documented. Flag if you want the annotation automation wired against the current runtime-provisioning location.
• make -C go lint could not be run here: make and the kube-api-linter plugin are unavailable in this environment. Lint was run with the repo's .golangci.yaml minus the plugin (0 issues); kube-apilinter only lints api/v1alpha CRD paths, untouched here.

Refs: #2148

[MartinForReal]

Hi @supreme-gg-gg, friendly ping for review when you have a moment. This exposes the gen_ai token usage Prometheus metric for the Go ADK runtime. Ready for another look. Thanks!