feat(a2a): emit aggregated turn token usage on the terminal status update

[QuentinBisson]

Supersedes #2332, which the stale bot closed and GitHub refuses to reopen. This is a reimplementation, not a rebase: both executors were rewritten to delegate to the ADK library's own A2A executor, so the manual event loop the original PR hooked into no longer exists.

What this adds

Every terminal status update (completed, input-required, failed) carries a kagent_usage_total metadata entry with the token usage summed across the task:

{
  "promptTokenCount": 300,
  "candidatesTokenCount": 50,
  "totalTokenCount": 350,
  "modelVersion": "gpt-4o-2024-11-20"
}

Partial (streaming chunk) events are skipped, since each LLM call reports its usage on the final non-partial event. Resumed tasks seed the accumulator from Task.metadata, so a task that spans several executions (HITL input-required cycles, follow-up messages) reports a task-lifetime total instead of the last segment only. The value is a running total re-emitted on each terminal update, so consumers take the latest value rather than summing across executions.

The value is serialized with the same genai UsageMetadata mapping as the existing per-event adk_usage_metadata / kagent_usage_metadata entries, plus modelVersion, so consumers can share one parser (with the caveat that modelVersion is not a field of the genai type). Tests in both languages assert that equality directly.

totalTokenCount is derived from prompt + candidates + thoughts when no provider reported one. The Anthropic models report per-call input and output counts without a total, and a zero total next to non-zero counts reads as "no tokens were used".

a2a-go and a2a-python both merge terminal status-update metadata into the stored task, which is how the total is persisted and how the next execution reads it back.

How it differs from #2332

• Go: aggregation hangs off a RunnerProvider that wraps the ADK runner instead of AfterEventCallback. The callback only fires for ADK events that convert to an A2A artifact, and the input-required processor consumes the paused tool-call event, which is exactly the event carrying the usage of the call that paused. Wrapping the runner keeps the original vantage point: every ADK event, before conversion. The provider mirrors the ADK default provider (it has to, since RunnerProvider replaces RunnerConfig), and a test reflects over adka2a.RunnerConfig so a field added upstream fails the build instead of being silently dropped. Stamping uses the new AfterExecuteCallback hook, which covers all three terminal states in one place.
• Python: aggregation runs from an ADK runner plugin (on_event_callback), for the same reason. The A2A after-event interceptor only runs for ADK events the converter turns into at least one A2A event, and the event whose long-running function call is stripped before conversion produces none. The plugin sees every ADK event the runner produces, matching the Go vantage point.
• Test intent is ported rather than the test code; the Go tests now run end-to-end through the executor (including a full HITL pause/resume cycle) instead of unit-testing the accumulator in isolation.

Known limits

• A cancelled task, or a client that disconnects mid-stream, does not reach the terminal status write, so the tokens of that in-flight execution are not added to the total. Earlier executions of the task keep theirs.
• toolUsePromptTokenCount is not aggregated. It is inside totalTokenCount, so the total stays correct, but the breakdown fields do not sum to it on Vertex.
• modelVersion is last-writer-wins over summed counts, so a task that uses more than one model reports a total that cannot be split back into per-model cost.

Verification

• go build ./adk/... ./core/..., go test ./adk/..., go vet, gofmt
• uv run pytest packages/kagent-adk/tests/unittests (390 passed) - the 9 failures in models/test_tls_e2e.py are pre-existing in this environment (verified against a clean checkout) and unrelated

[Copilot]

Copilot encountered an error and was unable to review this pull request. You can try again by re-requesting a review.

[QuentinBisson]

Rebased onto main (bc92b68). Re-verified locally after the rebase: go build, go vet, and the touched Go/Python test packages pass.

The three remaining red checks are not from this branch, they reproduce identically on main at the same commit (run 32034286380):

• go-unit-tests: the go/api/v1alpha2 and go/api/v1alpha3 CEL tests fail in setup, not in assertions. setup-envtest cannot write the 1.36.2-linux-amd64 assets ("unable to create file kubectl/etcd from archive to disk"), so the control plane never starts. Looks like parallel test packages racing on the same go/bin/k8s store.
• upgrade-tests (prev-stable) / rolling-upgrade-tests (prev-stable): UPGRADE FAILED: context deadline exceeded in TestUpgrade/upgrade_with_helm.