test(telemetry): replace hand-written telemetry expectations with record/replay goldens

[RKest]

Please ensure you have read the contribution guide before creating a pull request.

Linked issue

• Related: telemetry: default to the experimental OpenTelemetry GenAI semantic conventions, with a one-env rollback #1634 (this is the test half of feat(telemetry)!: default to experimental GenAI semconv, with ADK_TELEMETRY_SCHEMA_VERSION_OPT_IN=otel_semconv_1_36 rollback #1635, split out so it can be reviewed on its own)

Problem:
The functional telemetry tests compared each run against hand-written Go literals in internal/telemetry/telemetrytestcase. Only one content-capture mode was covered, and every telemetry change meant hand-editing expected span trees. The expectations could also not be compared with adk-python's recordings.

Solution:
Record/replay goldens, following adk-python's test_functional.py:

• Golden: internal/telemetry/functionaltest/testdata/<test_id>.json, where the id is <scenario>/[<variant>-]<capture>-<schema>, holds the span tree, with each log record nested under the span it was emitted in. The digest shape (root_span / name / attributes / status / children / logs, plus the PRESENT sentinel) follows adk-python's _digests.py, with one difference: a JSON-string attribute is wrapped as {"JSON_STRING": …}, so it stays distinct from a structured one.
• Replay: plain cmp.Diff between the recording and the golden. Recording fails on a log record emitted outside any span, and a golden that no case records fails the suite.
• Re-record: go generate ./internal/telemetry/functionaltest.

Scenarios are typed, and data is kept apart from setup:

internal/telemetry/functionaltest/scenarios/
  testcasedata/          data only
    testcasedata.go      Scenario interface, Capture, SchemaVersion, Case, Matrix
    toolcall/            llmagent + function tool; Failure: inference-error, tool-error
    streaming/           two streamed chunks
    delegation/          agent-as-tool
    dynamicworkflow/     static node, dynamic node, agent nodes, cached node;
                         Failure: static-node-error, dynamic-node-error,
                         first-agent-error, second-agent-error
  testcaseimpl/          turns a Scenario into agents, models and tools
    common.go            Run (real Runner), Cases, the Scenario → builder switch
    toolcall.go …        one builder per scenario

Each scenario package has its own strongly typed Failure (or none) and a Matrix() covering every variant under:

• every capture mode: no_content, span_only, event_only, span_and_event;
• both ADK_TELEMETRY_SCHEMA_VERSION_OPT_IN values: otel_semconv_1_36 and otel_semconv_1_44.

That comes to 80 goldens. main does not read the schema variable yet, so each otel_semconv_1_36 / otel_semconv_1_44 pair is identical here. #1635, which starts reading it, then changes only otel_semconv_1_44 goldens, which shows exactly what telemetry changes and that the legacy format does not.

Behavior change

Nothing. This PR changes tests only.

Testing Plan

Unit Tests:

• All unit tests pass locally: the AGENTS.md module loop (go mod tidy -diff, go build, go test -race -shuffle=on, golangci-lint run v2.3.1) is clean in both modules.

Coverage kept: each deleted expectation has a golden that pins at least as much:

• AgentWithToolCase → toolcall/no_content-otel_semconv_1_36.json
• AgentWithToolCaptureContentCase → toolcall/span_and_event-otel_semconv_1_36.json
• the five Workflow*Cases → dynamicworkflow/*no_content-otel_semconv_1_36.json

The goldens pin the legacy tool arguments and responses by value; the old expectations had PRESENT there.

Determinism: 5 re-recordings under -shuffle=on produced byte-identical goldens.

With your source change reverted and your tests kept, which test fails?
Not applicable: there is no source change. The goldens pin span names, attributes, status codes, the tree shape, and log event names, bodies and attributes. Like adk-python's digest, they do not pin span events or status descriptions.

Checklist

• I have read the CONTRIBUTING.md document.
• I have performed a self-review of my own code, plus one fresh-context review pass; its findings are addressed.
• I have commented my code, particularly in hard-to-understand areas.
• I have added tests that prove my fix is effective or that my feature works.
• New and existing unit tests pass locally with my changes.
• I have manually tested my changes end-to-end. (Not applicable: tests only.)
• Any dependent changes have been merged and published in downstream modules.

[knapg]

overall, lgtm!

[wolo-lab]

One thing to fix before merge, and it is about the tests rather than the code: the new check that fails a recording when a log record is emitted outside any span has no test that exercises it.

Before this change the digest helper dropped such records silently by design (span_digest.go at the merge base, lines 102-109). Now BuildDigest fails instead, and the description lists that as one of the suite's guarantees. Replacing that t.Fatalf with continue leaves the whole suite green, and internal/telemetry/telemetrytest has no test file. So a later edit could quietly bring back the old drop. The check does work: a throwaway test that emits a record with no matching span hits the fatal. One thing to plan for when adding the test is that BuildDigest takes a concrete *testing.T, so asserting that it fails needs an error return or a testing.TB seam.

[RKest]

One thing to fix before merge, and it is about the tests rather than the code: the new check that fails a recording when a log record is emitted outside any span has no test that exercises it.

Before this change the digest helper dropped such records silently by design (span_digest.go at the merge base, lines 102-109). Now BuildDigest fails instead, and the description lists that as one of the suite's guarantees. Replacing that t.Fatalf with continue leaves the whole suite green, and internal/telemetry/telemetrytest has no test file. So a later edit could quietly bring back the old drop. The check does work: a throwaway test that emits a record with no matching span hits the fatal. One thing to plan for when adding the test is that BuildDigest takes a concrete *testing.T, so asserting that it fails needs an error return or a testing.TB seam.

Reverted the old behavior thanks

[wolo-lab]

LGTM, thanks!