Test gaps and decisions that the test audit found #83

The test audit on the test/test-audit-campaign branch found these gaps. They existed before the audit, and the audit did not remove their proof.

Decisions:

• The second parameter of telemetryDev in @telemetry-dev/ai-sdk and @telemetry-dev/tanstack-ai (overrides), and MediaWrapperOverrides, have no caller outside the tests. Two options: document them as public options, or change the tests to decode the OTLP requests from a fetch spy and remove the parameter in a major release.
• postOtlp and maybeGzip are public exports of @telemetry-dev/otel, but only the package itself uses them. A removal is a breaking change.

Coverage:

• No test runs the ai@6 integration through the real ai-v6 dispatcher. Every ai@6 test sends hand-built events. Done in test(ai): cover the AI SDK paths the test audit found #169.
• These AI SDK paths have no test: the gateway relabel in the ai@7 and media paths, the ai@7 step without onStepStart, executeTool, wrapGenerateVideo, wrapGetBatchStatus, wrapCancelBatch, the transcription duration, the video counts, and the wrapper passthrough without a key. Done in test(ai): cover the AI SDK paths the test audit found #169.
• The onConfig phase guard in packages/tanstack-ai/src/middleware.ts has no test. Supported engines do not call onConfig after onStart. Done in test(sdk-tanstack-ai): cover the onConfig phase guard #162.
• No release check makes sure that each exports and publishConfig.exports target is in the packed tarball. Done in ci(sdk): check that each npm tarball has the targets of its exports #155 and ci(sdk): fail npm tarballs with no export or bin targets, and print the counts #160.
• The Python core has no wire test for on_error with a server that returns an error. Done in fix(sdk-python): report a series of failed exports once #87.
• In the ai@7 integration, the object and embed step ends set the provider on the step's duration and token metrics, and no test checks it. If either recorded the end event's raw provider, the metrics would say gateway instead of Vercel AI Gateway and every test would still pass. The gateway table that test(ai): cover the AI SDK paths the test audit found #169 added to packages/ai/tests/integration_v7.test.ts could check each metric's gen_ai.provider.name. Done in test(ai): check the gateway relabel on the ai@7 step metrics #209.
• In the ai@7 integration, the execute_tool duration metrics take their provider from the root (metricBase in recordQueuedMetrics in packages/ai/src/integration_v7.ts), and no test checks it. If that provider were dropped or recorded as gateway, every test would still pass. No row of the gateway table runs a tool, so a row with a tool call could check it.

Python compatibility code:

• The Bedrock, Google GenAI, LiteLLM, and OpenRouter integrations keep fallbacks for span handle methods (capture_masked, record_output_chunk) that each supported core version has. The Anthropic and OpenAI integrations no longer have them.
• CI tests each Python package against the core in this repository, also in the lowest-bounds job. If the telemetry-dev floor of an integration goes below the version that has these handle methods, no test fails.
• In the Python core, CaptureBudget.from_client and the public metrics_recorder parameter of TelemetrySpanProcessor have no caller.