fix(adk): fix llm span metadata (provider, tools, errors)

[starfolkai]

Summary

Brings the Google ADK integration in line with .agents/skills/sdk-integrations/SKILL.md and cuts the test suite over to real VCR-backed coverage.

Integration behavior (tracing.py)

• metadata.provider="google" and metadata.tools on every llm span. Tool definitions are lifted out of input.config into metadata.tools, preserving Google-native shape.
• _capture_config is now a strict allowlist — no more dict(config) pass-through. Fields outside the allowlist (including tools/tool_config) are dropped from input.config automatically. _extract_tool_metadata is the single source of truth for what goes into metadata.tools.
• error=e instead of error=str(e) in the tool and MCP-tool wrappers — the Exception instance flows through to the log pipeline.
• Trim raw ADK objects from span metadata. invocation_context, model_response_event, parent_context, and the **kwargs spread across every wrapper are dropped; only small named fields (app_name, user_id, session_id, state_delta, agent_name, flow_class, tool_class, model, provider, tools) remain.
• Drop _get_field and bt_safe_deep_copy(tools). Production ADK always passes Pydantic objects, so plain getattr(x, field, default) is enough. The log pipeline serializes Pydantic tool objects at send time (SKILL: "do not over-serialize").

Tests

• test_adk_braintrust_integration now filters spans by context.span_origin.instrumentation.name == "adk-auto", asserts {task, llm, tool} types are covered, and validates metadata.provider, metadata.tools, and no tool leakage into input.config.
• test_adk_captures_metrics asserts metadata.provider="google" and the direct_response call-type span name.
• test_adk_max_tokens_captures_content now also asserts the sampling config (max_output_tokens, temperature) is captured in the LLM span input.
• test_adk_sync_runner_run_does_not_duplicate_invocation_spans now also asserts thread-bridge context propagation — every adk-auto span shares the outer parent's root_span_id.
• Delete test_adk_mcp_tool.py entirely (the two remaining patcher-wiring tests were duplicative of the auto-instrument subprocess test) and drop its line from noxfile.py.
• Delete fake-LLM, mock, and non-ADK unit tests — 15 tests removed. Every branch they covered (_determine_llm_call_type, _capture_config allowlist, _serialize_content/_serialize_part, _serialize_pydantic_schema, bt_safe_deep_copy, sampling config, thread bridge, PDF attachment, MCP-in-flow parenting, async context) is exercised by an existing VCR test.

Final test count: 13 tests, 12 VCR-backed + one plain-Python regression for the _create_thread_wrapper double-invoke bug.

Test plan

• nox -s "test_google_adk(latest)" — 13 passed
• nox -s "test_google_adk(1.14.1)" — 13 passed
• No cassettes re-recorded (one orphaned cassette, test_adk_agent_metadata_with_attachment.yaml, removed on both matrix versions)

🤖 Generated with Claude Code

Created by abhijeet

Slack thread