PydanticAI adapter stringifies token usage instead of using the usage field

[aniketwattamwar]

Problem

sdk/python/src/ecp/adaptors/pydantic_ai.py already reads real token counts, then formats them into a string and appends it to evaluation_context:

usage_str = f"Usage: {usage.input_tokens} input, {usage.output_tokens} output tokens ({usage.requests} requests)"
self.captured_thoughts.append(usage_str)

This predates the structured usage field. Two consequences: the audit record's aggregated token totals stay empty, and graded text has token counts glued onto it — which can trip a does_not_contain grader for no reason.

Proposal

python Result(..., usage={"input_tokens": usage.input_tokens, "output_tokens": usage.output_tokens})

Acceptance criteria

• Usage reported via the structured field, not appended text
• Aggregated into the audit record's usage totals
• Fixture regenerated
• Audit other adapters for frameworks that expose usage

[mikemikimike]

Implemented and submitted as #31.

The PydanticAI adapter now returns input_tokens and output_tokens through ECP's structured Result.usage field instead of appending a Usage: string to evaluation_context. The seeded PydanticAI conformance fixture and adapter tests verify both the structured audit data and clean evaluator-visible context.

Validation completed locally:

• SDK adapter/conformance tests: 11 passed
• Full SDK test suite: 37 passed
• Runtime audit usage tests: 14 passed
• Ruff check: passed

The PR's CI and Tests workflows were created for head 3b73c397335b1f0d2e77de7f4c0230d9cf7a5c04, but both completed as action_required with no jobs, consistent with upstream fork approval gating. No code failure was reported; the run URLs and the formatter baseline limitation are recorded in the PR description.