fix(f1_radio_rag): give the assertion judge the retrieved messages

[fschlz]

Problem

The f1_radio_rag test suite checks groundedness assertions — e.g. "grounded in the provided radio messages", "does not invent events absent from the messages", "refers only to drivers named in the messages".

But Opik's assertion judge (run_tests) builds its prompt from only the trace input and output. The task returned the retrieved messages as a separate top-level context key:

return {"input": query, "output": output, "context": context}  # context never reaches the assertion judge

So the judge saw query + summary but never the messages it was supposed to check against. Every groundedness assertion was effectively unverifiable → trivial passes and unstable verdicts.

(The evaluate() path was already fine: ContextRecall / Hallucination declare context as a first-class argument and read the top-level key directly.)

Fix

Add _suite_task, which folds the retrieved messages into input so the judge can see them, and point run_tests at it:

def _suite_task(item: dict) -> dict:
    result = answer(item["query"])
    return {
        "input": {"query": result["input"], "messages": result["context"]},
        "output": result["output"],
    }

• evaluate() keeps _rag_task — the metrics read top-level context, so they are unaffected.
• expected_output is deliberately kept out of input (per the Building Test Suites docs), so the judge can't use it to pass assertions that should fail.
• README documents that the assertion judge reads only input / output, hence the fold.

Verification

• python -m py_compile clean on evaluation.py.
• uv run f1rag eval dry-run path unchanged (no creds needed).
• With creds, the groundedness assertions now receive the radio messages in the judge prompt.

🤖 Generated with Claude Code