context-graph-eval: judge scores were attributed to the wrong questions

[antejavor]

Part of #297

What was wrong

runner._judge_group paired deepeval's result.test_results with the input goldens by position (zip(goldens, result.test_results)). deepeval returns test results in completion order under async concurrency (AsyncConfig(max_concurrent=4)), so every question was scored with some other question's metrics.

Evidence

• Reproduction: a fake metric with random latency, 12 cases, concurrency 4. They came back in the order q-1, q-2, q-3, q-5, q-0, q-8, …. The new regression test fails on the old code and passes on the fix.
• In saved runs: 11–24 questions per 100-question run had a Contextual Recall reason quoting a different question's expected answer. This held for both a Sonnet and an OpenAI judge, and for text-search, graph-agent and hybrid runs alike.

Impact

• Run totals were roughly right: a permutation within each judged group (abstention and non-abstention are judged separately) preserves the count. They aren't exact, because enforce_retrieval_floor then zeroed scores against the wrong questions' payloads.
• Every per-question and per-type result was wrong. That includes eval: the judge scores identical answers differently, which is what makes the passing set unstable #324's per-question pass rates and its "zero questions passing in all three runs" reading, which should be re-measured.

Fix

PR #373, 7aa430b: test cases carry the golden's name, and results are matched by it.