fix(search-eval): skip llm-rubric assertions when judge API key is absent

[Copilot]

When OPENAI_API_KEY is unset, promptfoo throws on every llm-rubric assertion grading call. This doesn't just fail the LLM-judged metrics — it wipes out all namedScores for the affected test, including the deterministic IR metrics (MRR, Recall@k, nDCG@k). Result: 59/59 tests counted as errors, all metrics at 0.000, spurious 100% regression detected, job exits 1.

Changes

• judges.ts — adds two helpers:

  • judgeApiKeyVar(judge) → returns the env var name (OPENAI_API_KEY, ANTHROPIC_API_KEY, GOOGLE_API_KEY) for a resolved judge
  • isJudgeKeyAvailable(judge, env?) → returns true iff that key is present and non-empty

• run.ts — instead of always passing the static promptfoo.yaml, now generates a runtime config at results/promptfoo.runtime.yaml (already gitignored) before each promptfoo invocation:

const judgeAvailable = isJudgeKeyAvailable(judge);
if (!judgeAvailable) {
  console.warn(`⚠  ${judgeApiKeyVar(judge)} is not set — LLM-judged metrics will be skipped.`);
}
writeFileSync(PATHS.runtimeConfig, buildRuntimeConfig(judgeAvailable), "utf8");

When judgeAvailable is false, buildRuntimeConfig produces a config without the three llm-rubric assertions and the defaultTest.options.provider block. The deterministic javascript assertions (ir-metrics, structural) always run regardless of key availability.