feat(evals): add LLM-powered failure analysis to eval CI

[Maahir Sachdev (maahir30)]

When evals fail in CI, it's often unclear why from the raw pass/fail counts alone. This adds a lightweight post-eval step that uses an LLM to analyze each failure's trajectory and produce a human-readable explanation, surfaced directly in the GitHub Actions summary.

What changed:

• Reporter extension (libs/evals/tests/evals/pytest_reporter.py): The pytest reporter now captures per-test failure details (test name, category, failure message including the agent trajectory) into a failures list in evals_report.json. Only populated for failures — zero overhead on passing runs.

• Analysis script (.github/scripts/analyze_eval_failures.py): A new script that reads the failures from the report, sends each one to an LLM (Claude Haiku by default) for analysis via asyncio.gather (all failures analyzed in parallel), and writes the results to $GITHUB_STEP_SUMMARY and a failure_analysis.json artifact. Exits immediately with zero cost when there are no failures.

• CI workflow (.github/workflows/evals.yml): Two new steps in the eval job — one to run the analysis script, one to upload the analysis artifact. Uses the existing ANTHROPIC_API_KEY secret, no new secrets needed.