Explain Agent Eval Failures Without Leaking Tests

Implement summarize_failures(failures) that produces a readable summary of hidden-test failures without exposing the raw detail text. failures is a list of dicts: {"category": str, "detail": str}. The detail field may contain sensitive test content that must not appear verbatim in the output. Return a dict mapping each category to a summary dict: {"count": int, "failure_modes": int, "sample_modes": [str]} where failure_modes is the number of distinct failure detail strings and sample_modes is up to 3 example snippets (first 50 chars) to help debug without revealing the complete hidden test.