Fix metric bugs

[gabegma]

Summary

Bug fixes hit while debugging a stuck rerun loop on a remote benchmark job.

1. Judge response parsing — 8e6342b

parse_judge_response used extract_and_load_json (first match wins). For metrics like faithfulness, judge responses often contain incidental JSON fragments in prose (e.g. [] from tool-call references), so the first object wasn't the real answer.

Now iterates every JSON value, flattens lists, and returns the largest dict — preferring ones with a top-level rating. Adds a unit test for the multi-object case.

2. Pipeline robustness on incomplete trial dirs — 3aa4073

• _load_context substituted "{}" for a missing result.json, producing a cryptic Pydantic error. Now raises FileNotFoundError like the existing scenario_db checks.
• _run_and_pipeline's outer except Exception: raise was both hiding bugs and cancelling the whole iteration on a single corrupt artifact. Now catches FileNotFoundError / json.JSONDecodeError narrowly (record marked invalid, iteration continues); other exceptions still propagate.

3. response_speed skips on degenerate input — d3dc008, 9bb88e5

Both error branches (all latencies filtered out by the 0 < v < 1000s sanity range, or latency_assistant_turns empty because the agent never produced an overlapping turn) now return score=None, skipped=True. response_speed is diagnostic — these cases are already flagged by task_completion, conversation_progression, turn_taking.

4. turn_taking scores 0 on no-overlap — 9bb88e5

When no turn has both user and assistant audio (agent didn't participate), this is a turn-taking failure, not missing data. Now scores 0.0 instead of erroring.

Test plan

• pytest tests/unit/metrics/{test_metrics_utils,test_response_speed,test_turn_taking}.py
• pytest tests/unit/orchestrator tests/unit/metrics/test_runner.py tests/integration/test_evaluation_mode.py — 94 tests green