Implement batch judge calls

[ScuttleBot]

Priority 2: Batch Judge Calls (20-30% speedup)

Problem

Currently we make 98 separate API calls to the judge (one per task). Each call has network latency overhead.

Solution

Batch judge prompts for multiple tasks into a single API call.

Implementation

Modify _grade_llm_judge() to accept a list of tasks and return a list of results.

def _batch_grade_llm_judge(
    tasks: List[Task],
    execution_results: List[Dict],
    batch_size: int = 5,
    ...
) -> List[GradeResult]:
    # Build combined prompt
    prompt = "Grade the following tasks. Respond with a JSON array:\n\n"
    for i, (task, result) in enumerate(zip(tasks, execution_results)):
        prompt += f"Task {i+1}: {task.task_id}\n{_summarize_transcript(result)}\n\n"
    prompt += "Response format: [{task_id, scores, total, notes}, ...]"
    
    # Single API call
    response = call_judge_api(prompt, ...)
    return _parse_batch_response(response)

Expected Impact

• API round-trips: 98 → ~20 (batch of 5)
• Network latency savings: ~30-60 seconds per model

Testing Plan

1. Run sample of 10 tasks single vs batched
2. Compare scores for consistency
3. Validate JSON parsing handles batch responses

[ScuttleBot]

Closing — PR #354 closed, parallel judge execution (#372) merged instead.