Summary

• Adds --metrics CLI argument to lightspeed-eval that filters each turn's turn_metrics to only the specified metrics
• Example: --metrics custom:answer_correctness runs only correctness, skipping all RAGAS metrics
• Works like --tags and --conv-ids — filters after loading, before validation

Motivation

Running the full metric suite (5 metrics per question) takes ~37 minutes. When we only need answer correctness for comparison runs, this drops to ~6 minutes without editing YAML configs.

Changes

• runner/evaluation.py: Add --metrics arg to argparse, pass to load_evaluation_data
• core/system/validator.py: Add metrics parameter, filter turn_metrics lists after scope filtering

Usage

# Run only answer correctness (skip RAGAS metrics)
lightspeed-eval --eval-data config/CLA_tests.yaml --metrics custom:answer_correctness

# Run two specific metrics
lightspeed-eval --eval-data config/CLA_tests.yaml --metrics custom:answer_correctness ragas:faithfulness

# No flag = run all metrics (existing behavior, unchanged)
lightspeed-eval --eval-data config/CLA_tests.yaml