feat(m3-real-calibration): add batch mode to run-calibration.js, run real calibration on all 6 pools, replace heuristic baselines with measured values

- Add --batch flag with incremental progress-saving to partial-results file
- Add --label-field arg for explicit label field selection
- Raise default K_TRIALS to 5 for finer per-item difficulty resolution
- Add prompt truncation for long code-context prompts (>1600 chars)
- Fix ANSWER: format vs JSON instruction conflict in buildCalibrationPrompt
- Fix judgePrediction to handle Yes/No string predictions for boolean labels
- Run real stratified-sample calibration on all 6 pools (K_TRIALS=5, CONC<=4, solver=claude-sonnet-4-6)
- Replace all heuristic baseline values with measured_sample values (n=10-25 per pool)
- Record explicit exemptions for all 6 pools: claude-sonnet-4-6 achieves ceiling-level baseline performance (>70%) on all candidate pool items, leaving insufficient headroom for skill improvement measurement
- All decisive split rules preserved: >=1 target + >=1 distractor + 0 ambiguous