docs: refresh brittle prompt examples for newer models

[voidborne-d]

Summary

• refresh lessons 3 and 4 across all notebook variants so they no longer assume the baseline prompt will always fail on newer Claude models
• reframe the exercises around reliability and reasoning quality, which still demonstrates the prompting technique even when the first answer is correct
• add a small validation script to catch future regressions in the model-drift notes

Testing

• python3 scripts/check_model_drift_notes.py

Fixes #86