[GSoC] Build NER training dataset from annotated rules

Tasks:
- Parse .RULE files and extract extract required phrase markers
- Propagate annotations (with CLI commands from #3924 ) to unannotated rules (with matching text) to expand the dataset
- Improve composite rules (AND/OR expressions where some aren't marked)
- Convert to BIOES labels
- Split into train/val/test by rule family
- Validate dataset quality

Timeline: Weeks 1-2 (current)

Deliverable: BIOES dataset ready for training (train/val/test splits)