Define next-step research ideas

## Summary
Document promising research directions discovered during MVP development for future exploration.

## Success Criteria
- [ ] Ideas collected from experiment insights
- [ ] Each idea has clear hypothesis
- [ ] Rough effort/impact estimates
- [ ] Prioritized list created
- [ ] Documented in docs/research-directions.md

## Candidate Directions (to be refined)

### Model Improvements
- [ ] Adaptive confidence threshold (learned, not fixed)
- [ ] Diff-conditioned generation (use git diff as input)
- [ ] Multi-language support
- [ ] Longer context (CHANGELOG generation)

### Data Improvements
- [ ] More diverse data sources
- [ ] Synthetic data augmentation
- [ ] Active learning for edge cases

### Deployment
- [ ] Git hook integration
- [ ] IDE plugins (VSCode, JetBrains)
- [ ] CI integration (auto-generate from PR diff)

### Experiments
- [ ] User study on commit quality
- [ ] A/B test: human vs ccgram commits
- [ ] Ablation: which n-gram order matters most

## Format
Each idea should include:
- **Hypothesis**: What we think will happen
- **Experiment**: How to test it
- **Effort**: Low/Medium/High
- **Impact**: Low/Medium/High