feat: enriched training data (125 examples, 7 answer templates) + 8-layer LoRA, val loss 0.122

- Enriched training data: 125 examples (65 QA + 60 negatives)

- 7 unique negative response templates (was: 1)

- File-path-specific QA responses with commit subjects

- Increased LoRA: 8 layers (2.6M params), 300 iterations

- Val loss: 0.192 -> 0.122 (best yet, vs PR#48 0.125)

- Model now produces context-rich responses with commit history and actionable advice

Author: bearmug