v0.6: +15.2% accuracy with robust extraction (90.2% overall)

Key improvements over v0.4:
- Adversarial %: 71% → 93% (+22%)
- Multi-step: 32% → 92.3% (+60.3%)
- Position Sizing: 81.7% → 98.3% (+16.6%)
- Overall: 75% → 90.2% (+15.2%)

Fixed evaluation extraction bugs that caused 17% false negative rate.

adapter_model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:2eddb001a78738266b300b45fb2f9cd2116bccb31ea348b681a84c91dd8313bc
+oid sha256:3370282a181685795e77e0ac9e5f0d6883e87fc87fff6f32f836da85b18dfa83
 size 533009160