docs: add LoRA-per-task design document

[abrichr]

Summary

• Literature-backed design for task-specific LoRA adapters with runtime routing
• Covers architecture (registry, router, PolicyAgent with dynamic LoRA), training pipeline, tiered data collection strategy
• Identifies correction flywheel (PR openadapt-evals#116) as natural training data source: corrections → SFT data → LoRA
• Compares LoRA-per-task vs demo-conditioning with hybrid strategy
• 30+ paper references (ShowUI-Aloha, SeeClick, LoRAHub, S-LoRA, etc.)
• Positioned as one experiment track within the broader experimentation framework

Test plan

• Review doc for accuracy and completeness
• Validate against current codebase (existing infra referenced)

🤖 Generated with Claude Code