v4.1.2

Fixes

• Training pipeline: Resolved UnboundLocalError on X variable after del — training was completing successfully but failing at the metadata save step, causing the entire run to report failure
• Lint: Removed unused mem_model_mb variable (ruff F841)
• Model card: Updated Hugging Face model card with correct v4.1 architecture specs (45M params, 384 hidden dim, 6 heads/2 KV heads, 4 experts)

What's included

• Mamba-2 SSM + GQA + MoE hybrid architecture (~45M parameters)
• BalancedDirectionLoss (Huber + BCE) for joint regression and direction accuracy
• EMA weight averaging, cosine warm restarts, gradient accumulation
• Automated training via GitHub Actions every 6 hours
• 44 technical indicators, multi-timeframe data ingestion
• Models deployed to Hugging Face Hub: meridianal/ARA.AI