Release v0.1.0

[0.1.0] - 2025-10-09

Added

• Initial release of rlbidder
• Rust-powered data pipeline with Polars Lazy API for processing massive auction logs
• SOTA offline RL algorithms: IQL, CQL, Decision Transformer, GAVE, GAS
• Classic control baselines: BudgetPacer, PID, Fixed CPA, Stochastic CPA
• Modern transformer stack with FlashAttention (SDPA), RoPE, QK-Norm, SwiGLU
• HL-Gauss distributional RL implementation
• Efficient ensemble critics using torch.vmap (faster than loop-based)
• Numerically stable stochastic policies (SigmoidRangeStd, BiasedSoftplus)
• Parallel online evaluation with ProcessPool and round-robin agent rotation
• Interactive Plotly dashboards for campaign monitoring and market analytics
• RLiable metrics for statistically robust algorithm comparison
• PyTorch Lightning training infrastructure with multi-GPU support
• Draccus type-safe dataclass-to-CLI configuration management
• Local experiment tracking with Aim
• Decision Transformer inference buffer for online deployment
• Comprehensive data preprocessing pipeline with scikit-learn-style transformers
• Production-ready auction simulator with multi-agent support

Features

• Data Processing: Symlog, Winsorizer, ReturnScaledReward transformers
• Evaluation: Multi-seed, multi-period evaluation with parallel workers
• Visualization: Campaign health, budget pacing, market structure, scatter analysis
• Scalability: Stream processing, mixed precision training, gradient accumulation