v0.9.4-benchmark
Validated speedups vs Stan: - H (hand-coded): ~9x (2 configs) - A (autodiff): 3.8x-27.6x (33 configs) - N (numerical): GP deferred (5 configs) Features: - Native HMC/NUTS with autodiff gradients - ICAR/BYM2 spatial (13-22x faster than Stan) - RW1/RW2/AR1 temporal (6-14x faster) - ZI/Hurdle models (5.7-10.3x faster) - benchmarks/ folder for reproducible validation Known issues: - GP spatial heisenbug (see feature/gp-heisenbug branch)