ILT Checkpointing & RDP Vertex Decimation: Survive Preemption, Cut OASIS Bytes
`LevelSetILTModel.predict(checkpoint_dir=, save_freq=, resume_from=)` periodically `torch.save`s the mask logit, Adam state, and best-loss tracker — SLURM preemption or a CUDA crash on a multi-thousand-iter run no longer wipes prior progress, and resume-vs-uninterrupted determinism is pinned by test. `export_oasis_mbw(vertex_tolerance_nm=...)` runs anchored Ramer-Douglas-Peucker simplification on each sampled curvilinear polygon — default `0.0` keeps bit-exact academic behaviour; positive values cut full-chip OASIS data volume and the MBMW shot/byte budget without measurable wafer-image change.