v0.1.0 publishes the complete experimental result, not a production optimiser.
The compiler analyses deterministic torch.export graphs without model-specific annotations, derives named-input dependency regions, synthesises semantically equivalent execution strategies and gives reusable regions canonical cache identities. Portable tests cover guard preservation, every tensor boundary, aliases, effects, numerical configuration and adversarial cache collisions.
Automatic factoring succeeded exactly on CLIP and EfficientSAM: all 17 generated strategies for each model were bit-exact against their packaged monolith. The general runtime hypothesis failed. With correctness-safe pruning, CLIP retained 17 strategies and 22 canonical functions; process peak RSS reached 14,456,012,800 bytes while only three functions were live, slightly above the reproduced 14,405,255,168-byte eager control. Physical residency followed the packaged portfolio rather than live handles.
The surviving value is narrower: compiler analysis, exactness auditing, execution-plan diagnostics and experiments with a deliberately tiny fixed strategy set.
See docs/residency-negative-result.md and evidence/README.md for the claim boundary and raw evidence map.