@@ -9,10 +9,12 @@ Most recent change on the top.
 ## Unreleased
 
 ### Changed
+- [Breaking] `matscipy>=1.3.0` is now required, because `matscipy<1.3.0` returns nonzero cell shifts along nonperiodic directions for atoms outside the provided cell (https://github.com/libAtoms/matscipy/pull/315), giving wrong edge vectors from the default `matscipy` neighborlist backend for such nonperiodic systems. Since `matscipy>=1.3.0` requires NumPy 2, NumPy 2 and PyTorch >= 2.3 (the first release supporting NumPy 2) are now also required.
 - `nequip-compile --device cpu` now defaults to scalar C++ codegen (Inductor's `cpp.simdlen=0`) instead of vectorized codegen, which is 1.5-2x faster for NequIP and 6-11x faster for smaller Allegro models in CPU inference, and avoids the CPU AOTInductor compilation failure on PyTorch 2.10/2.11 entirely. Controllable with `ATEN_CPU_CAPABILITY=<your ISA>` or `--inductor-configs cpp.simdlen=<width>`.
 - PyTorch compatibility docs: the CPU + AOTInductor compilation failure affects both PyTorch 2.10.0 and 2.11.0 (fixed upstream in 2.12), now avoided with default NequIP scalar codegen compilation.
 
 ### Fixed
+- e3nn modules built after any NequIP model on PyTorch >= 2.10 being compiled with the deprecated `torch.jit.script`: `conditional_torchscript_mode` restored only e3nn's legacy `jit_script_fx=True` flag on exit, which e3nn maps to `jit_mode="script"`, overriding the `"eager"` default. `jit_mode` itself is now restored
 - `--inductor-configs` values were passed to `torch` as strings, so numeric options mis-compared. Values are now YAML-parsed, as for `nequip-package modify` modifier kwargs.
 
 ## [0.19.1]