Changelog

A minor release in which the CPU view work of 3.12.0 reaches transposes and slices along any axis: x.T, x[:, 1:3], split along a later axis and expand share their input's buffer, and every op reads a strided view correctly. Alongside it the wheel's floor moves down — one wheel now covers macOS 15 and every later release — and a batch of silently wrong results is fixed. - CPU transposes and slices are views. permute, transpose, .T, .mT, a slice along any axis (x[:, 1:3], x[:, i]), split / chunk / unbind / narrow on a later axis, and expand return views of their input's buffer. A write through one reaches the tensor it came from, and autograd follows a recorded write through it (CopySlices splices by element positions as well as runs). expand's view repeats elements, so a write through it is refused. Metal keeps copy semantics, and diagonal, unfold, broadcast_to and step slices still copy. - Every op reads a strided view correctly. TensorImpl::storage() hands a non-dense CPU tensor a packed copy of its elements, cached against its version; the audit's layout axis found about a hundred ops reading the wrong bytes before that, and none after. - macOS 15 is the floor. The wheel is built for macOS 15.0 and installs on 15 and later. CI builds and tests on a macOS 15 runner on every push, and a release publishes only once the built wheel has been installed and exercised there. - `arange` of integers gives int64, as the reference does, and ranges past 2^53 are built exactly. - Mixed int and float operands meet at their common dtype in cat, stack, where, clamp, outer and einsum instead of raising — and clamp(int_t, 0.5, 2.5) no longer answers in int64. - `lucid.tensor(t)` copies and honours dtype= and device=, F.one_hot returns int64, and the distributions derive probs from logits and back. - float64 `erfc` keeps double precision, and erfc holds its relative precision in the tail. - CPU complex arithmetic works: +, -, /, broadcasting, and casts between real and complex.