smoke — minimal tensor/autograd sanity demo

The smallest complete Fucina program (main.zig): two tracked variables — x = [2, 3] (shape 1x2) and w = [4, 5] (shape 2x1) — a dot contraction, a sumAll loss, one backward(), and the gradients read back out. It exercises ExecContext, tagged-dimension Tensors, the autograd graph, and gradient retrieval in under 50 lines.

No weights, no flags (zig build run is kept as an alias for the same step):

zig build smoke

Expected output (loss = 24 + 35; grad_x = w, grad_w = x):

loss=23
grad_x=[4, 5]
grad_w=[2, 3]

Shared knobs

The ReleaseFast/-Dcpu build discipline and global thread/BLAS knobs are documented once in docs/RUNNING-MODELS.md; this demo is instant either way and takes no arguments.