Explicit product-family verification, 2026-09-14.

- cargo check --workspace --tests: passed, no warnings.
- cargo test -p ipu-codegen --release: 314 passed, 5 ignored; one doctest passed.
- New semantic profiling regression sums useful and physical FLOPs across
  packed batch matrices and checks that splitting does not multiply the count.
- Small 64-tile F16 MLP (Repeat 3) and FP8 ViT (3 layers, fused QKV) both passed
  hardware/reference checks. Commands, logs and binary profiles are adjacent.
  Numerical and modeled cycle results are unchanged from copy-compute:
  MLP max abs 0.000183, modeled total 32613 / exchange 18087;
  ViT cosine 0.997459656, max abs 0.189453, modeled total 331874 / exchange 120708.

These are randomized semantic fixtures, not trained full-size model performance
measurements. No kernel instruction implementation changed in this slice.