Autoresearch-ANE slashes val_loss on Mac

This is a meaningful systems result, not just a prettier loss curve: the project gets much more real training done on consumer Apple hardware once compilation stops dominating the wall clock.

• –The repo says the dynamic pipeline compiles 10 ANE kernels once at startup, then stages weights dynamically, boosting throughput from roughly 120 to about 1340 steps per 5-minute run
• –The ANE backend is its own training stack in Objective-C with TinyStories and `val_loss`, so the numbers are impressive but not directly comparable to Karpathy's CUDA `val_bpb` baseline
• –Keeping the agent's edit surface mostly to `ane/experiment_config.h` makes autonomous overnight experimentation much more plausible on a laptop
• –If these gains hold up, Apple's Neural Engine starts looking less like an inference novelty and more like a viable playground for small-model research loops