Training moves to TPU

The Kerven Vega 1 training pipeline now runs on a TPU v5e-8, granted through Google's TPU Research Cloud.

The practical reason is dull and decisive: PyTorch's TPU support is a translation layer, and JAX's is not. Rewriting the fine-tuning loop in JAX and Flax cost a week and bought back the compute budget of a small company.

Synthetic data generation is next. Twenty to thirty thousand examples of AI-written code with real, injected, verifiable bugs, then supervised fine-tuning, then GRPO against a reward that can actually be checked.