Two-Day Kernel Debug Traced to a 10x Learning Rate Mismatch in Defaults

A developer spent two days debugging a browser-based machine learning model whose training loss plateaued at 2.45, initially suspecting WebGPU kernel issues. The root cause turned out to be a default learning rate mismatch: the browser used 3e-3 while the Python reference used 3e-4, a tenfold difference. Kernel parity tests validated numerical operations but never compared configuration defaults, allowing a correctly implemented model to appear broken. The same investigation uncovered two more gaps — a tiny 863-byte default corpus masking model capacity, and a Memory64 benchmark that missed the actual browser bridge used in production. The fix involved adding end-to-end checks across configuration, data, and integration boundaries to ensure test coverage matched real-world workflows.