The content-hash-in-telemetry point is underrated, once every trace carries the prompt hash, "why did quality drop last Tuesday" turns from a research project into a filter query. I also agree with the two-layer test split, deterministic golden renders on every commit and non-deterministic eval runs only on prompt or model changes matches what has actually held up in prod for me. One thing I would add is keeping the eval set versioned alongside the prompts, otherwise a passing suite six months later means something completely different.