A reproducible test: Claude reviews the same PR twice, with and without context

Most of what I read about LLM code review is "we tried it, it works." That's not enough to be useful. I want to know: same model, same PR, two different review contexts - does the output actually chan