The raw-slice score collapse suggests checking vector health before a row enters the index. Finite values, expected dimensions and a nonzero norm can be asserted without knowing whether a query will retrieve the right snippet; a NaN vector should become an explicit embedding failure, not a searchable record.

Would you add one deliberately invalid vector and a near-zero vector to that smoke test? Pairing those with the working float32 path would show that the indexing boundary rejects broken numerical output, rather than relying on a later similarity cutoff to hide it.