Predicts the query's fingerprint from its peaks and ranks each candidate
SMILES by Tanimoto agreement. The ranking rests only on the subset of bits
the predictor actually learned; the rest sit at an untrained 0.5 prior and
would otherwise order candidates by molecular size. That subset is a small
fraction of the fingerprint, so `bit_coverage` in the response is the
authoritative live number — read it, and the `warning` that accompanies a
low one, before acting on the order.

Measured on open structure retrieval against MassSpecGym candidate pools
(median 256 candidates): the correct structure comes first about 3.5 % of
the time, against a 2.4 % permutation noise floor. The lift is real but
roughly one percentage point, so the right answer is NOT at rank 1 in about
96 % of cases. Use it to cut a long candidate list down, never to name a
compound.

If you already suspect what the compound is, use covams_search_db instead:
matching the spectrum against the packaged 169k-spectrum library is a far
stronger signal, and dereplicating a known compound is the common case.