`experiments_json` is a JSON object keyed by nucleus, e.g.
{"1H": [{"ppm": 1.18, "num_protons": 3, "multiplicity": "t"}, ...],
 "13C": [{"ppm": 18.0, "multiplicity": "CH3"}]}.

Give `molecular_formula` when you have it from HRMS. Without one the search is
unconstrained and much weaker -- measured on a toluene peak set, the correct
structure does not reach rank 1.

Measured limits, so you do not spend credits discovering them: candidates are
assembled from a fragment library and capped at two fragments each, so fused
ring systems are not reachable (naphthalene returns zero candidates) and
three-fragment targets return a wrong top hit (paracetamol). It has been
exercised on synthetic peak lists, not yet on measured spectra.

Scoring follows `nucleus`. If that nucleus is absent from `experiments_json`
another one is substituted rather than failing, and the response says so in
`nucleus_scored` / `nucleus_source` / `nucleus_warning`.