covnmr_verify without the file: this closes the loop from pasted
publication data (covnmr_parse_nmr_string) or a peak list held in the
Sketcher to a structural verdict, with per-atom assignments. The verdict
and `evidence` are the same as covnmr_verify's (read those, not
`match_score`, which is coverage). Give intensities when the peaks were
picked from a spectrum: weak lines are then treated as noise; without them
every listed line is taken as the compound's, and a structure that leaves
most of them unexplained is refuted.

`peaks_json` is a JSON array of peak dicts, e.g.
[{"ppm": 58.4, "nucleus": "13C"}, {"ppm": 18.2, "nucleus": "13C"}].
`experiments_json` is the alternative shape, a JSON object keyed by
nucleus: {"13C": [...], "1H": [...]}.

THE PEAK LIST OUTRANKS `nucleus`. A list parsed from a "13C NMR ..." header
carries its own nucleus, and that is evidence the way a FID's pulse
sequence is. Measured on ethanol's own literature 13C list: scored as the
default 1H it returns match_score 0.000; scored as the 13C the peaks declare,
1.000 coverage and carbon evidence. The response always reports which list was
used in `nucleus_scored` / `nucleus_source` / `nucleus_warning`. Set
`nucleus` only to state one the data does not carry.

Bounded: `unassigned_peaks` at `max_unassigned` (500) and the incoming list
at `max_peaks` (2000). Both untruncated counts are reported alongside
(`n_unassigned_peaks`, `n_peaks_supplied`), so a clipped list can never be
mistaken for a short one.