model(mouthfeel): broaden the tingling head — it barely generalizes

[rvnminers-A-and-N]

The generalization test (fire the head over all 8.8k molecules, count hits that were NOT in its training set) exposed that tingling was memorizing: trained only on Zanthoxylum sanshools it produced 0 novel hits — a lookup table wearing a 1.00 AUROC.

Broadening the positives with paresthesia amides from other genera (Echinacea tetraenoic isobutylamides, Anacyclus, Heliopsis, Piper — deliberately varied chain lengths and unsaturation) took it from 11 to 18 positives and 0 → 2 novel hits (one of them piperine). Real generalization, but a thin margin, and still the weakest of the five:

head	fires	trained	novel
cooling	32	14	16
pungent	24	12	12
warming	21	12	9
astringent	19	11	8
tingling	16	18	2

Next steps to consider: more structurally-varied public paresthesia amides (further Echinacea/Acmella/Zanthoxylum species, synthetic tingle agents with food clearance); check whether the physicochemical block is even helping here vs a pure fingerprint; and consider whether a shared alkylamide structural feature would let tingling borrow signal from the pungent/warming amides. Documented as a known limitation in docs/MOUTHFEEL.md.

[rvnminers-A-and-N]

Re-measured with the committed tool from #258 (training/audit_generalization.py --mouthfeel), which now reports discovery at two thresholds:

head	AUROC	n_pos	fires	novel @0.5	novel @0.35
tingling	0.999	18	20	2	4
astringent	0.950	11	19	8	24
warming	0.984	12	21	9	13
pungent	0.970	12	24	12	14
cooling	0.937	14	30	16	39

All five generalize — the whole modality is off zero. The Echinacea/Anacyclus/Heliopsis/Piper amides did their job: tingling went from memorizing to finding 2 molecules nobody labelled.

Keeping this open rather than closing it. Two novel discoveries is real but it is the thinnest head we ship, and it sits well behind cooling at 16. The paresthesia-amide class has more public chemistry to draw on — Zanthoxylum beyond the sanshools already listed, more Piper and Achillea alkylamides, and the synthetic tingle agents with food clearance. Worth another round of positives to see whether it can reach the pack.

Live check on the deployed app: spilanthol reads tingling 1.0, confident — so the head is firing correctly on its canonical agent, not just scoring well in cross-validation.