Broaden the memorizing aroma heads, and measure whether it worked

[rvnminers-A-and-N]

Closes #256.

Two things here: a measurement instrument, and the data work it made possible.

The instrument

training/audit_generalization.py fires every head over the whole corpus and counts the hits that were not in its training set. That count is the head's discovery power, and a cross-validated AUROC cannot substitute for it — CV only ever asks about molecules inside the labelled set, so a head fit on twelve molecules sharing one scaffold scores 0.99 while discovering nothing.

Scoring at one threshold turned out to be wrong, and this is the part worth reading. A head with 13 positives against 2400 negatives is calibrated conservatively even with balanced class weights, so novel@0.5 = 0 conflates two opposite problems:

• learned nothing — the real failure, fixed with structurally diverse positives
• learned the class but is shy about saying so — fixed with more positives, or nothing at all

The audit now probes at 0.35 as well and returns a three-way verdict. The distinction changed the answer materially:

verdict	count
generalizes at 0.5	153	median 6 novel discoveries
under-confident	12	found real unlabelled molecules just below the bar
memorizing	2	turmeric, celery — nothing at any threshold

Read at 0.5 alone, all 14 looked like the same problem. pine finds 8 unlabelled molecules at 0.35, rosemary 9, freesia 9 — those heads had learned their class all along.

The data

All twelve heads named in #256 had no entry in CURATED at all — their positives came solely from the weak-labelled HSDB odor corpus, which is exactly why they were thin enough to memorize. Added well-established character-impact chemistry for each, chosen for structural spread rather than volume (the #247 method): pinenes/camphene/carene/bornyl esters for pine, cineole/terpinen-4-ol/aromadendrene for eucalyptus, phthalides for celery, indole/cresol/benzyl esters/cinnamyl alcohol for narcissus.

Result, stated honestly:

• five moved off zero at 0.5 — narcissus 0→4, frankincense 0→2, costus 0→2, fennel 0→1, allspice 0→1
• five were under-confident, and sharpened — pine, eucalyptus, rosemary, freesia, elemi
• two did not move at any threshold — turmeric, celery. Both are genuinely narrow single-scaffold classes (bisabolanes, phthalides). More molecules will not fix that; it needs a better model, not a longer list.

AUROC fell for several heads (pine 0.965→0.905, frankincense 0.916→0.829). That is the right direction — a broader class is harder to fit, and a score dropping while discovery rises is what learning instead of memorizing looks like.

Corpus now trains 167 aroma heads, up from 164; no head was dropped.

Also

docs/METHODS.md still described the aroma model as deferred to licensed odor data, several hundred heads after it shipped. Corrected — what remains gated is depth, not vocabulary — and documented the two-threshold method alongside it.

Provenance

All added associations are public-domain flavour/essential-oil chemistry resolved to canonical SMILES via PubChem, consistent with the existing supplement. No restricted compilation was consulted; the excluded-sources list in docs/DATA-SOURCES.md is unchanged.

Verification

• ruff check training/ tests/ clean
• full suite: 43 passed
• downstream artifacts rebuilt against the new heads: master_enrichment (8,855 rows), profile_index (8,836 × 178 dims — 167 aroma + 6 taste + 5 mouthfeel), flavor_map (8,836 points)
• needs sudo systemctl restart flavormancer to deploy — the running app still holds the old 164 heads in memory

Filed #257 separately: reading pine's raw positives surfaced chlorinated pesticides, and 105 industrial organohalogens are training flavor heads corpus-wide. Not touched here.