Build the 73 indicative aroma heads up to confident (precision, not AUROC, is the bar)

[rvnminers-A-and-N]

Calibration in #261 surfaced this, and it is the single biggest quality signal we have on the aroma roster: 73 of 167 heads never reach 50% out-of-fold precision at any threshold. They fire well above base rate — real evidence — but they are right less than half the time when they fire, so they ship as indicative rather than confident.

The point of this issue is that indicative is a starting state, not a resting state. Every one of these should be built up until it can make a confident call.

Why AUROC hid this. AUROC is computed on ranking and is insensitive to class imbalance; precision is not. Both of these are true of the same head:

head	AUROC	out-of-fold precision
ginger	0.979	0.10
rosemary	0.897	0.06
passionfruit	0.895	0.06
woody	0.812	0.07
fatty	0.854	0.17
coffee	0.960	0.50 (just clears)

A head can rank its positives above its negatives beautifully and still be swamped by false positives, because there are 200x more negatives than positives. Nothing we had been reporting would have shown this.

What actually fixes it — in order of expected effect:

1. More positives per head. This is the same lever as model(aroma): 16 heads are memorizing, not generalizing — broaden their positives #256 and model(mouthfeel): broaden the tingling head — it barely generalizes #247, and it is the one that moves precision. A head with 11 positives cannot learn a decision boundary sharp enough to beat a 200:1 imbalance. ginger (11), passionfruit (10), violetleaf (11), woody (15) are the obvious first targets — all have deep, public, uncontroversial chemistry.
2. Better negatives. Right now every unlabelled molecule is an implicit negative, which is wrong: the HSDB corpus is sparsely annotated, so a genuinely gingery molecule with no odor text counts against the head. Curated hard negatives, or treating unlabelled as unknown rather than negative, would raise precision without touching the positives.
3. A better representation (research(model): GNN for structure→property vs the RandomForest heads #199). Where 1 and 2 stall, this is the GNN case: a 2048-bit Morgan fingerprint may simply not separate these classes. Precision on the indicative 73 is a much sharper benchmark for a GNN than AUROC, which the forests already score highly on.
4. Calibrated probabilities. CalibratedClassifierCV (isotonic/Platt) over the forests would make the scores mean what they say, which may let some heads clear the floor without any new data.

Explicitly NOT the fix: dropping the 73 heads, or hiding their molecules from chips, palette match and the map. They keep everything — they are simply labelled honestly. That is settled; see the reverted gating in #261.

Measuring progress: training/audit_generalization.py prints precision and the capable/indicative split per head. Target is the count of confident-capable heads going up from 94/167, with novel discoveries not collapsing — a head that reaches the floor by refusing to fire has not improved.

Baseline as of #261: 94 confident-capable, 73 indicative, median threshold 0.46.

[rvnminers-A-and-N]

Lever 4 (probability calibration) tested first, and it does not work here. Recording the negative result so nobody spends a day on it.

CalibratedClassifierCV (isotonic and Platt/sigmoid) wrapped around the same forests, scored out-of-fold on a sample spanning the full range of n_pos:

head	n_pos	raw precision	isotonic	sigmoid
fruity	101	0.40	0.62	0.60
floral	39	0.14	0.25	0.12
fishy	34	0.46	0.33	0.43
gardenia	29	0.33	0.21	0.33
marine	10	0.21	0.14	0.15
cardamom	10	0.12	0.50	1.00

The pattern is the problem: it lifts the one head that already has data (fruity, 101 positives, would clear the floor), makes several mid-size heads worse, and the cardamom 1.00 is noise — at 10 positives a single true positive reads as perfect.

Isotonic regression needs data too. It cannot rescue heads whose problem is a lack of data; it just adds variance where variance is already the issue. Worth revisiting only for heads above ~100 positives, which are not the ones in trouble.

So: data is the lever, as ranked. Now working #1 — structurally diverse positives for the indicative heads with deep public chemistry. 38 descriptors added to CURATED covering citrus, almond, bready, earthy, cumin, oregano, thyme, dill, bergamot, violet, soapy, powdery, the berry/tropical set, the spice set, and the fatty/alcoholic ones. Supplement resolved to 2,157 associations across 165 descriptors (was 1,938 across 160). Retraining now; will report the capable/indicative split against the 94/73 baseline either way.

[rvnminers-A-and-N]

Progress update after #270 and #271.

	heads	confident-capable	indicative
when this issue opened	167	94	73
now	166	108	58

14 heads promoted, 15 fewer indicative.

Two boundaries on the method are now established, and they change what is left to do here:

1. Distinctive, not merely present. Adding molecules that occur across a whole category actively harms a narrow head. blackberry got ethyl butyrate / ethyl hexanoate / linalool and fell 0.730 → 0.626, off the roster entirely; stripping back to distinctive markers recovered it to 0.757. The same mistake was then repeated on chamomile (sesquiterpenes shared across many heads) — it got worse, 0.691 → 0.670, and cost four confident heads before being reverted.

2. Chemical coherence decides whether curation can work at all (#257). fatty looked like a broad umbrella but is structurally crisp, and sixteen molecules took it 0.21 → 0.51. sweet, ethereal and pungent are broad and incoherent — a sweet smell can be a lactone, a furanone, a vanillin, an ester or carbon tetrachloride — and moved zero on precision despite the same treatment.

So the 58 remaining split into two piles, and only one belongs to this issue. The coherent-but-thin heads are still worth curating here, one careful head at a time with distinctive chemistry. The incoherent ones cannot be fixed with data at all and belong to #199, where the prediction is now pre-registered.

Anyone picking this up: read docs/ACCURACY.md first, then work from the near-floor end of training/audit_generalization.py output. Verify each head individually — a batch that looks like a net win can hide a head going backwards, which is exactly how chamomile slipped through.