Promote fatty, and give every molecule a readable name

[rvnminers-A-and-N]

Closes #217. Follows on from #257 (closed with its finding recorded).

Two data efforts, each of which ended somewhere other than where it started.

#257 — the broad heads split cleanly

Calibration had already lifted sweet/ethereal/pungent over the 50% floor, but only by making them very reluctant to fire (thresholds 0.75 / 0.64 / 0.66), sitting exactly on 0.50. Added food-authorised character molecules to all three, plus fatty as a control:

head	n_pos	AUROC	precision
fatty	27 → 41	0.854 → 0.917	0.21 → 0.51	promoted
sweet	208 → 210	0.718 → 0.727	0.50 → 0.50	unmoved
ethereal	98 → 106	0.924 → 0.920	0.50 → 0.50	unmoved
pungent	289 → 295	0.792 → 0.802	0.50 → 0.50	unmoved

"Broad" was never the discriminator — chemical coherence is. fatty looks like an umbrella but is structurally crisp (mid-chain aldehydes, free fatty acids, lactones), so sixteen molecules more than doubled its precision. A sweet smell can be a lactone, a furanone, a vanillin, an ester — or carbon tetrachloride. No shared substructure means no fingerprint can find it, and curation can't touch it.

Hard boundary on the #247/#256/#262 method, pre-registered on #199 as a falsifiable GNN prediction. No industrial odorants were removed — the additive approach held.

#217 — the premise was wrong, and finding that out was the work

A full PubChem crawl on Title and IUPACName (the crawler had only ever asked for IUPACName, which is why inorganic salts and glycosides came back empty) resolved 41 of 784. I then theorised PubChem's SMILES matching was brittle and tested InChIKey lookup — it did worse, 1/12 against 2/12.

The remaining 743 aren't lookup failures. They're the long tail from the broad food/safety universe:

	count
multi-component salts	159
very large structures	197
peptide-like	100
carbon-free	13

PubChem indexes salt components, not the mixture, and most of these aren't flavour-relevant. No crawl will name them.

So every molecule now falls back to its molecular formula, with mixtures labelled honestly rather than pretending to be one substance — CaI2O6 (mixture: Ca+2 + IO3- + IO3-). A raw SMILES in the grid looks broken; a formula tells a chemist something real.

master_enrichment: 8,861 molecules, 8,861 named — down from 770 blank.

Also: a reverted failure

The chamomile rescue backfired. Its additions were sesquiterpenes shared across many heads (bisabolol, farnesene, germacrene D) — chamomile got worse (0.691 → 0.670) and the roster lost four confident heads. Same generic-molecule mistake blackberry taught us earlier. Reverted; 108 capable restored.

Verification

• ruff clean; suite 43 passed
• roster: 166 aroma heads, 108 confident-capable, 58 indicative
• rebuilt: master_enrichment (8,861), profile_index (8,842 × 177), flavor_map
• needs sudo systemctl restart flavormancer — models and all three artifacts changed