feat: mouthfeel map colours, taste namespacing, cross-listed terms, full head coverage

[rvnminers-A-and-N]

Closes #241. Part of #240. Follows #247.

Five related pieces, each verified against the running app rather than assumed — which is how the head-coverage bug below got caught.

Full head coverage — three heads were silently invisible. The design index only ever tagged sweet/bitter/umami, so sour, salty and tasteless were trained heads the UI pretended didn't exist. It now iterates every trained taste head, and all six are offered unconditionally (the rule aroma already used, so a head with few confident carriers stays selectable rather than vanishing). Verified live: taste:sour → acetic/citric acid, taste:salty → potassium chloride, taste:tasteless → 138 matches. Audit now reads taste 6/6, aroma 164/164, mouthfeel 5/5.

Tox heads accounted for, in the right role. The 12 Tox21 heads are not design targets — nobody formulates for assay activity — so making them pickable would be wrong. They surface instead as a "hide tox-flagged" filter plus a per-result ⚠ N tox flags marker, caution-only framing intact. All 187 heads now have a role.

Taste namespacing (taste:sweet), matching mouthfeel. sweet is both a taste head and an aroma head, and a bare tag conflated "tastes sweet" with "smells sweet"; bitter had additionally leaked into the aroma-notes picker. Verified distinct: taste:sweet → citric acid, coumarin · aroma sweet → benzaldehyde, butanol.

Map colours by mouthfeel — build_flavor_map labels each molecule with its dominant sensation; new mode beside taste/aroma/both with cool→warm swatches. Rebuilt to 8,834 points, 100 mouthfeel-labelled.

Cross-listing (#241). Rather than blanket-linking (most non-overlap is legitimate — there's no aroma head for "bacon", nobody orders "amber flavor"), aliases cover terms that are genuinely the same thing spelled differently: 5 auto-derived by normalisation (black pepper↔blackpepper, bread↔bready, butter↔buttery, cream↔creamy, malt↔malty) so new vocabulary keeps linking itself, plus 10 curated synonyms (blackcurrant↔cassis, tangerine↔mandarin, orange blossom↔neroli, cilantro↔coriander, chocolate↔cocoa, licorice↔anise, cheese↔cheesy, smoke↔smoky, peppermint/spearmint↔minty). Deliberately excludes lemon↔citrus — citrus is broader, and equating them would silently widen a search. Verified symmetric: both sides of each pair return identical match counts.

Tingling generalisation (#247). The head was memorising: trained only on Zanthoxylum sanshools it fired on exactly its own training molecules — a lookup table wearing a 1.00 AUROC. Broadened with paresthesia amides from other genera (Echinacea, Anacyclus, Heliopsis, Piper) for structural diversity rather than more of the same scaffold: 11 → 18 positives, 0 → 2 novel hits; pungent also improved 5 → 12 novel. docs/MOUTHFEEL.md gains a "Does a head generalize, or is it memorizing?" section with the measured table — a high AUROC alone proves nothing. #247 stays open to broaden tingling further; it's still the weakest of the five and the docs say so.

ruff check . clean and 25 unit tests pass.