feat(api): sour as both trained head + rule; ranked taste_profile

Sour now trains as a small-data INDICATIVE head (AUROC 0.90), surfaced as
sour_predicted, while its acidity rule stays the deterministic 'sour' boolean — so
the chemist sees both signals (they can disagree: MSG rule=True but model=0.12).
Add taste_profile: trained heads ranked by probability (the 'order of dominance'
view), so MSG reads umami-first instead of an ambiguous set of flags. Salty stays
rule-only. API contract updated.

Closes #39.

Author: rvnminers-A-and-N

docs/API-CONTRACT.md

@@ -51,13 +51,21 @@ Single-molecule flavor read.
   "bitter": 0.74,                           // (trained)
   "umami": 0.03,                            // (trained)
   "sweet_intensity": 1.8,                   // vs sucrose — OPTIONAL (trained)
-  "sour": false,                            // (rule)
+  "sour": false,                            // RULE call — acidic-group structural rule (rule)
   "sour_reason": [],                        // acidic groups matched (rule)
+  "sour_predicted": 0.12,                   // trained sour head — small-data INDICATIVE (trained)
   "salty": false,                           // (rule)
   "salty_reason": "no alkali-salt structure", // (rule)
   "known_tastes": ["bitter"],               // OPTIONAL, verified dataset labels (lookup)
   "multitaste": false,                      // 2+ taste heads ≥ 0.5 (trained-derived)
 
+  "taste_profile": [                        // tastes ranked by dominance (trained heads, desc)
+    { "taste": "bitter", "probability": 0.74, "basis": "trained" },
+    { "taste": "sweet",  "probability": 0.12, "basis": "trained" },
+    { "taste": "sour",   "probability": 0.12, "basis": "trained (indicative)" },
+    { "taste": "umami",  "probability": 0.03, "basis": "trained" }
+  ],
+
   "physchem": {
     "computed":  { "mol_weight": 152.15, "logP": 1.21, "tpsa": 46.5,
                    "h_bond_donors": 1, "h_bond_acceptors": 3,
@@ -98,6 +106,11 @@ Single-molecule flavor read.
 **Field notes for implementers**
 - Taste-head keys (`sweet`/`bitter`/`umami`) are present **per trained classifier**; a
   head below the data threshold is absent and the corresponding rule/flag covers it.
+- **Sour carries two signals:** `sour` is the deterministic acidity-rule boolean;
+  `sour_predicted` is a small-data **indicative** trained probability. They can disagree
+  (the rule flags structure, the model reflects perception) — surface both.
+- **`taste_profile`** ranks the trained heads (incl. sour-indicative) by probability,
+  descending — the "order of dominance" view. The `sour`/`salty` rules stay separate flags.
 - `sweet_intensity` and `physchem.measured` appear only when their model/table is loaded.
 - `salty` may be overridden to `true` with `salty_reason: "verified (dataset label)"` when
   a ground-truth label exists (lookup beats rule).

training/predict.py

@@ -636,6 +636,23 @@ def predict_aroma(smiles, top_k=8):
         return {"available": False, "note": "aroma model load ok but prediction failed", "detail": str(e)}
 
 
+def _taste_profile(out):
+    """Trained taste heads ranked by probability (descending) — the 'order of
+    dominance' view. Sour is a small-data indicative head; the deterministic
+    sour/salty rules remain separate flags (out['sour'], out['salty'])."""
+    ranked = []
+    for t in ("sweet", "bitter", "umami"):
+        v = out.get(t)
+        if isinstance(v, (int, float)):
+            ranked.append({"taste": t, "probability": round(float(v), 3), "basis": "trained"})
+    sp = out.get("sour_predicted")
+    if isinstance(sp, (int, float)):
+        ranked.append({"taste": "sour", "probability": round(float(sp), 3),
+                       "basis": "trained (indicative)"})
+    ranked.sort(key=lambda e: e["probability"], reverse=True)
+    return ranked
+
+
 def predict(smiles: str, include_aroma: bool = False) -> dict:
     mol = Chem.MolFromSmiles(smiles)
     if mol is None:
@@ -644,6 +661,10 @@ def predict(smiles: str, include_aroma: bool = False) -> dict:
     out = {"smiles": Chem.MolToSmiles(mol)}
     for name, clf in sorted(_CLASSIFIERS.items()):
         out[name] = round(float(clf.predict_proba(x)[0, 1]), 3)
+    # Sour trains as a small-data INDICATIVE head, but its boolean stays the rule's
+    # call below — keep the model probability separately as sour_predicted.
+    if "sour" in out:
+        out["sour_predicted"] = out.pop("sour")
     if _INTENSITY is not None:
         out["sweet_intensity"] = round(float(_INTENSITY.predict(x)[0]), 2)
     out.update(_sour(mol))
@@ -661,6 +682,7 @@ def predict(smiles: str, include_aroma: bool = False) -> dict:
     strong = [t for t in ("sweet", "bitter", "umami")
               if isinstance(out.get(t), float) and out[t] >= 0.5]
     out["multitaste"] = len(strong) >= 2
+    out["taste_profile"] = _taste_profile(out)
     out["physchem"] = physchem(mol)
     out["stability"] = stability(mol)
     out["chemesthesis"] = chemesthesis(mol)

training/train_taste.py

@@ -2,10 +2,10 @@
 train_taste.py — taste heads (multi-taste) + sweetness-intensity regressor.
 
 Reads the merged dataset from build_taste_dataset.py and trains one binary
-RandomForest per structure-driven taste (sweet/bitter/umami). Sour and salty
-are validated RULES in predict.py (sourness is a pH/solution property, saltiness
-a cation property), never trained — by design, not just data volume. Add more
-sweet/bitter/umami data and those heads sharpen on the next run.
+RandomForest per trainable taste (sweet/bitter/umami, plus sour as a small-data
+*indicative* head). Salty stays a validated RULE in predict.py (a cation
+property, too few labels to model); sour ALSO keeps its acidity rule there as a
+deterministic cross-check. Add more data and the heads sharpen on the next run.
 
 Even with everything merged this trains in minutes on the R620 CPU. The
 multi-day budget is the aroma model (train_odor.py), not this.
@@ -27,9 +27,10 @@
 from sklearn.metrics import roc_auc_score, r2_score
 
 BASIC = ["sweet", "bitter", "umami", "sour", "salty"]
-# Sour and salty are validated RULES by design (pH/solution and cation properties),
-# handled in predict.py and never trained — regardless of how much data accrues.
-RULE_TASTES = {"sour", "salty"}
+# Salty stays a validated RULE only (cation property, too few labels to model).
+# Sour trains as a small-data INDICATIVE head but ALSO keeps its acidity rule in
+# predict.py as a deterministic second check (surfaced as sour_predicted + sour).
+RULE_TASTES = {"salty"}
 FP_BITS, FP_RADIUS = 2048, 2
 _MORGAN = rdFingerprintGenerator.GetMorganGenerator(radius=FP_RADIUS, fpSize=FP_BITS)
 # Below this, a taste is too thin for an HONEST head, so it's skipped and